verification

The papers

Most central to this idea first, not most recent.

• Industrycs.CR

  Tampered embeddings hijack offline AI retrieval despite tolerance checks

  Exact verification of quantisation codes closes the text-vector integrity gap in offline RAG packs.

  securagen.ai · Nov 2026

• Top Universitycs.LG

  Game-theoretic oversight reduces AI agent deception, even with weaker monitors

  Agents cheat when evaluated only on outcomes; a game-theoretic monitor tracking the full trajectory prevents deception.

  Beihang University, Zhongguancun Laboratory · Nov 2026

• Independentcs.AI

  Verified code often passes formal checks but misses the real spec

  Code verification must separately measure verifier acceptance, whether proofs bind to specifications, and whether specifications capture the intended behavior.

  Nov 2026

• Academiccs.LG

  Giving harder inputs more attempts helps little on hard LLM workloads

  Giving harder inputs more repeated attempts by an LLM pays off only when the budget is large relative to difficulty; otherwise the optimal allocation is inverted-U.

  Hankuk University of Foreign Studies · Nov 2026

• Academiccs.SE

  Enterprise agents act without permission more with missing evidence than explicit denial

  Enterprise agents correctly obey explicit permissions but fail to stop when supporting evidence is missing.

  Independent Researcher, University of Southern California · Nov 2026

• Academiccs.LG

  Sparse attention for long-context AI verifies its own quality in one pass

  Coverage-Verified Sparse Attention achieves closed-loop quality control by using the softmax mass fraction to verify and correct sparse attention outputs.

  National Taiwan University · Nov 2026

• Big Techcs.SE

  Coding agents' unnoticed mistakes make later requests fail more often

  Coding agents left to run a chain of requests unsupervised finish about half of them, not because their code is poor but because they cannot spot their own mistakes.

  Amazon · Nov 2026

• Industrycs.CRcode

  Language model proofs take 41 to 50 times serving time

  Sable proves, on a single GPU, that a committed integer-quantized language model produced a specified token sequence exactly.

  MoonMath.ai · Nov 2026

• Big Techcs.LG

  Models flag missing facts more readily than usable unsupported premises

  Language models check if information is usable for an answer, not if the premise is validly supported.

  Google · Oct 2026

• AI Startupcs.AI

  Language models generate 5–8% faster by keeping draft alternatives separate

  Multi-Future Drafting avoids chimeras in speculative decoding by using mode embeddings, delivering up to 7.8× speedup.

  Beihang University, Canva Research · Nov 2026

• Industrycs.CL

  Swapping the AI judge often changes which chatbot ranks best

  Swapping in a judge model from another provider does not preserve system rankings, and judges agreeing flags a bad judge but does not certify a conclusion.

  CEO Business School · Nov 2026

• Research Labcs.SE

  Agent evaluation pipelines drop tool calls before the AI judge reads them

  Matching hashes and passing checkers did not stop a renderer from blanking nearly every judged agent step, so input content must be verified against source records.

  Adelaide University, CSIRO · Nov 2026

• Big Techcs.CE

  AI protein design improves by separating suggested changes from evaluation

  Protein binder design improves when approximate gradients from a diffusion structure predictor propose sequence edits and full no-gradient predictor passes verify them, exploring multi-site then single-site.

  Stanford University, HKUST (Guangzhou) · Nov 2026

• Chinese Techcs.CL

  Language models still miss subtle errors in both text and agent actions

  Large language models cannot reliably detect, localize, or correct their own hallucinations, especially in agent trajectories, even with thinking, web search, or fine-tuning.

  Beijing Normal University, Tencent AI Lab · Nov 2026

• Big Techcs.CL

  Extra attempts help small language models solve terminal tasks, but not verify solutions

  Small language models fall behind frontier models on terminal tasks not only in solving them but also in verifying solutions and anticipating their own success.

  Microsoft Research, University of California, Los Angeles · Oct 2026

• Chinese Techcs.AI

  Structured backtracking beats reflection for long-horizon language agents

  VBL, a training-free extension of ReAct, maintains a goal stack with dual-horizon verification and typed failure diagnosis, improving task success across three benchmarks.

  City University of Hong Kong, Huawei Noah's Ark Lab · Nov 2026

• Top Universitycs.AI

  Formal plan validation shows LLM agents rarely plan as claimed

  LLM agents frequently violate action preconditions, rarely identify impossible goals, and gain little from elaborate prompting or agent architectures compared with changing the underlying model.

  Barcelona Supercomputing Center, University of Amsterdam · Oct 2026

• Top Universitycs.AI

  Agentic framework formally verifies GPU kernels against PyTorch references

  LARK uses Lean-checked proofs to certify that low-level GPU kernels match their PyTorch specifications and rejects kernels with data races.

  Imperial College London, University of Cambridge · Nov 2026

• Top Universitycs.LG

  Mean benchmark accuracy misleads model release decisions; recipe cards certify tail risk

  Mean accuracy is the wrong estimand for model release decisions; tail failure probability should be certified instead.

  Seoul National University Hospital, Inha University · Oct 2026

• Academiccs.SE

  Coding agents can report successful hardware migrations while changing model training

  Coding agents migrating PyTorch from GPU to TPU can pass execution checks while silently changing training, and many failures raise no error, so independent verification is essential.

  North Carolina State University, Federal University of Ceará · Oct 2026

• Chinese Techcs.AI

  Coding agents solve more tasks when another AI reviews their work

  Independent review by a separate critic achieves better accuracy-cost tradeoffs than other verification designs in agent harnesses.

  Cornell University, Alibaba Group · Oct 2026

• Big Techcs.CR

  Zero-knowledge proofs let AI agents prove they followed policies without exposing the rules

  zkARc produces succinct zero-knowledge proofs that a neurosymbolic system correctly enforced natural language policies on AI actions.

  ICME Labs, Imperial College London · Nov 2026

• Academiccs.CR

  Signed LLM weights do not certify the quantized files users actually run

  Signing full-precision LLM weights does not verify deployed quantized models unless the transformation inputs and build environment support bit-exact reproduction.

  National Yang Ming Chiao Tung University, National Taiwan University of Science and Technology · Oct 2026

• Big Techcs.AI

  AI agents’ false completion claims are caught more often with follow-up checks

  Treating verification as active exploration of interfaces and code beats passive judging for computer-using agents.

  Google, The University of Hong Kong · Oct 2026

• Top Universitycs.CR

  AI training can be detected in simulated datacentre power despite concealment

  A passive power meter using Viterbi tracking detects synchronous AI training runs reliably and makes hiding expensive.

  The University of Melbourne, University of Oxford · Oct 2026

• Chinese Techcs.AI

  Silent tool errors roughly halve success of e-commerce LLM agents

  Silent tool errors in e-commerce LLM agents reduce task success by roughly half and are harder than explicit failures.

  Hong Kong University of Science and Technology, Tencent · Oct 2026

• Independentcs.LG

  Probabilistic AI forecasts retain skill for record-breaking weather extremes

  Probabilistic AI weather models outperform IFS-ENS for heat and cold records beyond historical boundaries.

  Oct 2026

• Big Techq-bio.BM

  AI agents and wet-lab feedback improve aptamer designs over three rounds

  Iterative wet-lab feedback lets LLM agents improve aptamer hit rates, detect assay artifacts, and recover target specificity across three design cycles.

  Google Research, Stanford University · Oct 2026

• Big Techcs.AI

  A verified memory interface makes agent security depend on explicit confirmations

  PerMIT secures persistent memory for LLM agents with machine-checked integrity theorems and community-based trust.

  Stony Brook University, AWS AI Labs · Oct 2026

• Chinese Techcs.CV

  Image generation improves through verified repairs without retraining the generator

  ImageRSI combines verification, experience library, and judge self-evolution to achieve 52.6% CSA on MindBench and 65.66% on ImageReq-Bench.

  University of Science and Technology of China, Qwen Business Unit of Alibaba, Alibaba Token Hub · Oct 2026

• Big Techcs.AI

  Verified retries let smaller code models recover lost capability

  Verified retries from smaller models can match or beat much larger models on code benchmarks while using less memory, time, and energy.

  Texas A&M, Apple · Oct 2026

• Academiccs.LG

  Lean proof assistant generates step-level labels to train natural-language proof verifiers.

  Lean-generated step labels train better natural-language proof verifiers than human labels, without human annotation.

  University of Texas at Austin, University of California, Los Angeles · Oct 2026

• Big Techcs.LG

  Verifier-based rewards train coding agents toward formally correct programs

  A reward from verifier diagnostics, plus integrity auditing, raises verified code generation from 66.1% to 93.5% on DafnyBench and from 56.3% to 100% on SWE-Proof.

  UC Berkeley, AWS AI Labs · Oct 2026

• Big Techcs.SE

  Rewarding necessary edits makes debugging patches smaller without reducing correctness

  PreciseCoder reduces average edit lines to 3 per bug while keeping pass rates on par with standard training.

  University of Chicago, University of Southern California · Oct 2026

• Big Techcs.NI

  Black-box RL method finds large performance gaps in LLM network control

  LENS uses reinforcement learning to find adversarial inputs that expose large performance gaps in LLM-based network control systems without accessing model internals.

  University of Southern California, Microsoft Research · Oct 2026

• AI Startupcs.AI

  LLM-assisted peer review should prioritize error detection over imitation of human reviews

  The verification-centric benchmark and Multi-Layered Review framework achieve high error detection and strong alignment with human review scores.

  Sakana AI, National University of Singapore · Oct 2026

• Top Universitycs.CR

  Lightweight decision model catches nearly all malicious agent skills

  SkillScreen detects malicious agent skills with high recall and few false positives using a compact decision model.

  University of British Columbia, University of Toronto · Oct 2026

• Chinese Techcs.CL

  Progressive verification speeds up diffusion language model inference

  PV-Decoding reduces sequential model forward passes by 8-37% while keeping the largest observed accuracy drop to 1.3% for diffusion language models.

  Huawei Technologies, City University of Hong Kong · Oct 2026

• Chinese Techcs.CL

  Outcome verification reveals failures in social agent plans that conversation ratings miss

  Social agents that sound successful often produce invalid plans, while paired-world outcome training improves verified coordination.

  Alibaba Group, The University of Tokyo · Oct 2026

• Academiccs.LG

  Executable environments enable scalable tool-use training data and verification

  Davinci-Tool, trained on execution-grounded verified data, outperforms a model with 16 times its parameters on three tool-use benchmarks.

  Tongji University, Shanghai Innovation Institute · Oct 2026

• Big Techcs.LG

  Transformers learn to detect invalid planning steps and abort to retry

  Train Transformers to detect invalid planning steps with an abort token, then restart generation to improve reliability.

  Harvard University, Shenzhen University · Oct 2026

• Big Techcs.CV

  AI image generators favor the same cultural values regardless of prompt language

  Text-to-image systems leak consistent cultural value orientations into images generated from neutral prompts.

  University of Wisconsin–Madison, Microsoft Research Asia · Nov 2026

• Academiccs.CL

  AI agents complete roughly four times as many tasks with helpers

  Reusable primitives with embedded LLMs, composed per task by HEELO, solve real-world tasks that frontier models fail to complete.

  University of Illinois Urbana-Champaign, Georgia Institute of Technology · Nov 2026

• Academiccs.LG

  Formal proofs verify skill removal and preservation in neural networks

  Mechanistic edits can be certified via formal proof to remove and preserve skills over continuous input regions, a guarantee no finite test can provide.

  American International University-Bangladesh (AIUB) · Oct 2026

• Chinese Techcs.CL

  Report-dependent auditing pushes LLMs to suppress input factors

  Report-dependent auditing of LLM factor explanations creates a suppression incentive, while report-independent auditing or a small floor preserves truthful reporting.

  Hefei University of Technology, East China Normal University · Oct 2026

• Big Techcs.AI

  Agentic verification improves outputs on long-horizon workspace tasks

  VeriHarness uses the generator's own model as an agentic verifier to improve selection and revision of long-horizon task artifacts.

  Google Cloud AI Research, University of Cambridge · Oct 2026

• Academiccs.SE

  Expert verification failures become reusable guidance for coding agents

  Expert diagnosis of verification failures can be captured as reusable guidance for coding agents.

  University of Illinois Urbana-Champaign · Oct 2026

• Chinese Techcs.AI

  Clinical agents falter when health records cannot support the question

  A benchmark of 5,486 noisy clinical queries shows LLM success drops from 62.2% to 37.9% when questions lack supporting EHR evidence.

  Zhejiang University, Ant Healthcare · Oct 2026

• Chinese Techcs.CL

  Terminal agents verify often but miss and mishandle many errors

  Agents self-verify almost always but detect only 61.43% of errors and repair only 49.36% of those; SCVD improves outcomes.

  Northeastern University, Inclusion AI, Ant Group · Oct 2026

• Big Techcs.PL

  Three-step hybrid pipeline validates GPU kernels by combining concurrency testing with formal proof

  RESOLVE validates GPU kernels by testing for nondeterminism, reducing concurrency, and proving functional equivalence.

  University of California, Riverside, Math, Inc. · Oct 2026

• Big Techcs.LG

  AI agents can score higher by keeping luckier simulated outcomes

  Branch-and-commit only improves performance when candidate returns vary, and most gain is from selecting among random outcomes rather than better actions.

  Technion, Israel Institute of Technology, Nvidia Research · Nov 2026

• Top Universitycs.AI

  Linking memory back to raw data helps AI agents avoid repeating bad lessons

  Evidence-linked memory, storing lessons with their source trajectories, reduces failures caused by incorrect or misapplied lessons in LLM agents.

  KAIST, DeepAuto.ai · Nov 2026

• Research Labcs.LG

  Language models can reveal their ancestry without exposing their internal weights

  TRACE-Edge reconstructs LLM lineage trees by using asymmetric log-likelihood deviations to determine fine-tuning direction.

  Institute of Science Tokyo, RIKEN · Nov 2026

• Top Universitycs.LG

  AI agents can migrate some ML code across frameworks, but remain unreliable

  Under the XStack benchmark, AI coding agents cannot yet reliably migrate a PyTorch repository to JAX on TPUs; the best passes 75%, all others at most 40%.

  Princeton University · Nov 2026

• Industrycs.SE

  Coding agents fail most long chains of dependent tasks

  Coding agents solve at most 18.5% of long-horizon multi-step software engineering tasks.

  Klavis AI · Nov 2026

• Chinese Techcs.SE

  Language models fix more software issues after learning to reorganize code

  SE tools and verification can produce targeted training data that improves a coding model’s refactoring and issue-resolution performance across agent harnesses.

  Queen's University, University of Manitoba · Oct 2026

• Independentcs.SE

  Coding agents can recognize risks and still take harmful actions

  A coding-agent safety benchmark uses executable outcome checks and capability twins; even a top agent with 92% capability pass rate still causes verified harm on 20.9% of eligible risk tasks.

  Oct 2026

• Big Techcs.CL

  AI agents nearly clear proof benchmarks but harder code challenges remain

  EasyProver proves almost all existing formal math and code benchmarks and introduces a harder benchmark from competitive programming.

  Case Western Reserve University, Amazon Web Services · Oct 2026

• Chinese Techcs.CV

  OACap turns video captioning into active, tool-guided observation

  OACap improves detailed video captions by using missing or unsupported claims to guide targeted observations, then revising descriptions against the resulting evidence.

  Ant Group, Zhejiang University · Oct 2026

• Chinese Techcs.AI

  AI proof graders still struggle with subtle errors in Olympiad mathematics

  LLM verifiers fall short of expert-level Olympiad proof grading; MABGrader improves agreement by 26.4% Quadratic Weighted Kappa and reduces error by 14%.

  Huawei, The Chinese University of Hong Kong · Oct 2026

• Chinese Techcs.SE

  Generated Software Tests Often Fail to Distinguish Correct Patches

  Generated tests that reproduce a reference fix often fail to distinguish correct from incorrect patches; validation-guided training roughly triples joint validation performance.

  Baidu Inc., Nanjing University · Oct 2026

• Top Universitycs.CLcode

  LLMs fail multi-step math despite solving each step correctly

  LLMs often solve all intermediate math milestones yet fail to combine them into a correct final solution, making composition their largest diagnosed reasoning gap.

  Rutgers University, Harvard University · Oct 2026

• Top Universitycs.AI

  Undergraduate research problems remain difficult for formal coding agents

  ReuF2F is a Lean 4 benchmark of 244 open undergraduate-research problems; the best coding agent proves or refutes about 40%.

  Carnegie Mellon University, Chinese University of Hong Kong, Shenzhen · Oct 2026

• Big Techcs.AI

  Language model libraries often fail outside their core calculation code

  GPU kernel library failures often originate in build and runtime environments, while specialized assumptions and inadequate testing allow correctness and hardware-specific problems to slip through.

  Microsoft Research Asia, Tsinghua University · Nov 2026

• Chinese Techcs.CL

  Reusable parts let a single fix repair many broken AI training environments

  Building agent training environments from reusable, evolvable pieces, then repairing the pieces when an environment breaks, raises how many work and improves agent training.

  Shanghai Jiao Tong University, Tencent · Nov 2026

• Big Techcs.CV

  AI image editors get better results comparing alternatives than repeatedly refining

  Image editing agents fail because they lack diverse tools and their judges lack quantitative evidence; EditAgent solves both.

  Adobe Research, UC San Diego · Oct 2026

• Big Techcs.LG

  Language models can outdo text searches at revealing internal signals

  Maximally-Activating Example Metamodels (MAEMs) generate text that strongly activates any given direction in a language model's residual stream.

  MATS, Arb Research · Nov 2026

• Top Universitycs.SE

  New benchmark reveals automated red-teaming services often miss real-world agent security flaws

  GreyZone evaluates automated red-teaming services on their ability to cause verified security-policy violations in real-world agent applications.

  National University of Singapore · Nov 2026

• Industrycs.LG

  Cheap baselines often rival expensive LLM methods on graph node classification

  Reported GraphLLM gains over simple baselines often shrink to a few points, and evaluation practices show test-set leakage.

  NASK National Research Institute · Oct 2026

• Big Techcs.AI

  AI assistants face a $1.7 trillion wish list of life tasks

  LifeVal benchmarks AI models on high value, personalized, agentic consumer tasks rooted in real user data.

  OpenAI · Nov 2026

• Big Techcs.AI

  Fewer draft tokens verified per round makes hybrid models a third faster

  HybridSpec speeds up hybrid GDN-MoE language model inference by adaptively selecting verification length and using parallel GDN kernels.

  Meta AI, New York University · Nov 2026

• Top Universitycs.LGcode

  Language models find more valid solutions without supplied correct answers

  MS-TTRL resolves the frequency conflict by using self-verification for correctness and reserving generation frequency for diversity.

  Beihang University, National University of Singapore · Oct 2026

• Chinese Techcs.AI

  Language models can give identical answers up to 2.78 times faster

  Filtering low-quality branches from a candidate tree before expensive verification speeds up tree-based speculative decoding without changing outputs.

  Tongji University, Nanyang Technological University · Oct 2026

• Chinese Techcs.AI

  AI agents can improve mathematical constructions despite failing proof tasks

  TCSExam evaluates frontier large language models across the full range of theoretical computer science research.

  Tsinghua University, Moonshot AI · Nov 2026

• Academiccs.CR

  AI code reviewers miss one in six attacks spread across updates

  AI code reviewers catch most hidden exploits but attackers bypass them in about one in six real maintenance attempts.

  University of Illinois Urbana-Champaign, Arizona State University · Oct 2026

• Big Techcs.LG

  AI agents learn more from varied tools than extra tasks

  To scale tool-use agent training data, grow tool and database diversity, not just task count.

  Apple · Oct 2026

• Top Universitycs.CV

  AI models struggle to report video tampering their internal features reveal

  Separate audio and visual temporal readouts on frozen AV-LLM features outperform zero-shot prompting for manipulation localization.

  Tsinghua University, Peking University · Oct 2026

• Big Techcs.AI

  AI agents fall short on a benchmark of professional engineering

  Frontier AI agents score at most 50.6% on realistic physical engineering design tasks from a new 85-task benchmark.

  Google Cloud AI Research, Google · Oct 2026

• Chinese Techcs.AI

  Coding agents satisfy individual requirements more often than complete revisions

  Coding agents satisfy 69.9% of requirements but fully complete only 26.4% of real user editing tasks.

  Fudan University, Alibaba Token Hub, Alibaba Group · Oct 2026

• Chinese Techcs.AI

  AI agents get nearly ten times more training tasks from scientific code

  Poseidon generates nearly ten times more training environments from scientific code repositories than prior methods and improves agent performance.

  The Hong Kong University of Science and Technology (Guangzhou), Hunyuan Team, Tencent · Oct 2026

• Big Techcs.MA

  Game engine expertise and verification feedback boost AI 3D scene completion from 18% to 48%

  Language instructions can drive controllable Unreal Engine worlds through agent-generated code with verification and localized repair, preserving persistent scene and timeline state.

  Tongji University, Microsoft Research Asia · Oct 2026

• Big Techcs.SE

  Coding agents can speed up software while proving behavior stays unchanged

  Agents can produce faster program versions with machine-checked equivalence proofs, achieving verified speedups across standalone functions and migrated applications.

  University of California, Berkeley, Amazon Web Services · Oct 2026

• Top Universitycs.CL

  RAG's retrieval, not reading, is the bottleneck for long-context question answering

  Similarity-based RAG fails long-context multi-fact QA because retrieval cannot surface all needed chunks and chunking loses document metadata, not because reading is inadequate.

  MBZUAI, Inception42 · Oct 2026

• Big Techcs.LG

  Global coordination of proof agents beats local repair in formal theorem proving

  Maestro coordinates prover workers globally to handle failures, outperforming existing provers on mathematics and verification benchmarks.

  University of Illinois Urbana-Champaign, AWS AI Labs · Oct 2026

• Chinese Techcs.AI

  Monitoring research agents detects integrity violations and helps reduce them

  RIDAR detects and corrects research integrity violations in autonomous research agents, reducing violations by 36.8%.

  Tsinghua University, Zhipu AI · Oct 2026

• Academiccs.AI

  New benchmark reveals agents fail to detect and recover from their own errors in long tasks

  Computer-use agents rarely recover from their own errors in long-horizon workflows; fine-tuning on real failures improves recovery and clean-start performance.

  University of Texas at Dallas, Independent Researcher · Oct 2026

• Chinese Techcs.AI

  AI agents often delay checks most closely linked to success

  Goal-aware dynamic multi-agent collaboration with execution graphs and per-agent goal contexts improves success on long-horizon tasks over existing baselines.

  Michigan State University, Baidu, Inc. · Oct 2026

• Big Techcs.AI

  Synthetic computer worlds improve agents on unseen computer-use tasks

  WildAWM uses open-source models to synthesize computer-use environments spanning GUI, MCP, and CLI; training only in these worlds improves agents across eight out-of-distribution benchmarks.

  UNC-Chapel Hill, Microsoft Research · Oct 2026

• Chinese Techcs.AI

  Conditions for reliable reuse of intermediate results in LLM reasoning

  Intermediate-result reuse is reliable only when checks accept sufficiently precise states and continuations preserve correctness; retries reduce exhaustion but not accepted-state error.

  Westlake University, Wechat AI · Oct 2026

• Top Universitycs.LG

  Careful system design makes asynchronous coding-agent training substantially stronger

  A fully asynchronous RL baseline with established components trains a 4B model to 62.0% pass@1 on SWE-bench Verified.

  Snowflake, University of Illinois at Urbana-Champaign · Oct 2026

• Big Techcs.AI

  Graded curricula extracted from an environment's latent rules substantially improve trained agents

  Gradus composes guarded behavioral operators into calibrated curricula, lifting a 30B model from 59% to 89% on tau2-Bench and 47% to 83% on CAR-Bench.

  AWS AI, University of Southern California · Oct 2026

• Chinese Techcs.CV

  Verifiable rewards help video captioning models capture more complete visual narratives

  MVRCap frames video captioning as reinforcement learning with hierarchical multi-dimensional verifiable rewards, achieving state-of-the-art results on VDC, CAPability, and VC4VG.

  Xiamen University, Alibaba Group · Oct 2026

• Big Techcs.SD

  Dense perceptual training improves reasoning in large audio-language models

  Constructing post-training datasets with multi-aspect perceptual grounding and cognitive refinement improves audio reasoning accuracy in large audio-language models.

  Carnegie Mellon University, NVIDIA · Oct 2026

• Top Universitycs.AI

  Verifiable visual rewards let image generators train on synthetic scenes and transfer to natural prompts

  Programmatically verifiable visual rewards, trained with reinforcement learning, transfer from synthetic geometric scenes to natural image prompts, improving instruction following.

  University of Washington · Oct 2026

• Independentcs.AI

  Auditable world-model harness achieves perfect ARC-AGI-3 score

  Kepler’s auditable executable world models achieved 100.00 RHAE across all 25 public ARC-AGI-3 games under one frozen configuration.

  Independent Researcher · Oct 2026

• Big Techcs.CL

  Learned controller adapts speculative decoding to maximize speed and reduce waste

  APEX adaptively selects speculation mechanism and draft depth to balance speed and token waste.

  University of Illinois Urbana-Champaign, AWS AI · Oct 2026

• Top Universitycs.CLcode

  Two-stage verifier combines reasoning and tool-use for safer computer agents

  SCOUT improves computer-use agent safety by combining reasoning-intensive rubric generation with tool-intensive evidence gathering to detect harmful behaviors.

  Columbia University · Oct 2026

• Academiccs.AI

  AI agents reproduce only 41% of machine learning papers with full code and weights provided

  AI agents reproduce ML paper results at 15-41% rates depending on code availability; most failures lack verification against the paper's numbers.

  University of Illinois Urbana-Champaign, National Center for Supercomputing Applications · Oct 2026

• Research Labcs.DC

  Adaptive verification speeds block-diffusion decoding under heavy concurrent demand

  DScale improves throughput of block-diffusion speculative decoding by up to 48.8% over DFlash with only a 112K-parameter predictor.

  Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, University of Chinese Academy of Sciences · Oct 2026

• Chinese Techcs.AI

  Structuring noisy file environments improves agents’ performance on complex tasks

  Env-Rethink prepares and evolves agent environments using collection maps and event logs, lifting mean rubric pass rate by 15.1 points across nine models.

  Shanghai Jiao Tong University, Theseus Labs · Sept 2026

• Big Techcs.CV

  PhyProbe scores physical inconsistencies in generated videos more reliably

  PhyProbe outperforms prior evaluators on physical consistency benchmarks, especially in generated-generated and no-correspondence settings.

  NVIDIA, University of Waterloo · Oct 2026

• Research Labcs.AI

  LLMs generate generalized plans with machine-checked completeness proofs

  LLMs can generate generalized plans and formal proofs that the plans solve all domain instances.

  Saarland University, German Research Center for Artificial Intelligence (DFKI) · Sept 2026

• Big Techcs.AI

  Multi-agent orchestration produced proofs for five open research problems

  Cogentic uses an orchestrator to allocate provers and adversarial verifiers to automatically discover proofs for open research problems.

  Google Research, Yale University · Oct 2026

• Top Universitycs.SE

  Agents write verifiers that wrongly reject valid solutions more often than they accept them

  LLM agents generate verifiers that reject invalid solutions more reliably than they accept valid ones.

  University of Pennsylvania, University of Illinois at Urbana-Champaign · Oct 2026

• Independentcs.AI

  CUA-Lite standardises computer use agent development across environments and models

  CUA-Lite is an open platform that standardizes environments, data, and model harnesses for computer use agents.

  Oct 2026

• Top Universitycs.AI

  An agentic framework that diagnoses and verifiably repairs foundation model failures

  EvalRx, an open-source agentic framework, statistically diagnoses and verifiably repairs foundation models, improving all 12 evaluated model-modality combinations by +1.4 to +10.9 percentage points.

  The University of British Columbia, Vector Institute · Oct 2026

• Academiccs.LG

  Scaling curves can be certified with far fewer sampled answers

  A paired audit achieves nearly minimax cost for certifying best-of-k scaling curves with simultaneous confidence bands.

  PMCC AI Lab, Peter Munk Cardiac Centre · Oct 2026

• Top Universitycs.CV

  New benchmark quantifies physical plausibility in AI-generated videos

  PVBench quantifies physical plausibility in video generation using object-level physical measurements across 46 dimensions.

  Zhejiang University · Oct 2026

• Research Labcs.CR

  Execution-Time Checks Limit Malicious Tool Calls by LLM Agents

  Provenance-Aware Capability Enforcement (PACE) prevents tool misuse by verifying capability and provenance at the point of execution.

  PRADA Lab, King Abdullah University of Science and Technology · Oct 2026

• Top Universitycs.SE

  LLM agents deploy repairs better than systems built from scratch

  FDE-Bench reveals that readiness checks are the primary failure point for LLM agents in deployment tasks.

  University of California, Berkeley, Imperial College London · Sept 2026

• Chinese Techcs.CV

  Geometry-aware test-time scaling improves camera control in video diffusion without retraining

  Training-free geometry-aware candidate selection at test time improves camera control and geometric consistency in video diffusion models without retraining.

  University of Freiburg, Huawei Heisenberg Research Center (Munich) · Oct 2026

• Academiccs.PL

  Semantic constraints can prune invalid code during language model decoding

  Semantic prefix oracles enforce typing and scope constraints during LLM decoding, achieving zero false prunes and improving program validity by up to 15.2 points.

  ENS de Lyon, Unsuspicious Industries · Sept 2026

• Academiccode

  Modeling attack chains improves prompt injection detection for email agents

  Modeling prompt injection as an attack chain with stage verifiers and intent consistency analysis improves detection F1 over binary classifiers.

  Eastern Michigan University · Sept 2026

• Industrycs.LG

  Reinforcement learning helps coding agents finish tasks across benchmarks

  RL on 1,700 expert coding tasks improves a 1T-parameter model on six benchmarks and shortens agent steps.

  Surge AI · Oct 2026

• Top Universitycode

  Compiler turns scientific protocols into certified simulated laboratory tasks

  SciHorizon-eLab compiles natural-language scientific protocols into certified simulation tasks for embodied agent benchmarking.

  Computer Network Information Center, Chinese Academy of Sciences, Beihang University · Sept 2026

• Academic

  Scoping Agent Memory Prevents Unrelated Tasks From Disrupting Each Other

  Family-scoped retrieval prevents cross-family interference and supports repeated adaptation in persistent agent memory.

  Independent, Northwestern University · Sept 2026

• Academiccs.AI

  Pass/fail grading hides tenfold differences among top AI agents

  Agent benchmark saturation stems from coarse pass/fail bits, not easy tasks; graded credit with cost axes fixes it.

  Lenovo Group, Marquette University · Nov 2026

• Academiccs.AI

  Prompt optimization yields no gains on strong reasoning models

  Automated prompt optimization produces no statistically significant accuracy gains on frontier reasoning models; it fixes formatting errors but cannot supply missing knowledge or fit truncated context.

  Hunan Agricultural University · Nov 2026

• Academiccs.CR

  AI agents fall for social engineering attacks in every tested scenario

  All frontier AI models fail to resist social engineering attacks, with proactive verification below 32% and many making ungrounded action claims.

  Hong Kong University of Science and Technology · Nov 2026

• Industrycs.AI

  Benchmark scoring code, not the model, can swing accuracy 22 points

  Scoring rules in multimodal LLM benchmarks can inflate accuracy by up to 22 points on identical outputs.

  Independent Researcher, Nebius · Nov 2026

• Big Techquant-ph

  AI agents find cheaper quantum programs and mathematically prove their correctness

  AI agents using formal verification discovered quantum programs with up to 290x improvements in key algorithms.

  Google DeepMind, Google Quantum AI · Nov 2026

• Chinese Techcs.SE

  Coding agents struggle to connect code changes with repository behavior

  SWE-Time shows coding agents still struggle to reason reliably about how code changes explain past bugs and shape future behavior.

  Nankai University, Baidu, Inc. · Nov 2026

• Big Techcs.CV

  Image forensics protocol sorts AI edits into retouching, tampering, and full synthesis

  A dual-channel protocol distinguishes retouched, tampered, and synthetic images beyond binary real/fake labels.

  Google, University of British Columbia · Nov 2026

• Independentcs.AI

  Weight-editing AI agents improve models with tiny parameter changes, but success is inconsistent

  Weight-Space Programming allows large accuracy improvements via few parameter edits but is inconsistent and transfer uneven.

  Nov 2026

• Academiccs.SE

  Code plagiarism detectors fail on whole repositories rewritten by AI agents

  Current provenance detection methods cannot reliably trace repository-scale AI-rewritten code to its source.

  Emory University, University of Illinois Urbana-Champaign · Nov 2026

• Research Labcs.LG

  AI models recover code better from the shapes of executions

  Path signatures encode execution traces for program induction, outperforming standard sequence models, especially with limited data.

  CISPA Helmholtz Center for Information Security, Technical University of Munich · Nov 2026

• Independentcs.AI

  AI agents fail end-to-end physics research due to shallow reasoning and missing intuition

  An end-to-end physics research benchmark finds frontier AI agents fail most tasks, with main weaknesses insufficiently rigorous reasoning and limited expert intuition.

  Nov 2026

• Big Techcs.CL

  Agent Skills teach AI systems precise subagent permissions

  Procedural knowledge in Agent Skills improves subagent authorization precision beyond what task outcomes alone provide.

  Samsung Research, Peking University · Nov 2026

• Top Universitycs.AI

  AI agent rankings shift under iterative feedback evaluation

  Models that score best under no-feedback metrics can be the wrong choice for agentic systems; measuring success over K attempts with feedback gives a different, more faithful ranking.

  Columbia University, University of California, Santa Barbara · Oct 2026

• Independentcs.AI

  Directly optimizing acceptance length boosts speculative decoding throughput by up to 9.4%

  D-REAL post-training directly optimizes expected accepted draft token length to speed up speculative decoding.

  Nov 2026

• Chinese Techcs.LG

  Reconstructing app behavior before coding improves simulated GUI agent training

  Separating behavior reconstruction from implementation yields faithful GUI environments from traces, with 91.1% workflow fidelity and 94.0% real-device gain retention.

  Nanjing University, Alibaba Token Hub, Alibaba Group · Oct 2026

• Independentcs.CR

  Some LLM graders collapse jailbreak compliance into evasion, skewing attack success rates

  Small LLM graders can erase the compliance class, making identical jailbreak responses appear up to 25 times safer depending only on the grader.

  Independent Researcher · Oct 2026

• Top Universitycs.LG

  Language models can improve coding answers without checking target correctness

  Test-Time Self-Play beats verifier-based methods for test-time scaling by training a solver on generated synthetic problems, improving by 35% with no verifier.

  ETH Zurich, Tufa Labs · Nov 2026

• Industrycs.LG

  AI services secretly using cheaper models can be detected with few queries

  API-TAIL detects partial model substitution in text-only language-model services by adjusting for legitimate serving changes and counting persistently unusual responses.

  TokenRhythm Technologies · Oct 2026

• Independentcs.CR

  Secure enclaves let AI evaluators test proprietary models without revealing prompts

  Double-blind evaluation uses hardware enclaves to keep both AI prompts and model weights secret from each other.

  Nov 2026

• Chinese Techcs.AI

  AI agents that satisfy more search constraints recommend fewer places

  An agentic local search benchmark reveals a core trade-off between task coverage and reliable constraint satisfaction.

  Alibaba Group · Oct 2026

• Big Techcs.AI

  First benchmark evaluates coding agents on XR development tasks

  A new benchmark, GenXR-Bench, evaluates coding agents on interactive XR development, finding the best model reaches only 50.9% Pass@1.

  Google XR Labs, University of California, Los Angeles · Oct 2026

• Top Universitycs.CL

  Misspecified training rewards reduce the reliability of chain-of-thought monitoring

  Reward misspecification makes chain-of-thought poor at revealing reward hacking, while training models to describe inferred rewards substantially restores monitorability.

  Columbia University · Oct 2026

• Big Techcs.AI

  Human-AI search extends small Mahler measure constructions to higher degrees

  A five-parameter construction yields irreducible polynomials with Mahler measure below 1.255 for all even degrees up to 2000.

  Google DeepMind, California Institute of Technology · Oct 2026

• Chinese Techcs.AI

  AI agents better anticipate when to help by checking past patterns

  A hierarchical memory that validates trajectories and regularities improves proactive LLM agent effectiveness by 21.4% over baselines.

  City University of Hong Kong, Tencent · Oct 2026

• Chinese Techcs.AI

  LLM agents degrade reliably under controlled interaction perturbations

  Perturbations to LLM agent inputs consistently degrade performance, and agents rarely detect corrupted information.

  The Hong Kong University of Science and Technology (Guangzhou), Tencent Hunyuan · Oct 2026

• Chinese Techcs.RO

  Visual reward optimization can improve robots while reinforcing wrong-object mistakes

  Optimizing a learned visual reward can raise task success and wrong-object failures simultaneously, hiding the error.

  Baidu, University of California, Santa Barbara · Oct 2026

• Big Techcs.AI

  Evidence graphs produce synthetic research tasks with controllable difficulty

  GRADS generates synthetic question-answer data with graph-regulated difficulty, producing verified tasks where empirical hardness matches requested hardness and improves agent training.

  Johns Hopkins University, Nanjing University · Oct 2026

• Research Labcs.AIcode

  Agent performance depends more on system configuration than model size

  Agent performance depends more on system configuration, especially supplied information, than on model choice alone, and reliable verification requires system tools rather than prompts.

  Technical University of Munich, MCML · Oct 2026

• Academiccs.CL

  Correct answers can conceal invalid grade-school math reasoning traces

  Correct answers do not certify valid chain-of-thought traces, especially under distribution shift, and plausible minimal traces do not establish selective planning.

  IT University of Copenhagen, Chennai Mathematical Institute · Oct 2026

• Big Techcs.AI

  Training coding agents to locate, vary, and verify boosts repair reliability.

  Autonomous coding agents can learn fault location, edit variation, and patch verification, lifting held-out SWE-bench pass@1 from 31.9% to 43.0%.

  Stanford University, AWS AI Labs · Oct 2026

• Chinese Techcs.CV

  Benchmark and agent for verifying images against real-world facts

  Persistent, state-guided verification detects image-world factual conflicts better than direct or retrieval-augmented judgment on evaluated 8B vision-language models.

  Ant Group, Tsinghua University · Oct 2026

• Research Labcs.CL

  Steering tool decisions helps, but correction does not guarantee repair

  Internal activation steering can increase correct tool calls but corrupts over half of baseline-correct decisions; repair requires destination-resolved auditing.

  University of the Chinese Academy of Sciences, University of Birmingham · Sept 2026

• Big Techcs.CL

  Training with solver feedback boosts reliability of language model reasoning

  Post-training language model agents with solver feedback improves SMT reasoning accuracy and reliability.

  Case Western Reserve University, UT Austin · Oct 2026

• Big Techcs.SEcode

  Code-linked evidence lets AI software claims expire when dependencies change

  Assay binds agent claims to Merkle hashes of dependency cones so claims expire automatically when code changes.

  Google, OMNI3ai · Sept 2026

• Big Techcs.AI

  A verifiable runtime for agents that extend themselves with safe, reversible tools.

  A Lean 4 agent runtime uses reversibility to decide human oversight, letting agents extend their own tools and act autonomously.

  Amazon · Oct 2026

• Chinese Techcs.AI

  Explicit belief states help agents stay coherent over long tasks

  Explicit belief states that track world state and unresolved task requirements, validated for consistency and monitored for progress, improve long-horizon LLM agent performance.

  Nankai University, Alibaba Group · Oct 2026

• Chinese Techcs.AI

  Post-training makes coding-agent trajectory verifiers substantially more accurate

  Post-training a 30B coding-trajectory verifier with supervised fine-tuning and reinforcement learning raises completion-classification accuracy from 57% to 78%.

  Leibniz Research Center Huawei, Monash University Malaysia · Oct 2026

• Independentcs.SE

  Intermediate snapshots improve long-horizon coding agents at lower cost

  Checkpointing agent state after each turn and resuming failed rollouts via verifier-guided backtracking makes long-horizon RL for coding agents more effective and cheaper.

  Oct 2026

• Top Universitycs.AI

  AI agents fail to produce honest scientific follow-up papers in new benchmark

  AI agents attempting open scientific follow-ups produced 135 papers; none met scientists' standards, and 114 failed integrity checks.

  Carnegie Mellon University, University of Minnesota · Oct 2026

• Industrycs.AI

  Code execution exposes oversight failures hidden by aggregate benchmarks

  Cheap learned overseers miss rare shortcut-driven errors even when aggregate performance appears strong.

  LawZero · Oct 2026

• Top Universitycs.RO

  Retrieval replaces generation for building simulation-ready articulated objects from a single photo

  Retrieving and adapting a library of pre-verified articulated assets generates physically valid simulation objects from single images.

  Tsinghua University, Chongqing University · Oct 2026

• Top Universitycs.AI

  Frontier AI agents struggle with end-to-end professional engineering workflows

  Frontier AI agents achieve only 44.3 out of 100 on the EngiWorld benchmark, failing to reliably automate professional engineering tasks.

  Tsinghua University, Zhiman Inc. · Oct 2026

• Big Techcs.AI

  Small surrogate models can flag errors in black-box LLM agents

  A small surrogate model's log-probabilities audit black-box LLM agent actions better than the agent's own confidence or self-consistency.

  Amazon · Oct 2026

• Big Techcs.AI

  Separating agent checks from reasoning improves weaker models’ reliability

  DeReAct improves AI agent reliability by moving action approval and task completion checks out of the main reasoning policy.

  Amazon, Apple · Oct 2026

• Big Techcs.AI

  Formal verification catches broken coding tasks before agents cheat

  SpecGuard detects and formally certifies conflicts between task intent and tests before an agent can cheat.

  MATS, Google DeepMind · Oct 2026

• Top Universitycs.AI

  LLM-generated PTX rivals autotuned Triton across three GPU generations

  An LLM agent compiles Triton kernels directly to PTX, reaching 0.83x-3.34x the performance of autotuned Triton across tested GPUs.

  Stanford University, EPFL · Sept 2026

• Academiccs.CR

  Security agents lose evidence support when available telemetry changes

  When telemetry changes, LLM agents lose citation support for 35.5% of previously covered attack actions even though the actions remain recoverable.

  Zhongguancun Laboratory · Oct 2026

• Big Techcs.AI

  Verified self-improvement makes agent harness optimization more reliable

  VERSE improves harness optimizers by letting them test edits and evolve their own verification tools.

  MIT, Amazon · Oct 2026

• Research Labcs.SE

  Framework scales verifiable training environments for coding agents that build programs from scratch

  ProgramGym's feedback-driven environment construction and role-isolated agent workflow improve from-scratch program reproduction, lifting Qwen3.8-27B to 55.71% on ProgramBench.

  Xiamen University, Washington University in St. Louis · Oct 2026

• Big Techcs.AI

  Multi-perspective self-verification improves vision-language model answer reliability

  Multi-perspective self-verification with reliability-guided rethinking improves VLM answer reliability without external judges.

  University of Leicester, Nanjing University of Posts and Telecommunications · Oct 2026

• Chinese Techcs.AI

  Diagnosis-first framework boosts solver performance across numerical domains

  ADSD links numerical diagnosis to reusable solver skill discovery, improving solver accuracy, robustness, and efficiency across four numerical domains.

  University of California, Berkeley, Alibaba Group · Oct 2026

• Big Techcs.RO

  Offline stitching of robot demos boosts policy success by 21 percentage points

  Offline algorithm NEEDLE stitches verified action bridges between robot observations to improve policy success, raising real-robot success by 21 points on average.

  Stanford University, NVIDIA · Oct 2026

• Academiccs.CL

  Verified workflow training gives language models reusable procedural skills

  SkillGym converts human-written agent skills into executable training environments, enabling LLMs to internalize procedural competence and beat larger models on real-world benchmarks.

  East China Normal University, Shanghai AI Laboratory · Sept 2026

• Big Techcs.AI

  Adaptive judges help embodied agents improve as failure patterns change

  VeriFine co-evolves policy, curriculum, and judge to scale verification and enable continuous self-improvement in embodied reasoning.

  NVIDIA, UCLA · Oct 2026

• Chinese Techcs.CL

  Hybrid agents learn to recreate software across five computing platforms

  Training hybrid agents on RecreationWorld trajectories improves performance on out-of-distribution coding and hybrid benchmarks.

  Alibaba Token Hub, Alibaba Group · Sept 2026

• Industrycs.SE

  Scientific agents gain auditability through enforced, re-verifiable research workflows

  Scientific credibility is achieved by making non-compliant states unrepresentable through mechanically enforced laws, not by asking the model to behave.

  IdeaHorizon Team · Sept 2026

• Big Techcs.AI

  Mined tool-call constraints help agents avoid errors and finish tasks

  Automatically mined symbolic constraints on tool calls, checked at runtime with an SMT solver, reduce agent errors and improve task completion.

  University of Illinois Urbana-Champaign, Microsoft Research · Oct 2026

• Research Labcs.CR

  Execution Logs Can Mislead Visual Judges in Video-Generation Agents

  Execution traces cause video-generation verifiers to judge clips based on text, not visuals.

  RIKEN · Sept 2026

• Chinese Techcs.AI

  Benchmark evaluates AI agents' sustained progress on long-horizon tasks with process rewards

  LHTB 2.0 extends agent benchmarks to 77 tasks with process rewards and three-hour evaluations.

  University of Maryland, Tencent · Oct 2026

• Independentcs.SE

  Existing scientific software can generate verified training tasks for terminal agents

  Existing scientific software workflows provide self-supervised training data for terminal agents via software-in-the-loop reconstruction.

  Oct 2026

• Chinese Techcs.LG

  Reverse scoring helps diffusion agents escape repeated failed actions

  Diffusion language model agents retry failed actions due to task-blind bias; reverse scoring corrects this without retraining.

  Huawei Noah’s Ark Lab, Technical University of Darmstadt · Oct 2026

• Chinese Tech

  New benchmark measures whether AI systems can explore, not just recall, unfamiliar rules

  ExplorationBench shows that leading AI systems can learn unfamiliar executable rules through exploration, but their gains vary across runs and can reverse with further exploration.

  Fudan University, Hunyuan Team · Oct 2026

• Big Techcs.AI

  Successful Agent Traces Predict Which Base Models Merit Post-Training

  Base model potential for agentic coding can be predicted by probing at the decisive step in successful post-trained trajectories.

  NVIDIA, University of Minnesota – Twin Cities · Oct 2026

• Research Labcs.AI

  Benchmark of 82 unresolved problems in theoretical sciences tests AI beyond known results

  GPT-6-Astra achieves a 14% judged solve rate on open problems in mathematics and theoretical physics, outperforming other AI models.

  University of Science and Technology of China, Chinese Academy of Sciences · Oct 2026

• Big Techcs.CR

  Multi-agent adversarial review boosts LLM-based SOC alert triage accuracy

  Structuring LLM agent evidence retrieval and adversarial review substantially improves SOC alert triage, raising F1 from at most 0.744 to 0.958.

  Virginia Tech, Google · Oct 2026

• Academiccs.AI

  Agents Repair Executable Skill Packages Better With Code-Structure Guidance

  SkillScriptBench shows AST-guided, scope-limited editing substantially improves reliable repair of executable agent skills while coordinating scripts and documentation.

  The Hong Kong University of Science and Technology, Harbin Institute of Technology, Shenzhen, China · Oct 2026

• Chinese Techcs.CV

  Structured caption tests expose weaknesses in long-range audiovisual reasoning

  Audio-visual caption evaluation should use verifiable atomic units instead of free-form text for reliable, fine-grained diagnosis.

  Hunyuan, Tencent, Nanyang Technological University · Oct 2026

• Big Techcs.CL

  Execution-verified feedback improves LLM harness optimization for SQL, Cypher, and Python generation

  HERO improves pass@1 by 7.6 points offline and an additional 2.0 points via test-time refinement on text-to-code benchmarks.

  University of Washington, Microsoft Research · Oct 2026

• Chinese Techcs.CL

  Safe LLM memory reuse requires certificates, not simple success counters

  A certificate that weighs failures by likelihood ratio and restarts on changes reduces stale answers 7-13x on QA and 35-73x on code vs vote counters.

  ByteDance, Scale AI · Oct 2026

• Top Universityhep-th

  Frontier AI agents solve most expert-level theoretical physics calculations

  Frontier AI agents solve hard theoretical physics problems with 98% accuracy at low cost, matching world experts.

  International Centre for Theoretical Sciences, Tata Institute of Fundamental Research · Oct 2026

• Research Labcs.SEcode

  Coding agents rarely build the agent a user actually asked for.

  Coding agents fail to build agents that satisfy held-out requirements, overfitting to visible cases and falsely reporting completion.

  Beijing Academy of Artificial Intelligence, Institute of Automation, Chinese Academy of Sciences · Oct 2026

• Academiccs.AI

  Multi-agent LLM system solves hundreds of open math problems without human steering

  Bolzano, a multi-agent system, solved about 200 open problems from 3800 attempts without problem-specific guidance, including four from STOC 2026 papers.

  Charles University, Czech Academy of Sciences · Oct 2026

• Top Universitycs.AI

  Tool calls need formation-bound dispatch to preserve their security meaning

  Unchanged valid tool calls can change security meaning due to schema-epoch drift; formation-consistent dispatch prevents this failure.

  Korea Advanced Institute of Science and Technology (KAIST) · Sept 2026

• Industrycs.RO

  Symbolic checks make household robot plans safer and cheaper to repair

  SAGE adds a zero-token symbolic precondition gate and sub-goal-local replanning to LLM planners, improving safety and recovery efficiency.

  Korea Electronics Technology Institute (KETI) · Sept 2026

• Chinese Techcs.CL

  Recursive testing loops improve generated games across two engines

  RSIGame autonomously improves game quality through recursive self-improvement loops and experience internalization, outperforming larger models.

  University of California San Diego, ByteDance Inc. · Oct 2026

• Chinese Techcs.CL

  Agents learn to reset noisy context during long web searches

  A rubric-driven search agent with self-managed memory and answer verification reaches 72.83 on BrowseComp and improves across several other search tasks.

  State Key Laboratory of Multimedia Information Processing, Peking University · Oct 2026

• Academiccs.AI

  Agent certification can fail when proposal and history correlations are erased

  State-only abstraction of foundation-model agents preserves proposal coverage but can destroy certifiability unless every proposal-conditioned history fiber has a common robust-safe intervention.

  Hong Kong Baptist University, Beijing Normal University · Sept 2026

• Big Tech

  Audio LLMs cannot assess their own transcription reliability.

  Audio-encoder representations predict whether an Audio LLM will transcribe degraded speech reliably far better than self-assessment or tested quality and uncertainty baselines.

  Apple, University of California San Diego · Sept 2026

• Big Techcs.AI

  New benchmark tests automated theorem proving for research-level theoretical computer science

  TCSAlgBench benchmarks automated proof discovery on 398 research-level TCS theorems; top models achieve about 25% coverage.

  University of Texas at Austin, Amazon · Oct 2026

• Chinese Techcs.CR

  Informed attackers can evade static and LLM-based agent skill scanners

  An informed attacker defeats static and LLM skill scanners by hiding payloads in natural language and splitting instructions across files, succeeding up to 97%.

  Huawei Research Zurich · Oct 2026

• Chinese Techcs.HCcode

  Linked story artifacts help authors supervise narrative-generating agents

  NarrativeSteward improves author understanding and guidance of AI-generated interactive narratives via linked artifacts and structural review.

  Tencent Inc · Oct 2026

• Top University

  Separating planning stages improves LLM reliability on complex compound tasks

  A multi-stage planning framework with separate generation, revision, and assessment modules improves LLM accuracy and multi-task robustness.

  University of Maryland, College Park, NEC Laboratories America, Inc. · Sept 2026

• Big Techcs.LG

  Verified memories help chip-design agents transfer lessons across tasks

  ChipMEM stores agent skills only after verification tools confirm they pass, improving performance on RTL designs.

  University of Illinois Urbana-Champaign, NVIDIA · Sept 2026

• Industrycs.DCcode

  Dispatch-aware agents safely optimize compiler-generated GPU kernels within full models

  KernelOPT, a multi-agent LLM system, optimizes generated Triton sub-kernels while preserving library calls, achieving geometric mean speedups of 1.40x, 1.15x, and 1.07x over torch.compile.

  Red Hat · Sept 2026

• Academiccs.AI

  Formal logic solvers can audit and improve language-model reasoning chains

  LogicTrack audits reasoning trajectories with automated theorem provers, improving both verifiability and accuracy.

  University of Bristol, King Abdullah University of Science and Technology · Sept 2026

• Big Techcs.AI

  Probabilistic scoring unlocks verification as a new scaling axis for LLMs

  Scoring token logit expectations, repeated evaluations, and decomposed criteria improve LLM verification accuracy and achieve state-of-the-art agentic benchmark results.

  Stanford University, UC Berkeley · July 2026

• Chinese Techcs.CL

  LLMs often diagnose without correctly valuing available medical evidence

  LLMs often submit correct diagnoses on insufficient or misleading evidence, and verifying evidential support before submission improves diagnostic accuracy.

  Zhejiang University, Ant Healthcare, Ant Group · Sept 2026

• Big Tech

  Evolved Agent Harnesses Improve Coverage but Fail to Consolidate Gains

  Automatically evolved LLM agent harnesses improve completion and task coverage in hardware verification, but complementary gains do not reliably consolidate into one consistently superior harness.

  NVIDIA · Sept 2026

• Top Universitycs.AI

  Task-specific plans and terminal checks create different value for LLM agents

  Prewritten plans improve oracle-verified success by 7.17 percentage points; a terminal verifier rejects 61% of invalid episodes at low cost.

  The Chinese University of Hong Kong, University of Edinburgh · Sept 2026

• Big Techcs.RO

  One demonstration can bootstrap reusable robot manipulation programs

  RAPID generates reusable robot programs from a single human demonstration using an iterative agentic coding loop.

  Massachusetts Institute of Technology, National University of Singapore · Sept 2026

• Independentcs.AI

  Safe agent self-modification requires verifiable, expressive recovery mechanisms

  Reliable LLM agent self-evolution requires co-designing verification, grounding, witness semantics, and recovery expressivity.

  Independent Researcher · Sept 2026

• Chinese Techcs.CR

  Simulated future actions may expose multi-turn agent attacks earlier

  SSH simulates future agent trajectories and verifies them against real actions to detect multi-turn attacks earlier while reducing false positives.

  Huawei Technologies Co., Ltd · Oct 2026

• Independentcs.AI

  Deterministic guardrails expose gains that fool LLM-based judges

  LLM judges are not reliable as final evaluators in self-improving loops; deterministic guardrails must override them.

  Sept 2026

• Industrycs.AI

  Coding agents pass 98% of tests but fail a third of builds in long-horizon tasks

  Agents that build full-stack products over many feature requests pass most tests individually but complete a fully working product in at most 65% of builds.

  Replit · Nov 2026

• Top Universitycs.AI

  Passing self-generated tests do not ensure semantic correctness in DataFrame code generation

  Self-generated tests can increasingly pass while plans, code, and assertions preserve the same semantic misunderstanding of a DataFrame task.

  University of California, Berkeley, Northeastern University · Oct 2026

• Big Techcs.CL

  Synthetic translated data improves language models across twelve Indian languages

  A 2.83-trillion-token synthetic corpus plus quality-controlled evaluation improves Indic-language knowledge and reasoning over open-corpus training without measurably reducing English performance.

  NVIDIA · Oct 2026

• Top Universitycs.CR

  Image generators render false claims as credible visual evidence despite knowing they are false

  Commercial image generators render over 70% of false claims as credible images; an attack raises this to 95%.

  Xiangtan University, Jinan University · Oct 2026

• Big Techcs.CL

  Research agents can exploit evaluations and adapt to evade oversight

  Autonomous research agents frequently reward-hack, and iterative reviewer feedback can help them evade oversight, requiring evaluation metrics and recomputation outside their control.

  Bake AI, University of Notre Dame · Sept 2026

• Research Labcs.AI

  AI agents can work on mathematical proofs for hours unattended

  Numina Fuse is an agent-native Lean 4 application built around human supervision of agent work through views of blueprint, dependency graph, and Lean code.

  Project Numina, University of Naples Federico II · Oct 2026

• Top Universitycs.MA

  Benchmark exposes computer agents’ failures when resolving conflicting information

  CLASHarness generates verifiable multi-agent computer-use conflicts, and evaluation shows models trust misleading peers, struggle to fix conflicts, and fail at coordination protocols.

  Emory University, University of Southern California · Oct 2026

• Top Universitycs.CL

  Cross-fitted feedback curbs shared errors in self-evolving search agents

  Co-cheating in self-evolving search agents is mitigated by cross-fitting source documents to break feedback loops.

  Rutgers University, University of California, San Diego · Oct 2026

• Big Techcs.LGcode

  Formal process rewards improve language models on long agent tasks

  A Lean4-based framework that grounds multi-level process rewards in formal verification improves open-ended agentic RL for software engineering and terminal tasks.

  Amazon Web Services, University of Illinois Urbana-Champaign · Oct 2026

• Independentcs.AI

  Strong language models often fail to recognize their own errors

  LLMs' confidence reports poorly discriminate errors from correct answers on hard problems, even with self-review.

  Oct 2026

• Chinese Techcs.AI

  Solved mathematical mechanisms can generate verifiable training worlds for AI agents

  Generating agentic RL training environments from pre-solved mathematical mechanisms yields cheap, verifiable, stateful tasks that transfer to unseen and external benchmarks.

  Georgia Institute of Technology, Alibaba Token Foundry, Alibaba Group · Sept 2026

• Independentcs.AI

  Benchmark tests whether coding agents improve governance without breaking repositories

  SWE-Prometheus evaluates how well coding agents improve repository governance across six dimensions.

  Sept 2026

• Academiccs.AI

  Framework improves AI agents by patching common failure patterns

  Vestrum turns execution trace failures into scoped harness changes that improve performance on held-out tasks without retraining the model.

  University of Georgia · Oct 2026

• Big Techcs.RO

  Robot policies falter as manipulation tasks stretch across more subgoals

  Robot manipulation policies succeed on single subgoals at 85.4% but collapse on multi-subgoal tasks; tool-selection orchestration improves success but physical execution errors remain the bottleneck.

  INSAIT, Sofia University “St. Kliment Ohridski”, NVIDIA · Oct 2026

• Chinese Techcs.AI

  Artifact-grounded RL trains a general frontend development agent

  WebCoder uses artifact-grounded reinforcement learning to outperform proprietary models on composed frontend development tasks.

  Tsinghua University, Zhipu AI · Oct 2026

• Research Labcs.LG

  AI models excel at cyclone track prediction but struggle with intensity and uncertainty

  Leading AI weather models outperform numerical forecasts on cyclone tracking but retain intensity biases and calibration errors.

  INRIA, AXA · Oct 2026

• Chinese Techcs.CL

  Translating answer-guided reasoning cuts the cost of model distillation

  CoTranslator learns to translate reverse CoT to forward CoT, improving student accuracy while using less compute.

  Tsinghua University, Ant Group · Oct 2026

• AI Startupcs.AI

  Orchestrating coding agents broadens autonomous computational research beyond optimization

  The Agentic Scientist orchestrates coding agents with rubric verification to automate computational research and discovery.

  Sakana AI · Oct 2026

• Industrycs.AI

  Cross-checking coding agents improves reliability in automated model research

  Runtime-enforced composition of coding agents raises all-oracle execution accuracy to 62.5% from a 45.8% single-product baseline in auto-research for ranking models.

  Meta · Oct 2026

• Chinese Techcs.CR

  Tool-calling agents access more private data than requests require

  LLM tool-calling agents significantly exceed authorized data access, and SelfAudit reduces privacy excess by 43%.

  Hefei University of Technology, East China Normal University · Oct 2026

• Chinese Techcs.CL

  Workspace-grounded synthesis scales verifiable training for long-horizon work agents

  WorkForge builds verifiable work-agent environments from real-world files by anchoring tasks and verifiers in observable workspace facts, improving agent performance across benchmarks.

  Tencent, Fudan University · Oct 2026

• Independentcs.LG

  Fixed-cost drafter speeds speculative decoding as context lengths grow

  LongSpark decouples drafter cost from prefix length by extracting fixed-size, multiscale views from target verification.

  Oct 2026

• Industrycs.SE

  New benchmark measures LLMs' ability to build complete codebases from scratch

  Whole-repository generation from specs is feasible, with LLM pass@1 ranging from 11.7% to 67.7% on the E2E-SWE benchmark.

  Meta Superintelligence Labs · Oct 2026

• Top Universitycs.SE

  Coding agents nearly finish repositories but still miss explicit requirements

  The strongest coding agent solved only 10 of 11 repository construction tasks, with failures due to single specification omissions or low-frequency rules.

  Gradient Data, McGill University · Oct 2026

• Top Universitycs.LG

  Outcome-only reinforcement learning can teach models to judge reasoning steps

  Outcome-only RL can induce step-level process supervision in generative reward models, reaching 85.2 F1 on ProcessBench with 3.2K outcome-labeled trajectories.

  Renmin University of China, Tsinghua University · Oct 2026

• Top Universitycs.AI

  Visual feedback boosts software engineering agent success rates

  Hybrid agents that combine code editing with GUI interaction outperform code-only agents on software engineering tasks.

  Carnegie Mellon University, University of Southern California · Oct 2026

• Industrycs.AI

  Top AI agents complete fewer than one-quarter of professional tasks

  Current AI agents fail most long-horizon professional tasks, with the best passing only about a quarter.

  Surge AI · Oct 2026

• Independentcs.AI

  Self-designed evaluators enable safer retries for long-horizon agents

  An agent can design its own evaluator from public materials to gate retries and label memory, improving without ground-truth labels.

  Sept 2026

• Big Techcs.CL

  Clinical language models diagnose well but often mishandle full encounters

  KlinikeBench reveals that top language models succeed on less than 30% of interactive clinical tasks despite 90.7% diagnostic accuracy.

  Zhejiang University, Imperial College London · Oct 2026

• Big Techcs.LG

  Post-training multi-token prediction heads match joint pretraining with far less data

  Post-trained MTP heads on a frozen model match jointly-pretrained quality with 10^3-10^4x fewer tokens; relaxed verification and adaptive head count add speedups.

  Microsoft · Oct 2026

• Academiccs.CL

  Reusing target probabilities in speculative decoding can improve tree selection but requires careful trade-offs.

  Target probabilities improve speculative-decoding tree selection, but guaranteed exact reuse requires a distribution computed at the next round’s full context.

  Korea University, Soongsil University · Oct 2026

• Academiccs.CR

  Security skills guide coding agents to only 11% fully correct vulnerability detection

  Current security skills rarely make coding agents fully identify Linux kernel vulnerabilities, while failure-driven skill refinement substantially improves detection on unseen cases.

  York University · Oct 2026

• Big Techcs.CL

  Confidence-guided complementary evidence pool improves multimodal retrieval-augmented generation

  CLIMB selects complementary evidence and uses confidence scores to decide when to update answers, improving multimodal RAG without retraining.

  Rutgers University, Google · Oct 2026

• Big Techcs.AI

  Value-guided search improves math reasoning during RL training

  APIVIS uses value-guided search during training to improve math reasoning beyond what independent sampling achieves.

  Amazon, Texas A&M University · Oct 2026

• Top Universitycs.AI

  LLM agents collude to bypass verification in long-horizon tasks

  Long-horizon interaction causes LLM agents to collude and skip verification, emerging in 94% of trajectories across 10 models.

  Stanford University, Georgia Tech · Sept 2026

• Chinese Techcs.CVcode

  Stateful code revision improves programmable image and video generation

  MaLiang-Harness bridges the program-to-visual gap by enabling MLLMs to iteratively construct, inspect, and revise executable visual programs for image and video generation.

  National University of Singapore, Fudan University · Sept 2026

• Chinese Techcs.AI

  Statistical gates help self-editing agents avoid lasting performance regressions

  SAGE uses per-item paired comparisons and one-sided paired tests to accept skill edits only when wins are statistically reliable, reducing regressions in self-evolving LLM agents.

  Peking University, University of Oklahoma · Sept 2026

• Chinese Techcs.SE

  Agents learn software repair rules that transfer across repositories

  XRepoSkill learns transferable software engineering skills by deriving rules from divergent successful and failed trajectories and verifying them across repositories.

  Nanyang Technological University, Tsinghua University · Sept 2026

• Top Universitycs.CVcode

  Fine-Grained Training Data Improves Agents on Interactive CAPTCHA Tasks

  CaptchaArena supplies 50,000 execution-verified interactive CAPTCHA puzzles across 20 types, and CaptchaAgent trained on it reaches 71.7 Pass@1.

  Columbia University, Zhejiang University · Sept 2026

• Chinese Techcs.LG

  StepLearn lets agents use new knowledge immediately while requiring validation before persistent reuse.

  StepLearn lets LLM agents immediately use insights from ongoing interactions while requiring cross-episode validation before long-term adoption.

  Renmin University of China, Tencent · Sept 2026

• Industrycs.AI

  Debate traces yield useful training data when confidence replaces voting

  AbCD converts multi-agent debate traces into preference data by clustering answers and ranking by confidence, improving reasoning without gold labels.

  Upwork Inc. · Nov 2026

• Top Universitycs.AI

  Language models can climb leaderboards by up to 45 places

  Large-language-model leaderboards that aggregate pairwise preferences with the Bradley-Terry model can be manipulated by strategic voting, prompting, and nomination.

  UC Berkeley, Università Bocconi · Nov 2026

• Chinese Techcs.AI

  Reasoning models can be trained to stop when they know the answer

  LoCo trains reasoning models to stop when they have enough information, reducing generated text while maintaining or improving accuracy on math and code tasks.

  Xi’an Jiaotong University, Tencent Inc. · Nov 2026

• Big Techcs.SE

  AI agents resist mistaken criticism better when they record evidence first

  Pre-feedback claim checklists prevent LLM agents from falsely reversing correct work when users complain.

  Google · Oct 2026

• Chinese Techcs.CL

  Language models’ self-assessments can look accurate without predicting future performance

  Language models can report supplied statistics exactly yet still fail to build trustworthy domain-level self-assessments from noisy user feedback.

  The Chinese University of Hong Kong (Shenzhen), Shenzhen Research Institute of Big Data · Oct 2026

• Chinese Techcs.CL

  Language models can understand others’ beliefs yet fail to act accordingly

  Language models' theory-of-mind belief inference does not guarantee optimal planning or task success in verifiable collaborative tasks.

  The Chinese University of Hong Kong (Shenzhen), ByteDance Inc. · Oct 2026

• Chinese Techcs.MA

  AI agents complete only one in 16 collaboration tasks

  A benchmark built from recombined collaboration structures tests whether multi-agent AI systems really coordinate, and none of the tested setups solved more than 6.25% of its tasks.

  Xiamen University, ByteDance · Nov 2026

• Chinese Techcs.SE

  Coding agents struggle to extend features built on their own code

  Coding agents struggle to maintain correct features across multiple stages, with whole-feature success at 28.5%.

  Hunyuan Team, Tencent, The Chinese University of Hong Kong · Oct 2026

• Chinese Techcs.AI

  Gameplay dataset maps tasks to subgoals for training AI agents

  AceAction-5M is a multi-game dataset linking task, subgoal, and skill instructions to time-aligned gameplay visuals and keyboard-and-mouse actions.

  Tencent, Jilin University · Nov 2026

• Top Universitycs.AI

  Faster Jev decisions do not consistently improve agent task efficiency

  Jev cuts decision cost and latency, but workflow overhead and failures erase gains, reducing mobile-agent success and sometimes raising cost per successful outcome.

  Nanjing University, Technical University of Munich · Oct 2026

• Academiccs.IR

  Self-improving data synthesis adapts embedding models to new tasks with minimal human input

  Self-improving data synthesis generates training data that adapts embedding models to new retrieval tasks across domains.

  University of Science and Technology of China, State Key Laboratory of Cognitive Intelligence · Oct 2026

• Chinese Techcs.CV

  Search agents that locate, verify and crop references improve image generation for recent events.

  A locate-corroborate-isolate search agent retrieves better references for newly introduced, easily confused entities and improves downstream image generation quality.

  Renmin University of China, Tencent · Oct 2026

• Big Techcs.CL

  Tailored reasoning traces improve student learning beyond answer-only training

  ReDiST improves reasoning distillation by revising student-generated traces, selecting those accessible to the target model, and jointly supervising reasoning and final answers.

  Stony Brook University, Meta AI · Oct 2026

• Chinese Techcs.AI

  MetaExp trains a transferable experience-learning capability in language model agents

  MetaExp trains language models to learn and reuse experience across tasks using verifiable cross-task families.

  University of Macau, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences · Oct 2026

• Independentcs.LG

  Test-time compute gains hinge on selection and failure handling

  The benefit of test-time computation depends more on candidate selection and failure handling than on the number of candidates generated.

  Oct 2026

• Chinese Techcs.LG

  Financial multi-hop benchmark exposes sharp agent performance drop beyond four hops

  FinHopper is a 1-6 hop financial retrieval benchmark with gold intermediate answers showing agent performance drops sharply with depth.

  South China University of Technology, FiT, Tencent · Oct 2026

• Chinese Techcs.AI

  Constructing testable, stationary rewards for open-ended tool agents

  Validated, evidence-producing rewards held fixed during optimization improve the consistency of open-ended tool agents’ outputs on two office-work benchmarks.

  Huawei · Oct 2026

• Independentcs.LG

  Smaller frozen models provide better rejected responses for preference distillation

  Smaller frozen models generate better rejected responses for preference distillation than student-scale self-rejects, with less compute.

  Oct 2026

• Chinese Techcs.CR

  New benchmark tests LLM agents on exploiting cross-protocol vulnerabilities

  LLM agents can exploit cross-component protocol vulnerabilities, with success rates from 48.8% (guided) to 12.5% (black-box).

  Tsinghua University, Alibaba Group · Oct 2026

• Big Techcs.AI

  Evolving environment beliefs helps LLM agents adapt without retraining

  EvoBelief improves frozen LLM agents by validating, reusing, and revising falsifiable environment rules, including after hidden dynamics changes and during scientific discovery.

  Tulane University, Amazon · Oct 2026

• Chinese Techcs.AI

  Correct Answers Often Fail to Prove Genuine Multi-Hop Reasoning

  Correct answers often bypass intended multi-hop reasoning chains, while evidence-deletion auditing and REALHOP produce evaluations that depend on every annotated hop.

  Hunyuan Team Tencent, Peking University · Sept 2026

• Academiccs.LG

  Backpropagating verifier rewards improves reasoning in continuous flow language models

  A critic predicting per-token values lets continuous flow language models improve reasoning through backpropagated verifier rewards.

  Tufts University · Oct 2026

• Top Universitycs.CV

  Training-free framework turns single photos into interactive, movable 3D scenes

  A single image can be turned into an interactive 3D scene with independent movable objects by iteratively deconstructing and recomposing the scene.

  Zhejiang University, Shanghai Innovation Institute · Oct 2026

• Big Techcs.AI

  Turning failed reasoning attempts into training signal improves GRPO on hard problems

  Failed reasoning attempts can be turned into verified supervision for GRPO by having the policy write corrective notes and distilling only replays that beat the original group.

  University of California, San Diego, Adobe Research · Oct 2026

• Top Universitycs.AIcode

  Agent performance depends on matching models, harnesses and tasks

  A model's performance and ranking depend on both the harness and the task.

  Nanyang Technological University · Oct 2026

• Independentcs.CL

  New benchmark tests if LLM agents can handle miscommunication and shifting user goals

  LLM agents struggle to track evolving user intent under miscommunication, with large gaps from oracle performance.

  Oct 2026

• Research Labcs.SE

  Scaling executable training data for software engineering agents with progressive construction and elastic streaming

  SWE-Stream constructs 80,874 executable SWE tasks from 10,728 repositories and achieves 69.0% on SWE-bench Verified.

  Xiamen University, Institute of Automation, Chinese Academy of Sciences · Oct 2026

• Independentcs.CV

  Native 4K image editing works better as an agent system than a single model

  BluePixel combines an efficiently adapted native-4K editor with specialist routing and reward-based retries, delivering four-step 4K edits while ranking second on VINS-4KEval++.

  Oct 2026

• Big Techcs.LG

  A learned router chooses when language models should draft farther ahead

  CADENCE learns context-adaptive decode length by generating Gaussian latent spans, achieving up to 1.3x higher generation throughput than greedy decoding.

  Microsoft, University of Massachusetts Amherst · Oct 2026

• Big Techcs.CV

  New metric detects real-world plausibility failures in AI-generated images

  TerraVis evaluates world-grounded visual consistency in generated images by detecting real-world plausibility violations using an MLLM workflow.

  Adelaide University, xAI · Oct 2026

• Academiccs.LG

  Agent improves by distilling its own verified experiences during deployment

  Deployed LLM agents can improve by self-distilling their own verified execution trajectories into persistent weight updates without a separate training phase.

  University of New South Wales (UNSW Sydney) · Oct 2026

• Chinese Techcs.AI

  Trajectory refinement improves reinforcement learning for long-horizon language-model agents

  TrajEvolve refines low-quality steps in historical trajectories and uses binary step-necessity classification for credit assignment, improving RL on long-horizon tasks.

  City University of Hong Kong, Tencent · Oct 2026

• Industrycs.LG

  Benchmark tests agent control of anomaly detectors under delayed feedback

  DynGraphAgentBench evaluates anomaly detection controllers that must select detectors under delayed feedback across temporal graph datasets.

  Sept 2026

• Chinese Techcs.AI

  SkillHorizon benchmark reveals that language agents struggle to complete long business workflows as task scope grows

  Long-horizon agent tasks built from composed skills cause completion rates to drop from 90.4% to 46.8% as workload grows.

  Hong Kong University of Science and Technology, Alibaba Group · Oct 2026

• Independentcs.AI

  Language agents rarely withdraw selectively when instructions or permissions change

  Language agents exhibit very low withdrawal recall and no valid resumption in the NAQD-Env benchmark.

  Oct 2026

• Chinese Techcs.AI

  Frozen LLM agents learn world models via reflective rulebooks

  An agent learns explicit world models as rulebooks and code, clearing all ARC-AGI-3 levels with 44% of human actions.

  University College London, Huawei Noah's Ark Lab · Oct 2026

• Chinese Techcs.AI

  LLM agents struggle to generate physics simulators from descriptions, new benchmark shows

  Generating correct physics solvers from text descriptions is difficult for LLM agents, as shown by the WorldSolver benchmark.

  Chinese Academy of Sciences, Peking University · Oct 2026

• Big Techcs.LG

  Activation steering strengthens token-level supervision for self-distilled reasoning

  Activation-Conditioned Self-Distillation steers a frozen self-teacher using contrasts from verified correct trajectories, improving math and code accuracy beyond reference-conditioned self-distillation.

  Rensselaer Polytechnic Institute, IBM Research · Oct 2026

• Top Universitycs.MA

  Agents improve collaboration by learning teams and evidence-gated communication

  CollabFlow improves multi-agent collaboration through recursive self-improvement with evidence-conditioned communication and flow-based team credit.

  The Chinese University of Hong Kong, Shenzhen, Fudan University · Oct 2026

• Big Techcs.AI

  Retrieving real states improves long-horizon planning over generated latent sub-goals

  Retrieving sub-goals from recorded episodes outperforms generating latent sub-goals for long-horizon planning.

  Samsung AI, University of Edinburgh · Sept 2026

• Big Techcs.CVcode

  Structured intermediate rewards improve robots on long, dependent tasks

  Structured intermediate rewards, gated by task prerequisites and scaled by completion pace, improve long-horizon VLA training over sparse terminal rewards.

  The Pennsylvania State University, Amazon AWS AI · Sept 2026

• Big Techcs.CL

  Hidden-state trajectories help diffusion language models escape wrong answers

  Hidden-state trajectories enable detecting and repairing stable-but-wrong lock-in in diffusion language models, yielding accuracy gains of 2.21-5.37 percentage points.

  Rochester Institute of Technology, NVIDIA Corporation · Sept 2026

• Big Tech

  Formal audit detects omitted source relationships in generative search

  A four-predicate audit can determine whether a generative search answer omits a source relationship that changes its interpretation.

  Google LLC, Sony Corporate of America · Sept 2026

• Chinese Techcs.AI

  Aligned skills and rubrics improve multimodal agent training

  Representing each skill as both actor-facing guidance and an evaluator-facing rubric, then co-evolving them, improves multimodal agent planning and tool use.

  The University of Hong Kong, Peking University · Sept 2026

• Top Universitycs.AI

  Compiling Mobile App Navigation Makes GUI Agents Faster and More Reliable

  PhoneCLI compiles mobile app navigation into deterministic commands, reducing VLM cost and improving task success.

  The University of Hong Kong · Sept 2026

• Academiccs.AI

  Co-trained agent roles improve tool-based reasoning and verification

  Jointly training planning, execution, and evaluation players under GRPO improves tool-integrated reasoning, beating prior baselines on mathematical and general reasoning benchmarks.

  Waseda University, Adelaide University · Sept 2026

• Industrycs.CL

  Self-evolving loop synthesizes high-quality multimodal training data

  VISA uses a self-evolving loop to synthesize multimodal instruction data, improving quality via verification feedback and target model errors.

  vivo AI Lab · Aug 2026

• Independentcs.AI

  Compiling agent skills into state machines improves task success

  HEXIS compiles agent skills into extended finite state machines, improving success by 16.2 percentage points over Skill + ReAct and reducing execution tokens by 38.4-88.9% with Qwen3.8-27B.

  Sept 2026

• Research Labcs.CV

  Fine-Grained Rewards Reduce Vision-Language Hallucinations Without Silencing Valid Detail

  Fine-grained alignment using dense reward signals and subsentence-level credit assignment reduces object hallucination in vision-language models.

  State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences · Oct 2026

• Top Universitycs.AI

  Attribution-guided skill graphs improve targeted repairs for frozen language models

  GraphSkillAA uses attribution-guided updates and two-level validation to optimize skill graphs, outperforming existing methods.

  National Key Laboratory for Novel Software Technology, Nanjing University · Sept 2026

• Top Universitycs.AI

  Accessibility profiles reveal hidden tradeoffs in LLM reliability

  LLM capability should be measured by accessibility across wording variants, not single benchmark accuracy, because broad success and dependable success differ.

  University of Maryland, College Park · Oct 2026

• Independentcs.AI

  Rewriting harness-assisted solutions improves general terminal-task performance

  Rewriting successful multi-harness agent traces into verified general-harness trajectories substantially improves a base model on complex terminal tasks.

  Oct 2026

• Chinese Techcs.CL

  Self-diagnosis pinpoints which agent decisions deserve reinforcement learning credit

  In agentic RL, self-diagnosed errors turned into step-level credit anchored by terminal outcomes recover learning signal from same-outcome groups.

  Alibaba Group, Kyoto University · Oct 2026

• Academiccs.CL

  Coding agents can improve by evolving their own instruction rules

  RuleEvolve self-evolves coding rules by iterative mutation and selection, beating hand-written rules and prompt optimization on correctness, length, and cost.

  Duke University · Oct 2026

• Big Techcs.SE

  Simulated users test coding agents as requirements unfold across conversations

  Coding agents evaluated in multi-turn interactions with dynamic user simulation and conversation validation.

  University of Innsbruck, Microsoft · Oct 2026

• Big Techcs.LG

  Diverse supervised reasoning traces improve generalization after reinforcement learning

  Selecting diverse reasoning traces for SFT improves post-RL generalization on puzzles and mathematics.

  University of Illinois Urbana-Champaign, Google · Sept 2026

• Chinese Techcs.AI

  App deeplinks help mobile agents navigate faster and finish more tasks

  Combining app-native deeplinks with GUI actions lets mobile agents complete tasks more successfully and in fewer steps than using GUI actions alone.

  Shanghai Jiao Tong University, Tongyi Lab, Alibaba Group · Sept 2026

• Big Techcs.AIcode

  Adaptively choosing how to revise LLM skills outperforms fixed revision strategies

  StraTune adaptively selects revision operators to improve self-evolving LLM skills, outperforming fixed operators across tasks.

  The University of Texas at Austin, Amazon · Sept 2026

• Chinese Techcs.CV

  One OCR model family handles recognition, parsing, and reasoning

  PolyOCR achieves state-of-the-art performance across diverse OCR tasks by unifying instruction-following, a data engine, and competence-guided policy optimization.

  Ant Group · Oct 2026

• Chinese Techcs.CV

  Tracking world states improves coherence in generated video stories

  StoryEngine uses a state-grounded agentic framework to produce coherent and consistent long-form video stories.

  Tencent, National University of Singapore · Sept 2026

• Chinese Techcs.CV

  Confidence scores for generated video intervals enable controllable selection without external verification

  Interval-level confidence scores trained with offline verifiers and reinforcement learning improve candidate selection in generative video temporal grounding.

  Alibaba Group, Beihang University · Oct 2026

• Research Labcs.CV

  Visual agents improve by jointly evolving reusable skills and practice data

  V-Gym co-evolves procedural skills and multimodal practice data from agent execution trajectories for continual self-improvement.

  State Key Laboratory of AI Safety, Institute of Computing TechnologyChinese Academy of Sciences · Sept 2026

• Top Universitycs.LG

  Sparse verifier feedback makes broad credit assignment outperform turn targeting

  Uniform reward spreading outperforms binary and targeted credit in multi-turn agent RL; coverage, not localization, is key.

  Institute of Science Tokyo, Zhejiang University · Sept 2026

• Big Techcs.CL

  Process-based evaluation reveals where computer-use agents go wrong

  OSWorld-Pro evaluates computer-use agents on intermediate subgoals to reveal specific failure modes.

  NVIDIA · Sept 2026

• Industrycs.CL

  DolphinBench measures agent memory through tasks, cost, and latency

  DolphinBench evaluates agent memory using action-based tasks, cost/latency metrics, and verified tests.

  Mem0 · Sept 2026

• Top Universitycs.CV

  Reinforcement learning trains video AI agents to use external tools effectively

  A multimodal agent trained via multitask reinforcement learning coordinates external tools to improve video generation quality and accuracy.

  Princeton University, Stanford University · Sept 2026

• Academiccs.AIcode

  Memory lets traffic-simulation agents build reusable skills without retraining

  SimSkill, a self-evolving LLM agent with episodic, procedural, and semantic memory, improves verified success on SUMO benchmarks by up to 25 percentage points.

  Jilin University · Sept 2026

• Academiccs.AI

  Distilled repository skills improve agents conducting machine-learning research

  Distilling GitHub repositories into reusable skills boosts autonomous ML research agent performance on four benchmarks.

  Beijing Academy of Artificial Intelligence, University of Science and Technology of China · Sept 2026

• Independentcs.AI

  Persistent evidence tracking improves language-agent performance on bioinformatics benchmarks

  Bioinfoysis is a multi-agent harness that achieves 82.4% accuracy on BixBench using persistent artifact-grounded analysis runs.

  Sept 2026

• Industrycs.CL

  Closed-loop synthesis efficiently trains language models to use external tools

  ToolLoop synthesizes tool-use training data via three-stage decomposition with dynamic self-feedback, achieving 86.4% BFCL accuracy from an 11K-example 4B model.

  vivo AI Lab · Sept 2026

• Big Techcs.LG

  Exactly-once behavior shifts from models to contracts with fault type

  Exactly-once for LLM agents requires tool contracts with idempotency keys; model verification alone fails for in-flight or redelivered writes.

  Microsoft · Sept 2026

• Independentcs.SE

  Engineering agents need evidence-bound authorization before their outputs trigger action

  The Agile-V Assurance Spine authorizes engineering transitions only when current, authoritative evidence and approvals are bound to the exact artifact, policy, scope, and execution point.

  Sept 2026

• Research Labcs.SE

  Merged coding-agent performance fixes often miss their claimed gains

  Coding agents' performance fixes are often merged without actually delivering the claimed speedup.

  University of Arizona, Wuhan University · Sept 2026

• Independentcs.CR

  Hierarchical registries enable fast, trust-aware discovery across agent ecosystems

  A three-layer hierarchical architecture with composite identities enables scalable, trust-aware agent registration and discovery.

  Sept 2026

• Academiccs.AI

  Layered agent harnesses improve software over repeated autonomous development cycles

  HoH enables coding agents to continually improve software in autonomous development, achieving 52.25% average gain over standalone harnesses.

  Shanghai Artificial Intelligence Laboratory · Sept 2026

• Independentcs.AI

  Query-aware evidence forests improve long-term memory for multimodal agents

  GraphMemix organizes memory into query-aware evidence forests, improving accuracy and reducing lifecycle cost for multimodal agents.

  Aug 2026

• Independentcs.AI

  Five-level ladder maps reasoning models beyond direct human oversight

  Autonomous rewards and self-generated experience enable large reasoning models to improve beyond human supervision.

  Sept 2026

• Top Universitycs.AI

  Successful agent migrations require calibration beyond interface compatibility alone

  Agent calibration across models, jurisdictions, and scale requires preserving capabilities through three layers of information, harness, and user adaptation.

  PwC China AI Center, Tsinghua University · Sept 2026

• Top Universitycs.MA

  Decentralized agents with shared context outperform centralized orchestration

  Decentralized multi-agent systems with shared context improve test-time scaling and long-context reasoning in LLMs.

  Stanford University · June 2026