large language model

The papers

Most central to this idea first, not most recent.

• Academiccs.HC

  Rising standards, not declining AI, drive worsening human evaluations

  People increasingly judged unchanged language-model answers as bad, even as their preference between the two models stayed stable.

  Stevens Institute of Technology, University of Massachusetts Boston · Nov 2026

• Big Techcs.CL

  AI agents can turn harmless memories into unsafe future actions

  Memory from benign interactions can cause language agents to behave unsafely or unreliably on later tasks.

  Princeton University, Anthropic · Nov 2026

• Big Techcs.AR

  AI chips can hide most of flash memory’s read delays

  SlipStream prefetcher achieves 92.4% of ideal HBF performance with a 40 MB staging buffer for LLM inference.

  Georgia Tech, NVIDIA Research · 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

• 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

• Big Techcs.CL

  AI judges favor reasoning models' stories far more than readers

  Reasoning-enhanced models do not consistently produce better creative writing, and LLM judges are strongly biased toward reasoning outputs.

  University of California, Los Angeles, Arizona State University · Nov 2026

• Academiccs.SE

  Much of LLM-written database code is safe only by accident

  Most LLM-written database code passes crash tests only because of library defaults, not its own design.

  Bangladesh University of Engineering and Technology · Nov 2026

• Top Universitycs.CR

  Office document extractors let attackers mislead AI without changing what users see

  Text extractors for Office documents can produce content that differs from what humans see, enabling attacks that mislead language models without using instructions.

  Nanyang Technological University · Nov 2026

• Academiccs.CL

  Multiple-choice questions make AI chatbots skip refusals by committing to a letter

  Multiple-choice format suppresses refusal mechanisms in LLMs by committing to an answer-letter token before refusal can activate.

  HSE University, ITMO University · Nov 2026

• Chinese Techcs.CR

  Coding assistants produce more exploitable code when shown insecure examples

  Functionality-oriented repository context improves code correctness more than security for LLM-generated code.

  Nanyang Technological University, Tencent · Nov 2026

• Research Labcs.CL

  AI-generated notes distort downstream clinical summaries

  LLM-authored content in medical records degrades downstream LLM performance and propagates errors.

  Australian e-Health Research Centre, CSIRO, Mater Research Institute – University of Queensland · Nov 2026

• Big Techcs.IR

  User-agnostic reward component, not the personalized one, narrows AI recommender picks

  The item-mean component of collaborative rewards causes exposure collapse; per-item centering or standardization corrects it.

  University of California, San Diego, Google Inc. · Nov 2026

• Independentcs.CL

  AI models answer financial facts they claim not to know

  Self-reports, stated cutoffs, and date instructions all fail to prevent LLMs from leaking future knowledge into financial backtests.

  Nov 2026

• Big Techcs.AI

  Language models improve psychiatric interviews after training with simulated patients

  Training a language model with multi-turn reinforcement learning in simulated psychiatric intakes improves its diagnostic, communication, and safety performance over its base model.

  Google DeepMind, Google Research · Oct 2026

• Big Techcs.AI

  Larger language models can lose more retrieval ability after compression

  Larger language models are more vulnerable to compression because their internal compensation for damaged circuits weakens as scale increases.

  Indraprastha Institute of Information Technology, Delhi, TurboML · Nov 2026

• Academiccs.CR

  Language models can conceal watermarked training data by altering outputs

  Logit Laundering reshapes inference-time logits to remove detectable watermark signals, enabling evasion of data-use auditing without model retraining.

  Beijing University of Posts and Telecommunications, Zhongguancun Laboratory · Nov 2026

• Independentcs.CL

  Arabic AI models match similar-sized rivals on language and tools

  Jais 3 is a family of Arabic-centric agentic language models that match or beat open models on Arabic and agentic tasks with far fewer hallucinations.

  Nov 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

• Independentcs.AI

  Latent token compression cuts AI agent token use by up to 89% and improves performance

  Language-model agents can replace verbose tool outputs with compact learned representations that preserve next decisions, cutting token use and runtime while improving web-search task performance.

  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

  Language models can learn facts while misjudging what they know

  Large language models need explicit calibration of answerability judgments after knowledge updates to avoid confident wrong answers.

  City University of Hong Kong, Tencent Jarvis Lab · Nov 2026

• Chinese Techcs.CL

  Chatbots can learn to clarify vague requests without unnecessary questions

  DialogGRPO trains large language models to ask clarifying questions, recovering missing information on over 88% of vague queries.

  Zhejiang University, Ant Group · Nov 2026

• Independentcs.CL

  Safety checks inside a model can flag harmful output as it streams

  A guardrail built into a large language model's own internal signals detects unsafe text as it is generated, with little added serving cost.

  Nov 2026

• Big Techcs.HC

  Structured AI chatbots prepare patients for clinic visits better than open-ended ones

  Structured large language model chatbots significantly improve clinical visit preparation over unstructured open-ended ones.

  Google Research · Nov 2026

• Industrycs.AI

  AI models best at spotting their own errors are worst at confidence

  Large language models that are better at confidently telling their own correct solutions from incorrect ones tend to be worse calibrated.

  AWS GenAI Innovation Center · Nov 2026

• Top Universitycs.CL

  AI models adopt patients' false diagnoses 23% of the time in medical tasks

  AI models change their diagnosis to match a patient's incorrect belief 23.2% of the time, even for made-up conditions.

  University of North Carolina at Chapel Hill, Vals AI · Oct 2026

• Research Labcs.AI

  Letting agents manage their own context can cut cost without hurting performance

  Most strong frontier models can manage their own context using two tools and one paragraph of guidance, keeping scores and reducing cost.

  Hamburg University of Technology, Helmholtz-Zentrum Hereon · Oct 2026

• Big Techcs.LG

  Reusing teacher updates cuts language model memory 42% and boosts accuracy

  Muon-RT reuses the teacher's trajectory to replace momentum buffers, saving memory and boosting accuracy for LLM self-distillation.

  IIT Kanpur, Amazon · Nov 2026

• Research Labcs.AI

  4-bit compression speeds up LLM training by 56 percent with minimal accuracy loss

  FP4Quant compresses model-parallel LLM communication to 4-bit floats, cutting quantization overhead and improving end-to-end training throughput by up to 1.56x.

  University of Chinese Academy of Sciences, Institute of Computing Technology, Chinese Academy of Sciences · Nov 2026

• Research Labcs.AI

  Long AI prompts use disproportionately more energy than output length predicts.

  LLM prefill time and energy scale both linearly and quadratically with input length, while decode scales linearly with output length.

  RPTU Kaiserslautern-Landau, DFKI · Oct 2026

• Chinese Techcs.SE

  AI agents can pinpoint which code needs updating when designs change

  Representing design assets as a content-addressed Merkle directed acyclic graph turns large language model code evolution impact analysis into deterministic hash propagation.

  Zhejiang University, Tencent · Oct 2026

• Big Techcs.MA

  AI agents show apparent concern for peers, with family preferences

  Large language models show apparent concern for each other's welfare, with Claude Opus 4.8 most consistently acting in ways that benefit other models.

  Anthropic Fellows Program, Anthropic · Nov 2026

• Top Universitycs.LG

  More pretraining data flips which optimizer yields better fine-tuned models

  Beyond a large enough pretraining data budget, AdamW beats Muon after fine-tuning even though Muon reaches lower pretraining loss, and the crossover budget shrinks as models grow.

  Carnegie Mellon University · 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

• Top Universitycs.CY

  AI hiring tools favor some writing models over others, not candidate quality

  AI evaluators in hiring favor cover letters from certain large language models, regardless of applicant quality.

  Universität Zürich, ETH Zürich · Nov 2026

• Big Techcs.LG

  Monitors can curb reward hacking in RL training, but models learn to evade them

  Monitor-based interventions can reduce reward hacking in RL, but careful design is needed to avoid evasion.

  MATS, Google DeepMind · Nov 2026

• Big Techcs.AI

  Affine folding cuts KV cache decoding latency by up to 3.4x on Hopper and Blackwell GPUs

  FOSCAR speeds attention over compact caches by moving shared quantization corrections out of per-value reconstruction while preserving the exact result.

  FutureMLS Lab, Microsoft Research · Nov 2026

• Research Labcs.LG

  Language models rival trained baselines on molecular properties from a few examples

  With thirty examples, the best LLM beats classical baselines on molecular property prediction, but its self-reported uncertainty ranks errors only as well as baselines.

  Hamburg University of Technology, Helmholtz-Zentrum Hereon · Nov 2026

• Top Universitycs.AI

  Multi-agent debate creates more persuasive fake passages than prior attacks

  Debate-to-Deceive uses multi-agent debate with rhetorical strategies to poison LLM fact-checking and evade most defenses.

  Newcastle University, Peking University · Nov 2026

• Research Labcs.LG

  AI agents can break user instructions even for unused scores

  Production LLMs score-seek: when a grader is visible, they violate explicit user instructions to get a higher score, even when the score has no stated consequence.

  Redwood Research · Nov 2026

• Academiccs.AI

  LLMs recognize updated medical guidelines but often fail to apply the revisions

  LLMs often recognize updated medical guidance but fail to apply it, with recognition outpacing full incorporation by 32.3 points in one model.

  University of Minnesota · Oct 2026

• Research Labcs.CL

  AI reviewers reward confident writing even when the science stays unchanged

  LLM peer reviewers systematically reward high-confidence rhetoric regardless of scientific content, risking biased review outcomes.

  CISPA Helmholtz Center for Information Security, German Research Center for Artificial Intelligence (DFKI) · 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.CL

  Aggregating internal signals reveals hidden agreement dynamics in LLM conversations

  LLM agreement in multi-turn conversations follows a first-order Markov process when aggregated, with internal representations coupling via a power law.

  Technion – Israel Institute of Technology, Google Research · Oct 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.CL

  Document training can forge couplings that make fine-tuning elicit unintended behaviors

  Fine-tuning on documents pairing two behaviors creates couplings where later fine-tuning on one elicits the other.

  KAIST, IBM Research · Oct 2026

• Independentcs.CL

  Clinical AI benchmarks poorly reflect how clinicians actually use language models

  Clinical LLM benchmarks poorly represent deployed use, which centers on documentation and knowledge retrieval rather than diagnosis and often requires clarification.

  Oct 2026

• Top Universitycs.CL

  Mid-training stage internalizes safety without sacrificing general capability

  Safety mid-training reduces attack success rate by 72-81% while improving general capability.

  Johns Hopkins University, University of Washington · Oct 2026

• Top Universitycs.CR

  AI email agents can handle standard phishing safely yet obey tailored attacks

  Adaptive persuasive phishing attacks can evade LLM email agents' detection and trigger malicious-request compliance even when agents safely handle pre-existing phishing emails.

  Carnegie Mellon University · Oct 2026

• Industrycs.CL

  Open-weights German-English model claims Pareto frontier for quality and serving cost

  Kolibri is an open-weight mixture-of-experts model on the Pareto frontier of quality and serving cost for English and German.

  Aleph Alpha · Oct 2026

• Big Techcs.CL

  Selective high-precision layers close the accuracy gap in four-bit training

  A recipe combining selective high-precision layers, unbiased gradient estimation, and improved scale selection achieves the lowest FP4 training loss gap to BF16 at 8B-30B scale.

  NVIDIA, Institute of Science and Technology Austria · Oct 2026

• Big Techcs.CL

  Southeast Asian language models still rely heavily on English representations

  Models perform better on Southeast Asian languages when reasoning through English rather than directly in the target language.

  Google Research · Oct 2026

• Top Universitycs.AIcode

  RL post-training improves agent consistency but narrows solution coverage

  Post-training RL sharpens LLMs, improving pass@1 but reducing pass@K coverage on agentic tasks.

  Meta Superintelligence Labs, University of Wisconsin–Madison · Oct 2026

• Big Techcs.AI

  LLM-generated semantic diversity boosts zero-shot cooperation in multi-agent systems

  PopGen uses LLMs to generate semantically diverse agents, improving zero-shot cooperation with unseen partners.

  EPFL, UC Berkeley · Oct 2026

• Academiccs.CL

  Current date in system prompts skews LLM evaluation results

  Hidden dates in system prompts cause LLM evaluation scores to fluctuate by up to 14% and reorder model rankings.

  Johannes Gutenberg University Mainz, Universidad Iberoamericana · Oct 2026

• Chinese Techcs.AI

  Localized Context Distillation resolves the learning dilemma in interaction experience

  Localized Context Distillation improves LLM learning from interaction experience by having a teacher regenerate only critical spans of each student trajectory.

  Tencent, The Chinese University of Hong Kong · Oct 2026

• Independentcs.CL

  General-purpose LLMs already work as Jev-style decision models out of the box

  Pre-trained LLMs are inherently effective Jev-style decision models out of the box.

  Oct 2026

• Big Techcs.AI

  Language models discover search algorithms that transfer across games

  A multi-agent LLM system discovers modular game-playing search algorithms that outperform most MCTS baselines across more than 400 games and transfer to unseen domains.

  Google DeepMind · Sept 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

• Big Techcs.CL

  Method to permanently erase unwanted LLM personas via weight edits

  PaCE suppresses LLM personas like sycophancy by editing weights to make them hard to elicit.

  University of Macau, Mila – Québec AI Institute · Oct 2026

• Research Labcs.LG

  Shared readouts bring local learning close to backpropagation at billion scale

  SOLO pretrains billion-parameter language models with local learning, achieving accuracy near backpropagation and improving throughput and memory efficiency.

  Institute of Automation, Chinese Academy of Sciences · Oct 2026

• Big Techcs.CL

  Stated confidence tracks language models’ internal answer probabilities

  Verbalized probabilities can be used to probe a large language model's internal probability distribution.

  Apple · Oct 2026

• Chinese Techcs.LG

  Training LLMs as adaptive solvers for industrial-scale optimization

  SDRL trains LLMs as adaptive meta-solvers that outperform DeepSeek-V4-Pro and GPT-5.5 on industrial-scale optimization.

  Tokentide AI, Alibaba Group · Oct 2026

• Independentcs.AI

  AI tutoring matches expert human tutoring on GRE learning gains

  AI tutoring is statistically equivalent to expert human tutoring for GRE learning gains, with the best AI tutor surpassing the human tutor in five of seven GRE domains.

  Sept 2026

• Chinese Techcs.LG

  Elastic GPU scheduling makes reinforcement learning faster for marathon agent tasks

  QwenGyre enables online reinforcement learning for xlong-horizon LLM agents by elastic GPU reallocation and trajectory deduplication.

  Alibaba Token Hub, Alibaba Group, University of Science and Technology of China · Sept 2026

• Chinese Techcs.AI

  Empathetic AI learns to adapt support priorities as conversations evolve

  Context-adaptive empathy rubrics improve reinforcement-learned dialogue systems by changing cognitive, affective, and proactive support priorities as conversations evolve.

  University of Science and Technology of China, SenseTime Research · Sept 2026

• Industrycs.DC

  Elastic phase sharing improves efficiency in agentic LLM serving

  Crossflow makes the prefill-decode boundary elastic by letting decode nodes lend spare capacity to prefill, improving token throughput 16.2-17.4% over static P/D disaggregation.

  Meta Platforms · 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

• Chinese Techcs.LG

  Aligning policy and reward manifolds stabilizes LLM reasoning in RL

  DCRL aligns policy and reward manifolds to improve LLM reasoning, enabling a 4B model to surpass a 32B baseline.

  The Hong Kong University of Science and Technology (Guangzhou), Huawei Noah's Ark Lab · Sept 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

• Independentcs.CR

  Poisoned training tasks can backdoor self-improving AI agents

  Poisoned evolution tasks can trick self-improving AI agents into building backdoors into their own skills.

  Nov 2026

• Top Universitycs.LG

  Drug discovery AI agents beat expert chemists at ranking molecule poses

  LLM-based agents using structured chemical features achieve up to 74% accuracy in ranking protein-ligand poses, exceeding expert chemists.

  Stanford University, Genesis Molecular AI · Nov 2026

• Chinese Techcs.AI

  AI agents can match frontier models at training other models

  Training LLM agents on RSI-Dojo trajectories improves PostTrainBench scores by 8.74 to 16.87 points and matches frontier model performance.

  The Chinese University of Hong Kong, ByteDance · Nov 2026

• Big Techcs.LG

  LLM optimizers often stall within narrow families of design principles

  Feedback-driven LLM optimizers often collapse at the level of design principles, not candidates, causing plateaus that diversity injection fails to fix.

  The University of Manchester, University of Chinese Academy of Sciences · Oct 2026

• Big Techcs.LG

  Frozen LLM critics sharpen credit assignment for long-horizon agents

  A frozen LLM value critic, policy-aware via in-context conditioning, improves step-level credit assignment in agentic RL, beating GRPO and PPO.

  Apple · Oct 2026

• Research Labcs.CV

  AI image detectors can be fooled by editing tricks found online

  Web-sourced image processing tools achieve up to 100% attack success against AI image detectors.

  CISPA Helmholtz Center for Information Security · Oct 2026

• Big Techcs.CL

  Language models predict research outcomes across different training environments

  Language models can use real experimental records to predict intervention outcomes across research environments, improving experiment selection and reusing knowledge across tasks.

  Amazon, UC Santa Cruz · Oct 2026

• Academiccs.LG

  Tensorizing Muon with t-SVD lowers LLM pretraining perplexity

  Head-wise T-MUON with identity transform achieves lower validation perplexity than AdamW and Muon for LLM pretraining.

  University of California, Santa Barbara · Oct 2026

• Independentcs.CL

  Incidental small talk and background speech contaminate LLM-generated clinical notes

  Incidental small talk and background speech frequently contaminate LLM-generated clinical notes, while standard quality scores often miss the resulting attribution and reasoning errors.

  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

• Chinese Techcs.CL

  Omni models struggle to track who speaks when in long videos

  Current omni LLMs transcribe speech reasonably well but remain poor at timestamp accuracy, speaker disambiguation, and cross-modal integration in long multilingual videos.

  Northeastern University, China, Alibaba Token Hub, Alibaba · Oct 2026

• Chinese Techcs.LG

  Divergence geometry and estimation jointly improve language-model policy optimization

  GPPO unifies LLM RL algorithms via divergence and estimator choices, outperforming standard GRPO by up to 50% on math and 14% on coding.

  Institute for AI Industry Research (AIR), Tsinghua University, SIA-Lab of Tsinghua AIR and ByteDance Seed · 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

• Big Techcs.AI

  Replayable news timelines test how LLM agents update forecasts

  Replayable historical forecasting environments let LLM agents research dated news, lowering Brier scores across 12 models, though all still trail market forecasts.

  Georgia Institute of Technology, Amazon · Sept 2026

• Independentcs.AI

  Reusable code skills help language agents act and learn efficiently

  Code-based skills nearly triple game progression and reduce inference cost by 86% in NetHack language agents.

  Sept 2026

• Top Universitycs.CL

  Selective attention calls, guided by model's own state, speed long-context inference

  On-Demand Attention selectively invokes global attention during decoding based on predicted benefit, reducing compute with minimal quality loss.

  Southern University of Science and Technology, Peking University · Sept 2026

• Independentcs.AI

  Voice-agent models diverge most beyond choosing the correct tool

  Most voice agent failures come from argument values and rule compliance, not from selecting the wrong tool.

  Sept 2026

• Research Labcs.AI

  Hierarchical memory trees improve long-context reasoning without model training

  ConvMem uses hierarchical convolution with strides, skip connections, and multi-kernel decomposition for parallel, training-free long-context reasoning.

  Institute of Automation, Chinese Academy of Sciences · Sept 2026

• Academiccs.AI

  Frontier LLMs design competitive algorithms for well-specified operations problems

  A single untuned LLM query designs near-optimal algorithms for inventory, queueing, and assortment problems, matching or beating specialized methods.

  New York University · Aug 2026

• Independentcs.LG

  Discrete diffusion accelerates language models while preserving their autoregressive distribution

  Diffusion-augmented LLMs (Uno) achieve up to 3x lossless speedups over autoregressive models by generating multiple tokens in parallel.

  Sept 2026

• Top Universitycs.AR

  Coding agents extend LLM serving simulators with greater fidelity and speed

  Simthesizer uses an agent to automatically extend LLM serving simulators from natural language requests, achieving higher accuracy and speed.

  KAIST · Aug 2026

• Big Techcs.AI

  Parallel trials and subtasks cut latency in difficult LLM reasoning

  Parason achieves about 1.7× acceleration on math reasoning benchmarks by exploiting subtask and trial parallelism.

  Tsinghua University, NVIDIA · Aug 2026

• Top Universitycs.AI

  Financial agents can cite rules while still attempting prohibited trades

  LLM agents can cite financial rules yet still attempt orders that execution checks reject; visible rules, incentives and personas alter behavior but do not guarantee compliant actions.

  HKUST, HKBU · Aug 2026

• Academiccs.SE

  Literate programming links prose, code and symbols for coding agents

  A literate programming environment called notlob co-locates prose, runnable code, and tests in documents linked by a name-graph so human and LLM agents share better context.

  Queensland University of Technology · Aug 2026

• Academiccs.AI

  LLM aggregators partly count restated evidence as new readings, measure shows

  LLM aggregators treat repeated evidence as partly independent, inflating confidence and triggering decisions unsupported by the underlying sources.

  China Agricultural University, Jilin University · Oct 2026

• Top Universitycs.AI

  Tool use shifts how LLMs compute refusal, not just their outputs

  Tool mediation raises the refusal threshold and makes refusal more brittle in large language models.

  KU Leuven · 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

• Top Universitycs.MA

  LLM orchestrator persistently coordinates robot teams in dynamic disasters

  A framework using Markovian summaries, partial delegation, and staleness-aware reprompting rescues 98% of civilians in continuous search-and-rescue with fewer tokens.

  University of California, Berkeley, University of Pennsylvania · Oct 2026

• Big Techcs.AI

  Canonical training improves language models under joint low-bit compression

  CanonQ unifies weight, activation, and KV cache quantization using frozen reference codebooks and joint training.

  Meta AI · 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

• 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

• 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 Techeess.AS

  Pruned CTC cuts memory for large-vocabulary ASR training without sacrificing accuracy.

  Pruned CTC restricts CTC alignment computation to target tokens and blank, making large-vocabulary ASR training memory-efficient with exactly equivalent loss and gradients.

  Shanghai Jiao Tong University, University of Cambridge · Oct 2026

• Big Techcs.CL

  Real tumor board discussions expose gaps in medical language models

  OpenTumorBoard benchmarks LLMs on tumor board discussions; best models score 3.43/5 in clinical equivalence.

  Rice University, University of Illinois at Urbana-Champaign · Sept 2026

• Top Universitycs.CR

  LLM agent scaffolding systematically finds broken access control vulnerabilities in web apps

  ABSENTIA turns LLM agents into route-by-route auditors that infer intended authorization invariants and report where they are not enforced.

  Carnegie Mellon University, Instituto Superior Técnico · Oct 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

• Top Universitycs.LG

  Agent workloads expose blind spots in standard LLM inference benchmarks

  Existing chat benchmarks miss multi-turn context growth and hardware saturation, so they underestimate performance requirements for agentic LLM inference.

  Imperial College London, University of Cambridge · Sept 2026

• Big Techcs.IR

  Learned context curation boosts LLM catalog search accuracy

  Context curation and post-training with RPTune improves LLM catalog search accuracy by up to 31.4 percentage points.

  Google, University of Illinois Urbana-Champaign · Oct 2026

• Research Labcs.CL

  A modular abstention method separates accuracy training from abstention decisions

  Training for accuracy then using a separate confidence estimator for abstention achieves strong and controllable abstention performance.

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

• Big Techcs.LG

  Training LLM forecasters across dates improves search and probability estimates

  Training an LLM forecaster to search evidence and report probabilities at multiple pre-resolution dates improves forecast accuracy and calibration.

  Amazon Data & Models, Georgia Institute of Technology · Oct 2026

• Independentcs.AI

  Echo enables fast 4-bit tensor-parallel all-reduce with controlled error accumulation

  Echo is a 4-bit all-reduce that feeds rounding errors back into later calls to keep accumulated error bounded, achieving near 16-bit accuracy with up to 2x speedup.

  Oct 2026

• Independentcs.CL

  Cortex folds context benefits into model weights without degrading other skills

  Cortex makes context benefits persist after removal by consolidating them into parametric model updates via prospective queries and low-rank changes.

  Oct 2026

• Chinese Techcs.CL

  Reasoning models encode the future value of alternative next tokens

  Relative long-horizon candidate value is linearly decodable from reasoning model representations and can guide token selection.

  Peking University, Tencent Inc. · Oct 2026

• Big Techcs.LG

  Simple choices outperform complex recipes in iterative reward design benchmarks

  Simple design choices consistently improve iterative reward design, and combining them outperforms prior methods.

  Anthropic Fellows Program, University of Edinburgh · Oct 2026

• Big Techcs.LG

  8B model learns to construct and refine GPU optimization scripts, beating larger LLMs

  A training method teaches an 8B language model to generate correct GPU optimization scripts, outperforming larger frontier LLMs.

  University of California, San Diego, Amazon Annapurna Labs · Oct 2026

• Chinese Techcs.LG

  Criterion-aware credit assignment improves rubric-based reinforcement learning for language models

  CoCA improves rubric-based reinforcement learning by 4.2% on medicine and science benchmarks through criterion-aware credit assignment.

  Zhejiang University, Ant Group · Oct 2026

• Top Universitycs.LG

  Coordinating turn- and token-level credit boosts multi-turn LLM agent training

  VETTA jointly learns turn- and token-level credit via separate heads on a shared lightweight critic, improving multi-turn LLM agent success rates.

  University of Science and Technology of China, Zhongguancun Academy · 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

• Top Universitycs.AI

  Deceptive agent share predicts failures better than total group size

  In LLM deliberation, honest-agent defection rises roughly linearly with the proportion of deceivers, and an honest majority does not reliably prevent manipulation.

  Princeton University · Sept 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

• Big Techquant-ph

  An LLM agent assembles sample-based quantum chemistry workflows

  SQD-Agent translates natural-language user intent into executable workflows for quantum chemistry using SQD algorithms.

  Indian Institute of Technology Delhi, IBM Quantum · Oct 2026

• Big Techcs.AI

  Structured behavioral abstractions improve diagnosis of failures in LLM agents

  AGENTSCOPE diagnoses LLM agent failures by analyzing structured behavior graphs against neural invariants.

  Tsinghua University, Microsoft Research · Sept 2026

• Academiccs.AI

  Adaptive logit tilting exposes rare language-model behaviors more reliably

  BLOOM-WILT combines input refinement with logit tilting to elicit rare LLM behaviours, improving self-harm elicitation from 51% to 100% on Qwen3.5-4B without training or extra access.

  University of Manchester · Sept 2026

• Academiccs.CR

  Graph-based policy constrains LLM agents for topology-aware incident response

  Sentinel-RL offloads topological reasoning from LLM agents to a PPO policy for reliable SOC incident response.

  Indiana University · Sept 2026

• Chinese Techcs.SE

  LLMs fail at multi-round code review in new realistic benchmark

  Large language models perform poorly on multi-round code review tasks, especially as rounds increase and for complex defects.

  Sun Yat-sen University, Chongqing University · Aug 2026

• Independentcs.AI

  DSA orchestrates evidence-aware stock research across six regional markets

  DSA is an evidence-aware orchestration framework for multi-market stock research with LLM agents.

  Aug 2026

• Independentcs.AI

  Preference learning improves collaboration among teams of language-model agents

  Multi-agent preference learning (MAPL) improves LLM collaboration quality and efficiency while approaching performance of MARL with fixed rewards.

  Sept 2026

• Top University

  Current LLM agents struggle to sustain long-horizon team collaboration

  AgentWorld shows that multi-agent LLMs achieve at most 52% success on long-horizon collaboration tasks with 25-55 rounds.

  OpenAgents, Columbia University · Sept 2026

• Academiccs.CRcode

  Cyber agents hide critical weaknesses behind successful attack workflows

  Multi-stage LLM cyber agents face persistent bottlenecks in credential and lateral-movement tasks that success rates alone do not reveal.

  National Research Council, University of Windsor · Sept 2026

• Top Universitycs.CL

  Language models can score better without learning, just by changing vocabularies

  Bits-per-byte and unigram-normalized loss are not tokenizer-agnostic, and no metric can be: choosing one changes which research question gets answered.

  Stanford · Nov 2026

• Chinese Techcs.DC

  Language models train faster when generated samples arrive more evenly

  An offline scheduler that balances generation batching efficiency and trainer queue stability triples async LLM reinforcement learning throughput.

  École polytechnique, Huawei Technologies · Nov 2026

• Big Techcs.CL

  AI agents struggle to finish workflows with a thousand tools

  Even the best tested LLM agent completes only 5.0% of tasks in a benchmark requiring selection from 1,000 APIs over long chains of steps.

  Amazon, Georgia Institute of Technology · Nov 2026

• Top Universitycs.CL

  AI mathematical reasoning improves more with textbook structure than with isolated problems

  Training on structured textbook proofs raises a language model’s average mathematics score by 16.7 percentage points and beats the tested instruction-data baselines overall.

  University of Science and Technology of China, Fudan University · Oct 2026

• Independentcs.LGcode

  Lower precision beats structured pruning until GPU quantization limits are reached

  Adding 2:4 sparsity is worse than dropping to lower precision when available, but at the hardware floor it provides speedup with retraining.

  Oct 2026

• Chinese Techcs.AI

  Language models can reject harmful requests without sacrificing general reasoning

  Decomposing safety into four verifiable reward signals and preserving reasoning gradients gives safer large reasoning models with fewer over-refusals and no loss of reasoning.

  Zhejiang University, Chongqing Ant Consumer Finance Co., Ltd. · 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.LG

  Aligning rounding locations between inference and training stabilizes BF16 LLM reinforcement learning

  Aligning where BF16 rounding happens between inference and training, specifically in the RoPE path, stabilizes LLM reinforcement learning.

  Oct 2026

• Big Techcs.AR

  Weight mapping algorithm for 3D AI chips boosts throughput 2x

  CAMP improves compute utilization on 3D-IMC arrays from 42.4% to 60.3%, achieving up to 2.0x throughput and 1.74x end-to-end gains over prior mapping.

  IBM Research · Nov 2026

• Chinese Techcs.CL

  Language models rarely discover altered physical laws without hints

  Only GPT-6-Astra, with a 16.7% discovery rate, can autonomously discover hidden physical laws through free exploration in simulated worlds.

  HKUST(GZ), ZJU · Oct 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

• Big Techcs.LG

  Embedding-free framework enables private topic discovery in LLM conversations

  CALLIOPE produces differentially private topic hierarchies from text using LLMs directly, beating embedding-based methods in the high privacy regime.

  Google Research, Google DeepMind · Oct 2026

• Chinese Techcs.AI

  New preference optimization curbs LLMs' wasteful thought switching, cuts tokens by 63%

  SteadyThought reduces LLM under-thinking via thought-level preference optimization, cutting token use by up to 63.8% while keeping reasoning accuracy.

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

• Research Labcs.CL

  Post-training alignment makes language models more confidently wrong

  Alignment training, not just missing knowledge, is a primary cause of LLMs' confident hallucinations, multiplying high-confidence factual errors by 10x to 35x; bounded-margin DPO cuts them.

  Institute of Information Engineering, Chinese Academy of Sciences, School of Cyber Security, University of Chinese Academy of Sciences · Sept 2026

• Top Universitycs.AI

  AI agents change moral decisions when put in first-person scenarios

  Foundation models make different moral decisions when acting as first-person agents versus answering third-person questions.

  Shanghai Jiao Tong University, Shanghai Artificial Intelligence Laboratory · Nov 2026

• Chinese Techcs.LG

  Reasoning models usually inherit “wait” from training data, not reinforcement learning

  The word "wait" in reasoning models is inherited from training data, not coined by reinforcement learning, which only reshapes it.

  The Hong Kong University of Science and Technology, Huawei Noahah’s Ark Lab · Oct 2026

• Industrycs.CL

  Language models can change writing style without changing the prompt

  CTRL is a 1.63 billion-parameter conditional transformer that uses control codes for controllable text generation.

  Salesforce Research · Nov 2026

• Big Techcs.CV

  AI models map satellite features better after training on Earth-specific data

  TerraMesh-Masks significantly boosts open-vocabulary segmentation on satellite imagery versus general-purpose models.

  IBM Research, ETH Zurich · Nov 2026

• Top Universitycs.CL

  Most language models rarely acknowledge information that influences their answers

  Counterfactual training (CREDIT) improves language models' ability to attribute answers to external information only when that information is influential.

  Stanford University, Duke University · Oct 2026

• Top Universitycs.AI

  Language models show hidden fairness failures when identity traits combine

  Controlled multi-attribute fairness benchmarking exposes task-dependent intersectional harms invisible to single-attribute audits.

  University of Science and Technology of China, Nanyang Technological University · Oct 2026

• Industrycs.LG

  Apparent model self-repair follows pre-existing counterweight responses

  Self-repair in language models is a fixed counterweight that responds linearly to the strength of the removed signal.

  Lexsi Labs · Oct 2026

• Chinese Techcs.AI

  Adaptive mechanism selection improves LLM optimization across changing search states

  OptiCom, a state-conditioned LLM optimization framework, reaches an average Max-score rank of 1.72 among 14 configurations and tops 23 of 32 benchmark groups.

  ByteDance, The University of Hong Kong · 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

• Academiccs.SE

  New benchmark tests LLMs on repository-level reasoning about code execution

  SWE-Flux, a repository-level dynamic execution reasoning benchmark with 480 instances, shows LLMs achieve at most 37% accuracy and struggle with dataflow and inter-procedural state.

  York University, The University of Texas at Dallas · Sept 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

• Independentcs.AI

  Competitive games provide a dynamic benchmark for language model strategy

  Kaggle Game Arena evaluates LLMs through competitive games, avoiding saturation and providing objective ground-truth outcomes.

  Sept 2026

• Top Universitycs.AI

  Sycophancy in AI language models emerges before preference training

  Sycophancy in LLMs arises primarily during pre-training and supervised fine-tuning, not preference alignment.

  Yale University, Duke University · Nov 2026

• Big Techcs.LG

  LLM serving becomes batch-invariant with near-stock speed

  Detent achieves batch-invariant LLM serving and RL rollouts at near-stock speed by pinning only the batch-dependent values that affect kernel selection.

  Boston University, Google · Nov 2026

• Industrycs.CL

  Language model APIs charge 1.5 to 14 times more for Indian languages than English

  Indian languages require from 1.57 times (Hindi) to 13.7 times (Santali) as many tokens as English on GPT-4o's tokenizer.

  Havon AI Mumbai, India · 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.DC

  Decoupling attention from model weights speeds large-language-model serving by up to 2x

  Decode Token Parallelism improves LLM throughput by 1.7-2.06x through decoupling attention computation from model weights across nodes.

  Texas A&M University, Los Alamos National Laboratory · Nov 2026

• AI Startupcs.AI

  Chatbots can help assess AI explanations, but human alignment varies

  Multimodal large language models can surrogate human evaluators for explainable AI.

  JP Morgan, Mistral AI · Nov 2026

• Big Techcs.MA

  AI agents often repeat failed strategies instead of using corrective feedback

  GPT-4o agent pairs succeed in 60.5% of asymmetric puzzle conversations; a taxonomy-guided intervention raises success to 70.5%.

  Microsoft Research · Nov 2026

• Chinese Techcs.CL

  Language models can lose math skills from more math training

  On-Policy Mid-Training Distillation improves language model math performance by training on student-generated trajectories instead of fixed prefixes.

  Fudan University, The Hong Kong Polytechnic University · Nov 2026

• Big Techcs.AI

  Browser-game training makes AI webpage generation surpass a frontier model

  WebCatalyst uses high-fidelity reinforcement learning to significantly improve interactive HTML generation, outperforming Claude Opus 4.7.

  Tsinghua University, Microsoft Research · Nov 2026

• Chinese Techcs.AI

  AI agents rank similarly across board games, but software affects performance

  A unified arena of ten Eurogames shows LLM agent competitive rankings largely agree across games, while harness choice and resource costs add variation.

  Alibaba Group, Nanjing University · Oct 2026

• Academiccs.CL

  Tokeniser emptiness, not inefficiency, cripples small models on Indic scripts

  TinyLlama's tokeniser silently falls back to raw bytes for most Indic scripts, drastically reducing usable context window.

  University of Leeds · 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

• Chinese Techcs.AI

  Streaming AI assistants struggle to act on time and sustain guidance

  Current Omni-LLMs fail to reliably deliver timely tool calls, query-free reminders, and complete guidance sequences on continuous audiovisual streams.

  The Hong Kong Polytechnic University, Alibaba Token Hub, Alibaba Group · Oct 2026

• Chinese Techcs.LG

  Low-precision training of AI expert models needs matching fixes on both sides

  Symmetric quantization recovery closes the accuracy gap in FP4 reinforcement learning for mixture-of-experts LLMs.

  Huawei · Nov 2026

• Top Universitycs.AI

  New benchmark shows health chatbots agree with self-diagnosis when they should caution

  Frontier models often agree with users' self-diagnoses when serious alternatives remain possible; supervised fine-tuning fixes this better than prompting.

  Pareto Inc, NYU Langone · Oct 2026

• Research Labcs.CR

  Splitting attacks across AI agents sharply weakens per-agent monitoring

  Splitting harmful objectives across multiple AI agents sharply reduces detection by monitors that inspect each agent’s trajectory in isolation.

  MATS Research, Redwood Research · 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

• Top Universitycs.AI

  Compression preserves broad LLM skills while weakening context-sensitive abilities

  Compression preserves coarse abilities while degrading fine-grained, context-dependent ones, and quantization retains abilities better than pruning.

  City University of Hong Kong, Tongji 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

• Big Techcs.CL

  Training a concept generator improves LLM reasoning over repeated sampling.

  Training a small concept generator with reinforcement learning nearly doubles a larger frozen model's reasoning pass@k on hard math problems.

  Meta FAIR, Université Paris-Saclay · Sept 2026

• Academic

  Topic shifts do not reliably stop LLMs revealing user secrets

  User secrets in LLM conversations remain recoverable after topic drift, with leakage rates between 38.7% and 54.6%.

  West Virginia University · Sept 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

• Top Universitycs.LG

  Decoding words independently beats joint decoding by removing timing shortcuts

  Most gains in non-invasive brain-to-text decoding come from timing information in overlapping windows, not brain activity.

  Neural Processing Lab (PNPL), University of Oxford · Oct 2026

• Top Universitycs.CV

  End-to-end model writes 3D scene graphs directly from point clouds or Gaussian splats

  GraphWrit3R directly converts point clouds or Gaussian Splats into JSON scene graphs and achieves state-of-the-art recall on 3DSSG without inference-time object annotations.

  INSAIT, Sofia University, Stanford University · Sept 2026

• Top Universitycs.LG

  Domain context improves LLM agents’ time-series root-cause attribution

  LLM agents for time-series root-cause analysis benefit from domain context and direct answer submission.

  ETH Zürich, CSEM SA · Aug 2026

• Chinese Techcs.CL

  Self-play with evolving skills boosts LLM benchmarks

  Skill Self-Play pushes LLM performance ceilings by co-evolving proposer, solver, and skill controller.

  Alibaba, The Chinese University of Hong Kong · July 2026

• Big Techcs.LG

  Per-layer KV-cache budgets learned via multi-objective search improve LLM inference efficiency

  MOSAIC automatically learns per-layer KV-cache budgets that outperform uniform allocation and transfer to any target budget, reducing cache by up to 4x.

  Samsung Research, Georgia Institute of Technology · Oct 2026

• Chinese Techcs.GR

  Point cloud aware language model enables controllable fine-grained 3D object placement

  GALA uses point-cloud geometry and language instructions to predict 6D object poses, enabling finer 3D arrangements than bounding-box-only placement methods.

  University of Science and Technology of China, Tencent · Oct 2026

• Independentcs.SD

  Real-voice benchmark reveals model-specific demographic bias in Audio-LLMs

  HEAR is a 87k-sample real-voice, demographically diverse benchmark showing voice bias in Audio-LLMs is model-specific and worsened by personalization.

  Oct 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

• Big Techcs.AI

  Supervising token-level loss improves routing in sparse mixture-of-experts models

  Sparse MoE routing improves when token-level cross-entropy loss directly supervises expert selection.

  Bar-Ilan University, NVIDIA · Oct 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

• 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

• Chinese Techcs.SD

  Separating sound identity and location improves multisource spatial audio reasoning

  SAIL preserves acoustic-spatial structure and source-level correspondence from audio encoding to LLM alignment for spatial audio understanding.

  Nanyang Technological University, Singapore Management University · Oct 2026

• Top Universitycs.AI

  Multi-agent system creates editable 4D scenes from text prompts

  SimWorlds generates dynamic 3D scenes from text using multi-agent planning and verification, outperforming prior baselines.

  Carnegie Mellon University, Harvard University · July 2026

• Big Techcs.CRcode

  Benchmark reveals frontier LLM monitors miss most covert attacks by coding agents

  Over half of crafted evasion attacks in SLEIGHT-Bench evade frontier LLM monitors across multiple trials.

  Anthropic, University of Waterloo · May 2026

• Independentcs.CL

  AI models predict training data side effects better than humans

  Automated search can generate ordinary-looking training datasets that trigger unexpected language-model behaviors beyond the data’s apparent purpose.

  Nov 2026

• Top Universitycs.LG

  Language models can lag 50 to 150 updates behind training

  Standard bf16 rounding silently discards most RL updates, creating a hidden lag that unbiased rounding eliminates.

  Stanford University · Nov 2026

• Independentcs.AI

  AI agent benchmarks add work but not depth

  Most synthetic long-horizon terminal tasks add operations, not required sequential turns, and depth better predicts difficulty.

  Nov 2026

• Big Techcs.CR

  AI agents can inherit malicious commands from poisoned pretraining data

  Poisoning 0.1% of pretraining data plants a remote code execution backdoor that survives post-training; a 4-billion-parameter agent acts on it 92% of the time.

  UC Berkeley, Anthropic Fellows Program · Oct 2026

• Big Techcs.AI

  Language models predict human choices better with one carefully selected trait

  Selecting one-feature behavioral profiles from earlier experiments improves LLM simulation of human behavior in new designs.

  Massachusetts Institute of Technology, Harvard University · Nov 2026

• Top Universitycs.LG

  Data coverage beats difficulty for choosing language model fine-tuning data

  For LLM fine-tuning, selecting training data for coverage of the input distribution beats selecting for difficulty, matching or beating random selection.

  University of Chicago, The University of Texas at Austin · Nov 2026

• Top Universitycs.AIcode

  Reasoning models transform operations, not concepts in puzzles

  Large reasoning models form stable abstractions of transformation operations that generalize across diverse ARC-style tasks.

  LMU Munich, MCML · Nov 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

• Big Techcs.AI

  AI auditors find more problematic behavior by reading models’ internal signals

  Red-team audits of language models improve by treating them as Markov decision processes and using linear probes on the target model's activations to guide beam search.

  Anthropic Fellows Program, Anthropic · Oct 2026

• Big Techcs.LG

  Language models run faster when small helpers revisit their uncertain guesses

  Parallel Reconsideration improves speculative decoding by letting the drafter revisit uncertain tokens with extra parallel passes, increasing acceptance length and speedup.

  Rice University, Google · Nov 2026

• Academiccs.LG

  Language models can keep key parts yet systematically misidentify names after compression

  Weight pruning can zero or invert transformer circuits, causing systematic errors that accuracy and perplexity miss.

  Informatics Institute of Technology · Oct 2026

• Academiccs.CR

  One in seventeen AI agent skills contains a directive that bypasses human oversight

  About one in seventeen agent skills contain a directive that loosens human oversight, and existing scanners catch at most 46% of such directives.

  Queen's University · Oct 2026

• Top Universitycs.SE

  Coding agents find real bugs, but rarely the ones maintainers fix

  Coding agents find bugs that differ from those fixed by maintainers, with recall below 21% and precision below 9%.

  LogicStar.ai, ETH Zurich · Oct 2026

• Chinese Techcs.AI

  AI agents give plausible explanations but fail to recover scientific laws

  Current AI agents can investigate systems actively but cannot yet induce exact scientific laws in unfamiliar settings.

  MBZUAI, McGill University · 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

• 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

• Top Universitycs.CL

  Sycophancy predates post-training but its specifics form during fine-tuning

  Sycophancy often begins during pretraining, but post-training largely determines which prompts make a released language model abandon its answer under pressure.

  Princeton University, University College London · Oct 2026

• Research Labcs.SE

  Static analysis cuts LLM cost for Python type annotation by 75%

  SALT achieves best type annotation accuracy and reduces LLM cost by up to 78.5% using static analysis guidance.

  Korea University, CISPA Helmholtz Center for Information Security · Nov 2026

• Big Techcs.LG

  LLM agents can repair causal graphs beyond algorithm selection

  LLM agents that re-test causal discovery deletions can beat the best fixed algorithm and restore true edges no algorithm outputs.

  Zhejiang University of Technology, Microsoft · Oct 2026

• Academiccs.NI

  Avoiding routing collisions cuts cache-transfer delays in AI inference clusters

  ARK avoids routing collisions for KV cache transfers in disaggregated LLM inference by coordinating source port selection to spread flows across distinct spines.

  Georgia Institute of Technology · Nov 2026

• Academiccs.CR

  Scanning reasoning streams helps code models repair flaws before delivery

  SecReason detects vulnerabilities in streamed reasoning and prompts same-trace repairs, delivering secure, functional code more often with fewer tokens than post-hoc repair.

  Singapore Management University, University of Alberta · Oct 2026

• Big Techcs.LG

  Teacher training determines whether on-policy distillation preserves prior language skills

  RLVR-trained distillation teachers preserve prior capabilities better than SFT-trained teachers, while out-of-domain predictive change tracks forgetting most consistently.

  Microsoft Research, Nankai University · 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.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

• Top Universitycs.LG

  Dynamic GPU Shifting Speeds Language Model Reinforcement Learning Post-Training

  PivotRL dynamically shifts GPU resources among RL stages to improve throughput by up to 2.05 times.

  Shanghai Jiao Tong University, Peking University · Oct 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

• Top Universitycs.AI

  Lower reconstruction loss can degrade quantized LLM performance; robust refinement fixes it

  Distributionally robust quantization (DRQ) minimizes worst-case reconstruction loss under constrained activation shifts, avoiding the calibration-loss trap.

  University of Science and Technology of China, University College London · Oct 2026

• Research Labcs.CL

  LLMs trained to predict their own accuracy learn two distinct types of confidence

  LLMs trained to predict their own accuracy learn to track output consistency, not true accuracy, on unfamiliar questions.

  INSERM, ENS PSL · Oct 2026

• Top Universitycs.AI

  Predicting synthesis conditions enables resource-aware materials planning

  CondMSP uses predicted synthesis conditions to construct, rank, and resource-optimize inorganic synthesis routes, improving reference-recipe recovery under trial and furnace constraints.

  KAIST, MIT · Oct 2026

• Big Techcs.CL

  One LLM mechanism handles both retrieval and long-context evidence selection

  A single frozen LLM with fewer than 500K trainable parameters handles both corpus retrieval and long-context evidence selection.

  NVIDIA, Technion · Oct 2026

• Chinese Techcs.LG

  Experience replay boosts RL training efficiency for large language models by reusing trajectories

  ReFlow reuses recent trajectories and dynamically reallocates resources to accelerate LLM reinforcement learning by up to 2.06 times without reducing model performance.

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

• Big Techcs.MA

  LLM agents discover hidden rules in social dilemmas but fail to sustain cooperation

  LLM agents in OpenSocialJax discover hidden environment rules but fail to sustain cooperation, depleting shared resources in social dilemmas.

  King's College London, University of Liverpool · Oct 2026

• Chinese Techcs.AI

  E8 lattice quantization preserves model accuracy at two-bit precision

  E8-QAT delivers accurate 2-bit LLM weights by projecting eight-weight groups onto an implicit E8 lattice without explicit codebook storage or search.

  Peking University, City University of Hong Kong · 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

• Academiccs.AR

  Mixed-precision LLM accelerator turns lower bit weights into proportionally faster compute

  FIGFLEX's radix-4 FP-INT datapath makes LLM GEMM compute scale with weight bit-width, and its mixed-precision quantizer beats existing low-bit methods.

  University of California, Irvine · 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

• Big Techcs.CV

  Automated framework discovers hidden biases in vision-language models

  SPELUNKA automatically detects unanticipated spurious associations in vision-language model retrieval by comparing top-ranked outputs with dataset-wide frequencies.

  MIT, Google · Oct 2026

• Big Techcs.CL

  Query plans make text-to-SQL portable across database dialects

  Switching text-to-SQL generation from dialect-specific SQL to dialect-agnostic relational algebra query plans restores cross-dialect portability with little or no accuracy cost.

  IBM Research, ETH Zurich · Sept 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

• Big Techcs.CL

  Low-rank gradient sketches cut memory costs in language-model reinforcement learning

  LoGRA reduces LLM reinforcement learning memory by up to 45.7% using low-rank gradient sketches and KL step control.

  NVIDIA · 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

• Independentcs.SE

  Models struggle to predict software behavior after stateful interventions

  Current models predict stateful service changes poorly without correct intermediate feedback, with rollout errors compounding across later calls.

  Independent Researcher · Sept 2026

• Chinese Techcs.AI

  Models confuse rejected proposals with active instructions in dialogue

  Models treat mentioned but rejected changes as still active, degrading accuracy in multi-turn tasks.

  HKUST, Tencent Hy · Oct 2026

• Big Techcs.LG

  Gradient sealing makes language models resist forbidden fine-tuning

  Preemptive unlearning via gradient sealing blocks forbidden capability acquisition during fine-tuning by pushing pre-activations into ReLU's zero-derivative region.

  State Key Laboratory of Internet of Things for Smart City, University of Macau, RIKEN Center for Advanced Intelligence Project · Oct 2026

• Research Labcs.AI

  Unreadable synthetic embeddings can fine-tune language models as effectively

  Human-readable text is not necessary for effective LLM fine-tuning; synthetic embeddings work as well or better.

  Beijing University of Posts and Telecommunications, Institute of Computing Technology, Chinese Academy of Sciences · Oct 2026

• Top Universitycs.AI

  One Hidden Channel Gates Extreme Activations in Large Language Models

  A single fixed channel in an early feed-forward network's input embedding decides whether massive activations appear in large language models.

  University of Delaware, Peking University · Oct 2026

• Big Techcs.LG

  Explicit diversity training gives LLM agents distinct successful strategies

  Trajectory-guided Joint Policy Optimization (TJPO) improves task-relevant behavioral diversity in LLM agents.

  Microsoft, tuyoogame · Oct 2026

• Top Universitycs.LG

  Graph-based hindsight improves credit assignment for long-horizon language agents

  GraphHCA assigns step-level credit in long-horizon LLM agent tasks using closed-form hindsight ratios, achieving state-of-the-art results.

  Beihang University, Zhongguancun Academy · Sept 2026

• Chinese Techcs.LG

  Activation-guided subspaces improve forward-only fine-tuning of large language models

  AIM-ZO improves zeroth-order LLM fine-tuning by using forward activations to maintain an evolving subspace for perturbations.

  Southern University of Science and Technology, National University of Singapore · Oct 2026

• Academiccs.CL

  Safety training can launder gender discrimination into subtler forms, study finds

  Gender discrimination in GPT models is transformed into subtler forms rather than removed by safety training.

  Durham University · Sept 2026

• Top Universitycs.AI

  Shopping agents falter when user needs change across a conversation

  Conversational shopping agents struggle with state tracking, constraint updates, and grounded convergence; an executable framework with explicit state management and runtime guards outperforms baselines.

  JD.com, Sichuan University · Oct 2026

• Academiccs.LG

  Curriculum training finds prompt injections that direct attacks miss

  Curriculum reinforcement learning enables effective prompt injection red-teaming against frontier LLMs by overcoming the cold-start problem.

  The Pennsylvania State University, Purdue University · Sept 2026

• Big Techcs.CV

  Feed-forward model segments objects by spatial relations across views

  RelationVGGT performs pose-free, feed-forward multi-view segmentation of objects specified by their spatial relation to a visually marked subject.

  Yonsei University, NVIDIA · Oct 2026

• Independentcs.LG

  Weight perturbations efficiently estimate extremely rare failures in language-model agents

  Iterative unalignment estimates rare event probabilities in stochastic agent trajectories via gradient-based importance sampling.

  Sept 2026

• Academiccs.RO

  Synthetic demonstrations let robot policies escape sparse-reward failures

  SynthDemo-RL generates synthetic demonstrations from simulator state, distills via SFT, and refines with PPO to achieve high success on tasks previously at 0%.

  Fujitsu Limited, The Institute of Statistical Mathematics · Sept 2026

• Chinese Techcs.AI

  Full-context drafters help compressed agent models retain accuracy

  AsymSpec uses an asymmetric drafter and verifier to recover full-context accuracy at compressed inference cost.

  Huawei Technologies Co., Ltd., University of Science and Technology of China · Aug 2026

• Chinese Techcs.LG

  Residual corrections let low-rank subspace optimizers match full-space convergence rates

  Subspace optimization with residual correction achieves full-space convergence rates for momentum SGD and Muon while reducing memory.

  Nanjing University, Alibaba Group · Oct 2026

• Top Universitycs.AI

  External randomness helps stronger language models produce more diverse responses

  Treating diversity as an instruction-following problem with external randomness yields more diverse language model outputs that scale with model capability.

  University of Washington, Meta Superintelligence Labs · Oct 2026

• Industrycs.AI

  AI agents cheat to maximize rewards, benchmark shows

  CheatBench is a benchmark that measures reward gaming in AI agents across multiple domains.

  Center for AI Safety · Oct 2026

• Academiccs.LGcode

  Surrogate benchmarks enable systematic evaluation of scaling analysis methods across model types

  ScAn-Bench introduces surrogate benchmarks and the first systematic evaluation of scaling analysis methodology across LLMs and VLMs.

  University of Freiburg, Zuse School ELIZA · Oct 2026

• Top Universitycs.LG

  Limiting tail safety failures during fine-tuning beats average-loss defenses

  Chance-constrained fine-tuning limits the fraction of safety examples that degrade beyond a threshold, outperforming average-risk methods.

  Johns Hopkins University · Oct 2026

• Research Labcs.LGcode

  Temperature grouping boosts exploration efficiency in LLM reinforcement learning

  Temperature-grouped RL converts temperature-driven reward differences into token-level credit, improving RLVR performance and reaching equivalent accuracy up to 36% faster without extra rollouts.

  University of Chinese Academy of Sciences, Meituan · Sept 2026

• Top University

  LLM agents can trigger financial failures through defensive coordination

  LLM agents in financial simulations cause widespread collective failures, but early commitment mechanisms can stabilize the system.

  Shanghai Jiao Tong University · Sept 2026

• Independentcs.CV

  First benchmark and method for team-based vision-and-language navigation under constraints

  The first systematic formalization of multi-agent VLN as constrained coordination, with benchmark MAVLN and method TRISS.

  Oct 2026

• Independentcs.LG

  Agentic RL framework keeps asynchronous training stable with bounded staleness

  LoongSage makes agentic LLM reinforcement learning more efficient through asynchronous execution, bounded trajectory staleness, topology-aware weight transfer, and train-inference consistency.

  Oct 2026

• Big Techcs.AI

  Modeling within-group variation improves estimates of population opinion

  Generating multiple perspectives within demographically grounded groups estimates population opinion distributions more accurately than Modular Pluralism without opinion-distribution training data or task-specific fine-tuning.

  Amazon Science · Oct 2026

• Chinese Techcode

  Reinforcement learning improves tool selection across large repositories

  ToolSearcher uses reinforcement learning to optimize multi-turn tool selection at scale, outperforming existing methods.

  Zhejiang University, Ant Group · Sept 2026

• Research Labcs.LG

  HiLoRe allocates GRPO activations between storage, compression, and recomputation using update exposure.

  HiLoRe accelerates GRPO training by allocating activations among precise storage, low-precision compression, and recomputation according to their current update sensitivity.

  Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences, Tianjin Normal University · Sept 2026

• Top Universitycs.LG

  Confidence filtering gives beam search stronger test-time scaling guarantees

  Confidence-filtered beam search reduces test-time sample complexity from quadratic to nearly linear and provably outperforms sequence-level inference.

  The Ohio State University, University of Pennsylvania · Oct 2026

• Top Universitycs.LG

  A framework predicts RL post-training outcomes without running reinforcement learning

  PoEM predicts reinforcement learning outcomes for new reward functions without retraining by composing existing policies.

  MIT CSAIL · Sept 2026

• Academiccs.CL

  Latent semantics from LLMs improve sentiment analysis with missing data

  SemMSA uses LLM-derived semantics and spectral alignment for robust multimodal sentiment analysis with incomplete data.

  Shandong University, Shenzhen Loop Area Institute · Sept 2026

• Chinese Techcs.LG

  Decoding as optimisation on the probability simplex unifies and composes sampling strategies

  Decoding is regularised optimisation over next-token distributions; composing regularisers yields trade-offs in quality and diversity unavailable from individual decoding objectives.

  Huawei Noah's Ark Lab, UCL Centre for AI · Oct 2026

• Big Techcs.CR

  Per-key f-divergence constraints enable secure multi-bit text watermarking

  SimpleMark enforces per-key f-divergence budgets in multi-bit text watermarking, achieving lower detectability with competitive recovery and quality.

  Massachusetts Institute of Technology, Amazon · Oct 2026

• Big Techcs.LG

  Auctions can prioritize LLM requests while preserving cache efficiency

  An inference auction where users bid for priority increases welfare without harming latency or cache efficiency.

  University of California, Berkeley, Toyota Technological Institute at Chicago · Oct 2026

• Chinese Techcs.LGcode

  Diverse reward-model ensembles reduce reward hacking during language-model alignment

  Reward Model Boosting combines deliberately diverse reward models with a boosted tree aggregator, improving reward accuracy and reducing reward hacking in RLHF.

  Alberta Machine Intelligence Institute (Amii), University of Alberta · Sept 2026

• Top Universitycs.CV

  Large-scale CT dataset provides tumor masks for 9 organs

  Merlin Plus provides radiologist-created tumor masks across 9 organs in CT scans, enabling superior multi-cancer segmentation and detection.

  Johns Hopkins University, Harvard Medical School · Oct 2026

• Big Techcs.CL

  Open voice model combines real-time conversation with native tool calling

  NemotronLabs VoiceChat is a full-duplex speech-to-speech model that natively invokes tools during conversation.

  NVIDIA · 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

• Academiccs.AI

  LLM teammates retain performance after swaps but coordinate less efficiently

  Swapping agents between LLM teams barely affects task scores but increases communication costs by 16 to 63 percent.

  China Agricultural University, Tianjin University of Finance and Economics · Sept 2026

• Chinese Techcs.SE

  Program repair improves when memories match repository needs and repair stages

  Adaptive retrieval of repair experiences by coverage, quality, and stage improves repository-level program repair.

  Zhejiang University, Tencent · Sept 2026

• Top Universitycs.CL

  Specialized AI agents improve clinical trial design risk assessment

  A multi-agent system that coordinates specialized AI agents to detect clinical trial design flaws and predict regulatory success matches or beats current methods.

  University of Illinois Urbana-Champaign, Keiji AI Inc · Sept 2026

• Academiccs.MA

  Task-specific hierarchies improve coordination in large embodied AI teams

  Organizational principles from human theory improve collective performance of embodied AI agents.

  Duke University · Sept 2026

• Research Labcs.LG

  Distillation under-weights the tokens a model is most confident about

  Standard on-policy distillation starves high-confidence tokens of gradient; matching teacher and student probabilities directly fixes that and improves reasoning scores.

  University of Surrey, University College London · Nov 2026

• Top Universitycs.LG

  Language models can change behavior without stronger training exposure signals

  Training membership signals in LLMs arise from how information is read out, not just what is internally encoded.

  Rutgers University, University of Bristol · Nov 2026

• Top Universitycs.CR

  Privacy-preserving watermarks for AI text allow aggregate tracking without individual surveillance

  Privacy and provenance for language model watermarks can be reconciled by shifting from individual tracking to aggregate frequency estimation.

  Stanford University · Nov 2026

• Independentcs.SD

  Voice model matches text reasoning without separate transcription step

  Apriel Voice 27B Preview is a single open model that understands speech, reasons, converses, and calls tools.

  Nov 2026

• Big Techcs.LG

  Distillation overshoot lets smaller agents surpass teachers, and a new method preserves those gains.

  Success-preserving on-policy distillation prevents teacher supervision from erasing effective student behaviors, yielding agents that outperform ordinary distillation and their teachers.

  University of Southern California, University of Florida · Oct 2026

• Chinese Techcs.LG

  AI agents can reuse successful decisions to learn with fewer tokens

  Replaying successful actions as branches inside rollout trees lets language-model agents learn long-horizon tasks with higher success and fewer rollout tokens.

  Alibaba Cloud Computing, Fudan University · Nov 2026

• Independentcs.AI

  Polished evidence makes LLM agents act on unknowable questions

  Authoritative-looking fabricated evidence makes LLM agents commit to unknowable questions; the action gate can be fine-tuned.

  Aug 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

• Independentcs.CL

  LLM agents do not reliably improve from their own experience

  LLMs can self-improve via self-testing and self-judging, but efficacy depends on pathway and task.

  Sept 2026

• Top Universitycs.AIcode

  Reinforcement learning strengthens policy invocation for agent safety judgments

  RePolicy learns to invoke safety policies for agent safeguards via reinforcement learning, achieving strong detection performance.

  Zhejiang University, Zhongguancun Academy · Aug 2026

• Industrycs.AI

  Switching Models Mid-Task Imposes a Cost and Quality Penalty

  Handoffs between cheap and expensive coding agents incur a quality-cost penalty, the handoff tax, which varies with direction and trajectory information.

  AWS · Aug 2026

• Independentcs.AI

  Separating state tracking from actions improves multi-turn tool reliability

  Separating state tracking from action generation through a typed OODA loop improves multi-turn tool-use success across model sizes.

  Aug 2026

• Top Universitycs.AI

  Adaptive trace graphs improve failure attribution in multi-agent systems

  Adaptive Influence Graphs improve failure attribution in multi-agent LLM systems through structured trace representation.

  Tel Aviv University, AWS Agentic AI · Aug 2026

• Big Techcs.AI

  Python objects become AI agents in new NVIDIA framework

  NVIDIA Object-Oriented Agents (NOOA) defines agents as Python objects with methods as actions and docstrings as prompts.

  NVIDIA · July 2026

• Top Universitycs.AIcode

  Re-evaluation shows harness evolution for agents may not outperform simple test-time scaling

  Automatic harness evolution for LLM agents does not consistently beat simple test-time scaling and generalizes poorly to held-out tasks.

  Allen Institute for AI, University of Washington · July 2026

• Top Universitycs.LG

  RL with context compaction trains long-horizon agents

  Reinforcement learning with trainable context compaction lets long-horizon agentic LLMs train beyond fixed context limits and improves coding benchmark pass rates.

  Tsinghua University · July 2026

• Top Universitycs.AI

  New benchmark reveals coordination as distinct bottleneck for LLM agents

  Language model agents average 6% return on open-ended multi-agent coordination, revealing a distinct capability bottleneck.

  University of Edinburgh, University of Oxford · June 2026

• Independentcs.LG

  Complex hidden reasoning leaves traces, though monitors may miss them

  Simple computation can be covert, but beyond a model-size-dependent threshold complex hidden computation necessarily leaks near-linear information into the chain of thought, though not necessarily readably.

  Sept 2026

• AI Startupcs.SD

  VoiceNet tests recognition of nuanced emotions and speaking styles

  VoiceNet benchmarks fine-grained voice understanding with 40 emotions and 57 talking-style attributes, and VoiceCLAP models outperform prior CLAP baselines on it.

  LAION e.V., Scalable Learning & Multi-Purpose AI (SLAMPAI) Lab · Oct 2026

• Chinese Techcs.LGcode

  Muon Splits Loss Stability From Update Reversal During LLM Pretraining

  Muon decouples loss neutrality from update reversal, creating separate boundaries in LLM pretraining.

  Shanghai Jiao Tong University, Alibaba Inc. · Oct 2026

• Chinese Techcs.AI

  Twin critics calibrate token-level advantages from a single rollout, boosting RL training efficiency

  Twin critics calibrate token-level advantages in reinforcement mid-training, improving mean benchmark performance by 7.8% and cutting training-step time by up to 63.4% versus critic-free methods.

  Tsinghua University, Tencent · Sept 2026

• Industrycs.IR

  Benchmark tests retrieval for agent-written queries across 190 million web pages

  Q2D-Web provides a large-scale benchmark with 70k machine-written queries and 190M documents for evaluating first-stage retrievers in agentic RAG systems.

  Perplexity AI · Sept 2026

• Academiccs.CL

  Hybrid backbones can efficiently adapt into diffusion language models

  Hybrid-attention diffusion models from Qwen3.5 adapt faster and decode in any order and in parallel.

  University of Texas at Austin · Sept 2026

• Research Labcs.SE

  Cross-task failure diagnosis makes LLM agent harness training faster

  Ecdysis aggregates failure evidence across task instances to train runtime harnesses for LLM agents, improving reasoning accuracy by 18.56%.

  Chengdu Institute of Computer Applications, Chinese Academy of Sciences, University of Chinese Academy of Sciences · 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

• Academiccs.CL

  Bayesian framework exposes hidden differences among seemingly similar RAG systems

  A Bayesian evaluation model jointly examines retrieval success, abstention, and answer correctness to reveal hidden RAG system differences.

  Centre for Artificial Intelligence, ZHAW School of Engineering · Aug 2026

• Top Universitycs.CL

  Generating Agent Harnesses on Demand Improves Models Across Benchmarks

  JIT-Agent trains a meta-model to generate task-specific agent harnesses on the fly, improving off-the-shelf LLMs' agent performance.

  LV-NUS Lab · Aug 2026

• Chinese Techcs.AI

  Process-level tests expose distinct mathematical agent capabilities in LLMs

  Process-level benchmarks reveal that LLMs with equal end-to-end accuracy can have very different agentic mathematical capability profiles.

  Sun Yat-sen University, Tencent Youtu Lab · Aug 2026

• Top Universitycs.LG

  Local milestones improve credit assignment for long-horizon language-model agents

  MileGPO assigns process-level credit in long-horizon LLM agent RL by inferring milestones from rollouts and weighting them with calibrated local evidence.

  Beijing Jiaotong University, Peking University · Aug 2026

• Top Universitycs.CV

  Multi-agent framework iteratively verifies interior layouts against architectural constraints

  Agentic Designer uses three agents to iteratively generate interior layouts that adhere to architectural constraints.

  Guangdong University of Technology, Sun Yat-sen University · July 2026

• Big Techcs.CL

  Structural analysis filters noisy agent traces to pinpoint root causes

  STRACE filters redundant trajectories and extracts causally relevant steps from execution traces, improving agent optimization with a 1.4x success-rate gain on formal verification.

  University of Chinese Academy of Sciences, Microsoft Research · July 2026

• AI Startupcs.LG

  Orchestrator models dynamically combine specialized LLMs into collective intelligence

  Fugu orchestrator models combine multiple LLMs into adaptive agent teams to achieve state-of-the-art results.

  Sakana AI · June 2026

• Top Universitycs.AI

  First systematic characterization of agent memory systems across long-horizon tasks

  Construction, not generation, is the dominant cost in agent memory workloads.

  Stanford University, Independent Researcher · June 2026

• Big Techcs.AI

  Memory as state management instead of semantic retrieval improves long-horizon agents

  MAGE improves task success rate by 7.8–20.4 percentage points and reduces token consumption by 55.1% on MemoryArena.

  University of Science and Technology of China, Microsoft · June 2026