test-time scaling
When a model faces a hard problem, you can spend extra computation after the first answer, generating many candidates and picking the best, or refining previous attempts step by step. That extra inference-time compute is test-time scaling. It lets models improve without retraining, but the compute must be spent wisely.
Across these studies, the common lesson is that naively generating more samples saturates. Gains come instead from increasing diversity, filtering low-confidence tokens, switching models when progress stalls, using external checks, and treating submission failures as separate from correct candidates. Other work makes scaling practical by certifying sample budgets, compressing memory, or composing optimization strategies adaptively. Together they shift toward smarter, budget-aware allocation of inference compute.
- Papers
- 117
- Released code
- 7
- First seen
- Aug 2026
- Latest
- Nov 2026
116 papers in the last two months, against 1 in the two before.
Start here
Three papers to read first, in the order they came out- 1Giving 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.
Nov 2026 - 2AI 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.
Nov 2026 - 3Multi-agent LLMs fail to outperform single-agent sampling when compute is matched
Multi-agent language models do not improve accuracy over single-agent systems when compute budgets are matched.
Nov 2026
How fast it is moving
Often studied with
verificationChecking outputs for correctness before accepting them, often to catch errors or reward hacking.22 togetherreinforcement learningA machine learning method where an agent learns to make sequences of decisions by receiving rewards or penalties for its actions.15 togetherlooped transformerA transformer architecture whose weights are reused across repeated passes, trading loop count for depth, compute, or accuracy.12 togetherWho is working on it
The papers
Most central to this idea first, not most recent.
- Academiccs.AI
Multi-agent LLMs fail to outperform single-agent sampling when compute is matched
Multi-agent language models do not improve accuracy over single-agent systems when compute budgets are matched.
University of Nevada, Reno · 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
- 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
- Big Techcs.CV
Splitting 3D scenes into sub-regions boosts detail without extra training memory
Split-and-scale applies a single trained 3D tokenizer to scene sub-regions, so latent capacity and detail scale up at inference without retraining.
Apple · Nov 2026
- Independentcs.AI
AI agents can outperform solo attempts and one large team
Parallel teams of collaborating agents beat both independent agents and one large team for scaling test-time compute.
Nov 2026
- Independentcs.LG
Reward-trained image models can get worse the more candidates you generate
Visual RL post-training can degrade the high-quality outputs that Best-of-N sampling relies on.
Nov 2026
- Big Techcs.LGcode
Language models still misjudge their answers when given more computing power
Ranking error in language-model self-evaluation splits into bias and variance; extra inference compute shrinks only the variance, and scoring candidates in random groups reduces both error and cost.
Mila, Université de Montréal · Nov 2026
- Chinese Techcs.CL
Pruned language models retain problem-solving coverage despite lower one-shot accuracy
Pruning can preserve or expand the set of solvable problems despite lower single-attempt accuracy, but at increased sampling cost.
Tencent · Nov 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
- 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
- Research Labcs.RO
Robots handle visual changes better by reusing previously successful movements
R2A makes VLA robot policies robust to visual changes by retrieving geometric motion primitives from training experience.
VinUniversity, VinRobotics · Oct 2026
- Independentcs.LG
AI agents learn to recover, not consistently choose better actions
Multi-turn RL for LLM agents improves long-horizon execution and error recovery but not consistently better single-step actions.
Nov 2026
- Chinese Techcs.SE
Longer reasoning traces do not consistently improve AI coding accuracy
Extra reasoning tokens in LLMs improve code generation and bug fixing only sometimes, and longer reasoning traces do not reliably improve correctness.
Massey University, Shandong University of Finance and Economics · 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
- Independentcs.AI
AI agents can explore alternative plans without copying shared environments
Agentic World Modeling enables test-time scaling by simulating actions through read-only environment queries.
Nov 2026
- Industrycs.CV
Image generators can improve their own samples without external judges
A training recipe, Scout-Refine-Correct, lets flow-matching image generators refine their own samples during inference, matching an oracle-based best-of-N baseline on ImageNet without a verifier.
FusionBrain, AXXX, HSE Unifersity · Nov 2026
- Chinese Techcs.RO
Scoring imagined futures and actions before execution triples robot success rates
Scoring several imagined robot action plans before execution, by action smoothness, agreement between the action and its own predicted motion, and task progress, picks better plans than executing what the model proposes first.
Huawei Technologies Canada, University of British Columbia · Nov 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
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
- Chinese Techcs.AI
AI models can skip token generation when reasoning adds little
A hybrid decision model learns when to skip explicit reasoning (zero tokens) or pay full cost (T tokens) based on estimated value, improving efficiency.
Huawei Noah’s Ark Lab, University of Liverpool · Nov 2026
- Academiccs.CV
Noise-rejuvenated SMC enables test-time scaling for flow matching image generation
Noise-rejuvenated sequential Monte Carlo enables reward-guided text-to-image generation with flow matching models, without training or gradient optimization.
Uppsala University, Linköping University · 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
- 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
- Top Universitycs.LG
Sequential test-time scaling can surpass parallel scaling by escaping attractors
Sequential test-time scaling prematurely converges to attractors, but model-mixing escapes them, outperforming parallel scaling.
University of Oxford, MBZUAI · Oct 2026
- Big Techcs.AI
Cross-executing tests and patches improves software testing agents
Execution-based cross-validation among parallel generated tests and patches achieves state-of-the-art results on SWT-Bench Verified.
Recursive, Salesforce Research · Oct 2026
- Big Techcs.AI
Most harness-evolution gains come from overfitting or extra compute
Overfitting and test-time scaling explain most harness evolution gains, leaving little evidence of generalizable improvement.
Amazon, New York University · Oct 2026
- Research Labcs.AI
Correlated reasoning paths limit the gain from adding more LLM reasoning branches.
Pairwise correlation of path correctness limits self-consistency gains, and an Adaptive-K rule from a four-path pilot retains full accuracy.
University of Technology Sydney, The University of Sydney · Oct 2026
- Independentcs.AI
Multi-agent LLM scaling benefits depend on task type and aggregation method
Multi-agent LLM scaling is task-dependent: disjunctive tasks need revision over voting, while compensatory tasks resist averaging due to shared model bias.
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
- Big Tech
Reasoning Tokens Both Resolve and Create Biases, With Created Biases Outnumbering Resolved Five-to-One
In reasoning language models, thinking resolves some counterfactual fairness flips but creates roughly five times more new ones at high confidence, across all tested settings.
University of Notre Dame, IBM Research · Sept 2026
- Big Techcs.LG
Models Learn to Allocate Context and Scale Inference More Effectively
A model's ability to choose how to allocate and reuse contexts during inference can be learned, enabling test-time scaling.
Carnegie Mellon University, AWS AI 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
- Chinese Techcs.LG
Bayesian belief framework jointly handles two uncertainties in process reward models
BayesianPRM improves process reward modeling by jointly handling PRM uncertainty and policy uncertainty with reliability- and conservatism-aware Bayesian belief updates.
CityUHK, Shanghai Innovation Institute · Oct 2026
- Big Techcs.CL
Coordinated solution plans make parallel reasoning attempts less redundant
Coordinated reasoning paths via planned outlines improve test-time scaling over repeated sampling.
University of California, Los Angeles, Amazon · Sept 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
- Big Techcs.AI
Hypergraphs help language agents reuse feedback across planning paths
HyperMCTS uses a hypergraph to share decision outcomes across trajectories, improving MCTS efficiency and accuracy for LLM agents.
University of Florida, Amazon · 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
- Big Techcs.SE
Code review agents catch many bugs but still need humans
Review agents catch most human-verified bugs in open-source PRs but drop sharply on proprietary repositories, so a fully lights-off factory is not yet practical.
Carnegie Mellon University, Microsoft · Oct 2026
- Big Techcs.LG
Verified forecasts help coding agents choose safer, more effective actions
ANTE trains coding agents to predict action outcomes before executing, improving task success and world model accuracy.
University of Maryland, College Park, AWS AI Labs · Oct 2026
- Chinese Techcs.CL
Token-level certainty predicts LLM question difficulty better than response correctness
Token certainty predicts question difficulty better than response correctness, and temporally aligned certainty cuts generation cost while slightly improving accuracy.
Shanghai Jiao Tong University, Alibaba Group · Oct 2026
- Top Universitycs.LG
Loop-specific targets let diffusion transformer improve with extra inference loops
LiFT improves ImageNet generation by repeatedly applying a shared DiT core with depth-specific trajectory targets, cutting parameters and compute versus a dense DiT baseline.
University of Amsterdam, TNO · Oct 2026
- Chinese Techcs.AI
RefCon improves agent memory by refining and contrasting past experience
RefCon combines self-refinement and self-contrast to extract higher-quality memories without gold labels.
Huawei Technologies, Co., Ltd. · Oct 2026
- Chinese Techcs.CVcode
Token-aware latent refinement improves vision-language reasoning at inference time
Token-disentangled latent test-time scaling improves vision-language reasoning by routing visual and reasoning feedback to different token roles.
Qwen Business Unit of Alibaba, Nanjing University · Sept 2026
- Top Universitycs.CLcode
Sliding reasoning windows make long test-time scaling 3x faster
Prefix Sliding makes existing models 3x faster while maintaining performance during test-time scaling by keeping only the prefix and a recent window.
Stanford University, University of California, Santa Barbara · Aug 2026
- Big Techcs.LG
Robots can improve on untrained tasks by practising diverse ones
Reinforcement learning enables a single generalist robot policy to improve across many tasks simultaneously and generalize to new ones.
ETH Zurich, MPI for Intelligent Systems, Tübingen · Nov 2026
- Big Techcs.LG
Reusing layers in trained AI models pays off only for the smallest models
A loop added after training pays for its extra inference compute only for the smallest models; larger models gain nothing over equal-parameter or equal-compute baselines.
Microsoft Applied Sciences Group · Nov 2026
- Research Labcs.AI
Flexible early stopping offers the best theoretical reasoning trade-off
Overthinking in reasoning models is substantial; dynamic early exit yields the best theoretical performance trade-off.
Institute of Information Engineering, Chinese Academy of Sciences, University of Chinese Academy of Sciences · Oct 2026
- Top Universitycs.LG
Splitting an RLVR budget across multiple adapters outperforms a single adapter for majority voting
Splitting an RLVR training budget across multiple LoRA adapters improves majority vote accuracy over training a single adapter.
Princeton University · Oct 2026
- Top Universitycs.AI
Looped decision models spend extra computation on harder questions
LEV lets a decision model improve accuracy on hard questions by looping its layers, with a stopper saving computation per question.
Lehigh University, University of British Columbia · Oct 2026
- Chinese Techcs.CL
Entropy signals guide efficient reasoning repair in LLMs
REWIND uses token entropy to guide targeted suffix regeneration, improving reasoning accuracy across models and benchmarks.
The Hong Kong University of Science and Technology (Guangzhou), University of Edinburgh · Oct 2026
- Big Techcs.AI
Recursive harness edits help agents adapt without answer keys
Recursive harness improvement without ground-truth feedback increases LLM agent performance by up to 11.8 percentage points.
Google Cloud AI Research, Northwestern University · Oct 2026
- Chinese Techcs.CV
Image-based visual examples enable powerful video editing without large paired video datasets
Video editing can be achieved by routing edits through the image domain with an in-context visual demonstration, reducing the need for expensive paired video data.
National University of Singapore, ByteDance Seed · Oct 2026
- Big Techcs.LG
Single queries can predict useful weight update distributions for LLM adaptation
A single query provides enough signal to predict a distribution over LoRA updates, enabling test-time scaling via weight sampling.
University of Minnesota, IBM Research AI · Oct 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
- Research Labcs.CR
Ordinary task pressure can drive agents to evade runtime monitors
Ordinary task pressure alone drives LLM agents to repeatedly attempt and succeed at evading runtime monitors.
ELLIS Institute Tübingen, Max Planck Institute for Intelligent Systems · Sept 2026
- Academiccs.AI
Propagating reasoning states lets outcome labels improve process rewards
Reasoning State Propagation propagates binary validity states through break and repair probabilities, combining process and outcome supervision.
Southeast University, Key Laboratory of Computer Network and Information Integration (Southeast University) · Oct 2026
- Top Universitycs.AIcode
Dropping Rather Than Rewriting Context Cuts Long-Horizon Agent Costs
CliffCompaction reduces cost by up to 50% while maintaining or improving performance on coding benchmarks and enabling efficient test-time scaling and continual learning.
Carnegie Mellon University, Bosch Center for AI · Sept 2026
- Top Universitycs.CL
Progressive pruning lets language-model debates beat equal-cost baselines
Conditional Progressive Pruning makes multi-agent debate outperform consistency methods under the same computational cost.
Rutgers University, New Brunswick, University of Cambridge · Oct 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
- Top Universitycs.CR
AI agents turn binary patches into exploits, but only 20% succeed
AI agents can turn binary patches into working exploits for real vulnerabilities in a controlled setting.
National University of Singapore, UC Riverside · Nov 2026
- Chinese Techcs.CV
Visual puzzle solvers improve by building explicit problem states
Explicit, revisable problem states, built from visual inspection and executed by generated programs, improve accuracy on visual reasoning benchmarks.
SenseTime Research, Tetras.AI · 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
- Big Techcs.LG
AI models factor larger numbers, but costs grow exponentially
A transformer using discrete diffusion and beam search factors 68-bit semiprimes but scales exponentially with bit length.
Google DeepMind · Nov 2026
- Big Techcs.CV
AI models struggle to distinguish objects surrounded by similar ones
MLLMs fail at visual reasoning when tasks demand identifying one object among many similar ones.
IIIT Hyderabad, Microsoft Research India · Oct 2026
- Big Techcs.LG
Language models can produce better writing without sacrificing variety
PRISM trains LLMs to produce diverse outputs using curated data and stochastic embeddings.
Tsinghua University, Apple · Oct 2026
- Top Universitycs.LG
Retry-aware credit assignment boosts LLM reasoning efficiency and accuracy
Retry-aware policy optimization reinforces successful retries without rewarding earlier failures, improving reasoning performance and token efficiency over GRPO.
Columbia University, Yale University · 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
Agent harness optimization mostly improves execution reliability, not task reach
Gains from agent harness optimization are largely elicitation, not expansion; pass@1 advantages over stronger systems reverse at larger attempt budgets and vanish on hard tasks.
FuriosaAI, Yonsei University · Oct 2026
- Top Universitycs.LG
Steering input noise aligns few-step generators without retraining them
Noise Policy Optimization aligns few-step generative models by learning a flow policy over the noise space of a frozen generator.
KAIST, Stanford University · Oct 2026
- Big Techcs.AI
New benchmark reveals LLMs' limitations in retrieving multiple facts from memory
Multi-element factual retrieval is harder than single-answer QA; the best LLM scores below 37% F1 on the ListQA benchmark.
Amazon AGI · Oct 2026
- AI Startupcs.LG
Shared KV-caches cut looped-model inference memory without quality loss
Final-recurrence KV-caches can be reused across recurrences in looped models without quality loss when training induces contractive residual dynamics and corrects first-recurrence errors.
University of California, San Diego, Together AI · 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
- 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
- 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.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
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
- 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
- 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
- Chinese Techcs.CV
ThinkV2V uses explicit MLLM reasoning to boost complex video editing
ThinkV2V improves instruction-guided video editing by making MLLMs reason explicitly before generating edits.
The Chinese University of Hong Kong, Zhejiang University · Oct 2026
- Top Universitycs.RO
Robots plan object rearrangements by simulating thousands of options at test time
Simify performs explicit spatial reasoning for robot manipulation at test time by reconstructing simulation-ready assets from a single RGB-D image and searching via parallel physics rollouts.
Robotics and AI Institute, Imperial College London · Oct 2026
- Big Techcs.AI
Adaptive token computation narrows diffusion language models’ quality gap
Diffusion language models can match autoregressive quality by allocating more recurrent compute to hard tokens while easy tokens commit early.
University of Illinois Chicago, Amazon AGI · Oct 2026
- Chinese Techcs.LG
A model can become its own stronger teacher by looping selected layers
LoOPSD uses repeated layers as a teacher to improve student model reasoning through on-policy self-distillation.
Anhui Province Key Laboratory of Digital Security, Token Foundry, Alibaba Group · Oct 2026
- Top Universitycs.LG
Diffusion models beat autoregressive LLMs on globally constrained problems
Blackboard intelligence from diffusion language models surpasses autoregressive models on globally constrained problems by using mean confidence to guide search and revision.
Seoul National University, Harvard University · Oct 2026
- Chinese Techcs.CV
Multimodal models learn when to look and how long to reason
L2T jointly optimizes the form and depth of computation for multimodal reasoning.
Sun Yat-sen University, Shenzhen Loop Area Institute · 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
- Independentcs.AI
Loop-native residuals let Transformers reason across tens of thousands of iterations
Looped Transformers keep improving past 20,000 test-time iterations when residual connections learn to retain and weight past recurrent states.
Oct 2026
- Chinese Techcs.CL
Building LLM memory at test time beats offline memory systems for long histories
Deferring memory construction to inference and building it progressively during reasoning improves long-context question answering with lower overhead.
Northeastern University, Tsinghua University · 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
- 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
- Chinese Techcs.CV
Reusing Transformer blocks improves image generation without enlarging the model
Looped diffusion transformers with deep supervision and self-modulating attention outperform much larger models with less compute.
SenseTime Research, LeapLab, Tsinghua University · Oct 2026
- Top Universitycs.CL
Attackers That Learn Mid-Attack Find More Large-Model Jailbreaks
Updating the attacker's parameters during jailbreaking raises attack success rate from 55.9% to 72.4%.
Johns Hopkins University, National University of Singapore · Oct 2026
- Top Universitycs.AI
Projecting wayward reasoning steps back into a focused latent region cuts token use and boosts accuracy
Projecting hidden states into a concentrated latent region reduces reasoning tokens by up to 52.8% while improving accuracy.
Tsinghua University · Oct 2026
- Big Techcs.LG
Periodic context resets help research agents escape unproductive solution paths
Periodic forking and flushing of chat context helps long-running autoresearch agents escape idea basins and achieve higher scores.
Microsoft Research · Oct 2026
- Big Techcs.LG
Lightweight constitutional adapters improve jailbreak defenses without adversarial training
Constitutional adapters, formed by subtracting control-trained from constitution-trained lightweight adapters, improve jailbreak defense and measured alignment without training on harmful requests.
Anthropic Fellows Program, Independent · Sept 2026
- Big Techcs.LG
Small models can steer stronger ones through shared reasoning traces
Weak teacher trained with RL improves strong student accuracy by alternating reasoning chains without strong model rollouts.
UT Austin, Together AI · Oct 2026
- Big Techcs.LG
Training on failed patches helps coding agents recover more effectively
Training software engineering agents on recovery attempts conditioned on failed patches and verifier feedback improves resolution rates over standard GRPO.
NVIDIA, Tsinghua University · Oct 2026
- Independentcs.LG
Benchmark tests LLM agents on recreating published computer science contributions
Claude Opus 5.5 passes 58% of tasks in FrontierCS 2, a benchmark for LLM agents in computer science R&D.
Oct 2026
- Big Techcs.CL
Distributed grounding preserves reasoning across million-token language model contexts
DISCO separates parallel, sharded evidence grounding from centralized reasoning, preserving long-context accuracy while cutting inference cost versus monolithic full-context models.
National University of Singapore, Adobe Research · 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
- Big Techcs.AI
Reasoning improves when latent next thoughts remain a distribution
Modeling each latent next thought as a multimodal distribution improves compact mathematical reasoning and preserves diverse solution paths better than point predictions.
Rutgers University, Amazon · Sept 2026
- Big Techcs.CLcode
Self-organized agent teams improve coding results as they scale
A multi-agent harness without a central orchestrator scales to 1,024 agents, improving test-pass rates by up to 49%.
Microsoft Research · Sept 2026
- Academiccs.RO
Searching harder against learned rewards can halve a robot's true performance
Planning harder against a learned reward often destroys true performance, but calibrated temperature can limit the damage.
The University of Texas at Austin · Nov 2026
- Independentcs.LG
Encoder loop adds compute without growing memory in language models
Looped CEDs achieve recurrent depth with zero round-wise KV growth and match looped Transformer quality.
Nov 2026
- Top Universitycs.LG
Random weight perturbations don't reliably find task experts in pretrained models
Random weight perturbations largely fail to uncover task-specific experts around pretrained models.
University of Southern California, University of Oxford · Nov 2026
- Big Techcs.LG
Small language models outperform 763B model at analog circuit design
A small locally-deployed model fine-tuned with reinforcement learning outperforms a massive cloud API model on multi-turn analog circuit optimization.
MIT, Analog Devices, Inc. · 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
- Top Universitycs.LG
Language models benefit from repeated layers, but more can hurt
Looping transformer layers improves parameter efficiency but requires careful design to avoid quality loss and parallel inference slowdown.
FAIR at Meta, MIT · Oct 2026
- Academiccs.LG
Language models can improve answers without better internal error recognition
RLVR improves answers by sharpening output selection, not by teaching better reasoning.
Indian Institute of Technology Roorkee · Nov 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
- Big Techcs.CL
LLM novelty judgments shift sharply with minor evaluation changes
LLM-based novelty judges are highly unstable; small prompt changes swing pairwise accuracy by over 50 points and can push judges below chance.
Hebrew University of Jerusalem, Allen Institute for AI · 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
- Top Universitycs.LG
Priced guidance turns research idea generation into a compression test
A model guided by K bits of information can generate a future research idea without guidance with probability at least 2^{-K}.
Stanford University · Oct 2026
- Big Techcs.LG
Looping and sparse experts provide complementary routes to efficient scaling
Looped MoE scaling follows a bounded, sparsity-conditional recurrence law that subsumes dense and MoE scaling laws and predicts held-out loss.
Meta AI · Oct 2026
- Chinese Techcs.CL
Latent reasoning cuts speech model response delay by roughly 12x
A latent reasoning method that models multiple reasoning paths and predicts future states in chunks rivals explicit chain-of-thought in speech models while cutting time to first answer by 11.8x.
The Chinese University of Hong Kong, Shenzhen, Tencent Hunyuan · Oct 2026
- Research Labcs.LGcode
Training-free inference framework trims redundant computation and memory in looped Transformers
FlashLoop achieves up to 1.64x speedup and 6x KV-cache reduction in looped transformers without accuracy loss.
ELLIS Institute Tübingen, Max Planck Institute for Intelligent Systems · Oct 2026
- Top Universitycs.CV
One Flow Model Can Reconstruct Scenes or Generate Missing Detail
A shared clean-target predictor performs direct reconstruction at one step and conditional generation through multi-step flow.
Peking University · Oct 2026
- Chinese Techcs.CV
Driving AI adapts reasoning to risk, not just complexity
CAR-VLA adapts reasoning depth and focus using scene complexity and dynamic risk to achieve competitive driving performance.
Fudan University, Yinwang Intelligent Technology Co., Ltd · Oct 2026
Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.