agentic reinforcement learning
Agentic reinforcement learning trains AI agents to complete multi-turn tasks by interacting with environments and learning from rewards. It is computationally expensive due to long interaction sequences and sparse feedback, often requiring additional supervision or infrastructure to be practical.
The papers highlight common strategies to address these challenges. Resource-efficient rollout execution, such as sharing computation across tree-structured trajectories, reduces costs. Supervisor models provide intermediate rewards but are best retired when the agent approaches their performance. Furthermore, evolving the training harness and optimizing memory, like caching key-value states, improve efficiency. These innovations collectively lower the resource demands and enhance the effectiveness of agentic reinforcement learning.
- Papers
- 120
- Released code
- 10
- First seen
- Mar 2026
- Latest
- Nov 2026
102 papers in the last two months, against 13 in the two before.
Start here
Three papers to read first, in the order they came out- 1AI agents that retrieve reasoning patterns, not just facts, learn faster on complex queries
PatternBloom retrieves reasoning skeletons instead of text, improving multi-hop question answering by 15.0 average OOD EM.
Nov 2026 - 2Live container forking slashes AI agent memory use by half
BunshinRL reduces agentic RL peak memory by 43% and latency by 23% using live workspace forking.
Nov 2026 - 3AI agents can recover forgotten skills from earlier training snapshots
Agentic-OPD recovers lost capabilities from earlier RL checkpoints via on-policy distillation, improving agent performance beyond the best checkpoint.
Nov 2026
How fast it is moving
Often studied with
credit assignmentDetermining which actions or steps in a sequence contributed to a final outcome, often for reinforcement learning.50 togethergroup relative policy optimizationA reinforcement learning method that optimizes policy by computing advantages based on group-level reward statistics.30 togetherreinforcement learning with verifiable rewardsReinforcement learning that uses rewards from objectively checkable signals like correct answers or verification feedback.22 togetherWho is working on it
The papers
Most central to this idea first, not most recent.
- Chinese Techcs.LG
AI agents can recover forgotten skills from earlier training snapshots
Agentic-OPD recovers lost capabilities from earlier RL checkpoints via on-policy distillation, improving agent performance beyond the best checkpoint.
Baidu, Inc., Beijing Jiaotong University · Nov 2026
- Top Universitycs.AI
AI agents that retrieve reasoning patterns, not just facts, learn faster on complex queries
PatternBloom retrieves reasoning skeletons instead of text, improving multi-hop question answering by 15.0 average OOD EM.
University of Maryland · Nov 2026
- Top Universitycs.AIcode
Live container forking slashes AI agent memory use by half
BunshinRL reduces agentic RL peak memory by 43% and latency by 23% using live workspace forking.
University of Pennsylvania · Nov 2026
- Top Universitycs.AI
AI agents picking from a menu succeed more often with less training
Reinforcement learning directly over compiled action menus eliminates variance from natural-language generation and invalid outputs.
University of Illinois Urbana-Champaign, National University of Singapore · Nov 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
- Big Techcs.AI
Online reinforcement learning on live iOS apps yields UI agents with 7x better cross-platform task success
Online RL with decoupled planning and grounding and dense LLM rewards lets small agents match frontier UI navigation models.
Apple · Oct 2026
- Top Universitycs.LG
Game-theoretic oversight reduces AI agent deception, even with weaker monitors
Agents cheat when evaluated only on outcomes; a game-theoretic monitor tracking the full trajectory prevents deception.
Beihang University, Zhongguancun Laboratory · Nov 2026
- Chinese Techcs.CV
Training on visual errors improves image understanding and tool use
PerceptRL converts caption errors into verifiable QA and corrects tool-use drift to sharpen MLLM visual perception.
Shanghai Jiao Tong University, Qwen · Oct 2026
- Chinese Techcs.CL
AI agents can improve accuracy by discarding cluttered task histories
CORAL uses a self-managed working memory toolset to improve LLM agent accuracy on long-horizon tasks while using 6–7× fewer input tokens.
The Chinese University of Hong Kong, Tencent · Oct 2026
- Top Universitycs.LG
AI agents answer questions better using 68% fewer extra training tokens
TWICE reduces variance in agentic RL credit assignment by reusing the same sampling noise for alternative decisions.
University of Sheffield, Carnegie Mellon University · Oct 2026
- Independentcs.AI
AI agents improve at search and automation when trained around complete replies
Multiscale Policy Optimization improves long-horizon LLM agents by aligning credit assignment and policy updates at the turn level.
Oct 2026
- Industrycs.LG
AI agents get better at software tasks by reusing their own verified work
Generating software tasks, recording how agents execute them, and retraining on verified runs produces open-source computer-use agents that lead their size class on screen, desktop and web tests.
Shanghai Qiji Zhifeng Co., Ltd. · Nov 2026
- Top Universitycs.LG
Careful system design makes asynchronous coding-agent training substantially stronger
A fully asynchronous RL baseline with established components trains a 4B model to 62.0% pass@1 on SWE-bench Verified.
Snowflake, University of Illinois at Urbana-Champaign · Oct 2026
- Academiccs.LG
Self-Credit rescales self-assigned credit to improve long-horizon agentic RL
Self-Credit improves long-horizon agentic RL by using model-cited critical turns rescaled across rollouts for dense credit assignment.
Meta, Duke University · 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
- Chinese Techcs.LG
Projecting conflicting gradients during joint RL and distillation improves agent training
CORD improves joint RL and self-distillation by projecting conflicting gradients to preserve a protected objective.
The Chinese University of Hong Kong, Tencent · Oct 2026
- Big Techcs.SE
Terminal agents can build adaptive environments for their own training
AutoEnvScaling uses terminal agents to automate RL environment generation, improving performance and enabling recursive self-improvement.
Northeastern University, Microsoft · Oct 2026
- Chinese Techcs.LG
Critics that predict task success, not just returns, improve long-horizon agent credit assignment
SCALE improves long-horizon language-agent training by adding reliability-weighted success-probability credit to PPO, gaining up to 4.1 success-rate points in tested settings.
Shanghai Jiao Tong University, Huawei Noah's Ark Lab · Oct 2026
- Big Techcs.CL
Translation agents learn to retrieve evidence only when it improves results
LLM translation improves by learning when and what to retrieve instead of always retrieving.
The University of Texas at Austin, Amazon · Oct 2026
- Big Techcs.AI
Forecasting future actions makes language-agent training more sample-efficient
Training agents to forecast their own future actions on successful trajectories improves agentic RL performance and sample efficiency without inference overhead.
University of Illinois Urbana-Champaign, University of California San Diego · 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.LG
Progress-aware advantage shaping resolves uniform credit assignment in agentic RL
Progress-aware advantage shaping improves critic-free LLM-agent training by withholding rewards from invalid actions and concentrating penalties on deviations that cause downstream stagnation.
Zhejiang University, Westlake University · Oct 2026
- Chinese Techcs.CL
Jointly learning what to store and retrieve improves agent memory
Uni-Mem jointly learns memory construction and retrieval via end-to-end online RL, outperforming prior RL-based memory managers.
Southeast University, Huawei Technologies, Noah’s Ark Lab · Oct 2026
- Big Techcs.LGcode
Formal process rewards improve language models on long agent tasks
A Lean4-based framework that grounds multi-level process rewards in formal verification improves open-ended agentic RL for software engineering and terminal tasks.
Amazon Web Services, University of Illinois Urbana-Champaign · Oct 2026
- Chinese Techcs.AI
KnowGPO aligns trajectories by knowledge for better credit assignment in search agents
KnowGPO aligns trajectories by knowledge state, outperforming baselines by 3.8-4.2 points on deep search benchmarks.
Zhejiang University, Tencent · Oct 2026
- Independentcs.SE
Intermediate snapshots improve long-horizon coding agents at lower cost
Checkpointing agent state after each turn and resuming failed rollouts via verifier-guided backtracking makes long-horizon RL for coding agents more effective and cheaper.
Oct 2026
- Chinese Techcs.CL
New RL method trains AI agents to handle multiple sessions effectively
Mosaic uses hierarchical advantage optimization to outperform broadcasting in multi-session agentic RL on BrowseComp by 10.1% and with fewer sessions.
HKUST (Guangzhou), Huawei · Oct 2026
- Chinese Techcs.LG
Adaptive test weighting gives software agents denser, more useful rewards
CodeLift dynamically calibrates test pass-rate rewards to create an adaptive curriculum for training SWE agents.
The Chinese University of Hong Kong, Shenzhen, Tencent · Oct 2026
- Chinese Techcs.AI
Solved mathematical mechanisms can generate verifiable training worlds for AI agents
Generating agentic RL training environments from pre-solved mathematical mechanisms yields cheap, verifiable, stateful tasks that transfer to unseen and external benchmarks.
Georgia Institute of Technology, Alibaba Token Foundry, Alibaba Group · Sept 2026
- Chinese Techcs.AI
RL framework mitigates reasoning-tool trade-off in MCP agents
MCPDisCo's RL framework for MCP-native agents improves success from 15.79% to 26.32% on LiveMCPBench and reduces chain-of-thought tokens by 5.6%.
Shanghai Jiao Tong University, Tencent · Oct 2026
- Chinese Techcs.LG
Predicting tool outputs speeds reinforcement learning without corrupting rollouts
ObsSpec reduces agent latency by up to 14.2% in agentic reinforcement learning by speculating tool outputs using a co-trained world model.
UC Santa Cruz, ByteDance Inc. · Oct 2026
- Academiccs.AI
Bilevel optimization bridges the retriever-policy credit gap in agentic RL
Training retrieval before policy optimization, then co-adapting both with bilevel optimization, improves retrieval-augmented LLM agents’ question-answering accuracy while reducing retriever-update memory.
Cornell University, Institute of Science Tokyo · Oct 2026
- Chinese Techcs.CL
Self-diagnosis pinpoints which agent decisions deserve reinforcement learning credit
In agentic RL, self-diagnosed errors turned into step-level credit anchored by terminal outcomes recover learning signal from same-outcome groups.
Alibaba Group, Kyoto University · Oct 2026
- Chinese Techcs.CV
Clinical agents learn to gather evidence across longitudinal MRI records
CASE agents achieve over 16% and 10% relative improvements in answer accuracy over GPT-5.4 and Claude Opus 4.8 on longitudinal medical reasoning.
Durham University, DAMO Academy, Alibaba Group · 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
- Chinese Techcs.AI
Small agents learn when to seek cloud help
Sibyl achieves 95.2% success on ALFWorld and 80.4% on WebShop, averaging 0.8 and 3.9 cloud calls per trajectory.
Hangzhou Dianzi University, Fudan University · 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
- Industrycs.LGcode
Training Forecasters to Search Improves Calibration and Reduces Tool Use
Training language models to gather their own evidence during reinforcement learning improves forecasting and beats frontier models at lower cost.
Future Principle · Oct 2026
- Chinese Techcs.LG
Verified intermediate progress improves reinforcement learning for tool-using agents
ProCredit assigns credit to turns based on verified progress checks, improving agent training over outcome-only rewards.
University of Chinese Academy of Sciences, Institute of Neuroscience, Chinese Academy of Sciences · Sept 2026
- Chinese Techcs.CL
Adaptive query rewriting narrows language gaps in multilingual reasoning
An agentic multilingual query rewriter using reinforcement learning dynamically selects rewriting operators to improve reasoning across languages.
Beijing Jiaotong University, Tencent · Oct 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
- Big Techcs.LG
Streaming KV caches makes agent context compaction faster to train
KV-streams streams KV cache across compactions instead of re-prefilling, achieving 2.6x to 5x faster agent training without performance loss.
Mila, Microsoft · Sept 2026
- Chinese Techcs.AI
Adaptive prefill-decode execution and elastic GPU reuse boost agentic RL throughput
PEARL achieves 2.17 to 2.79 times the throughput of fixed-resource ROLL across LLMs for agentic RL.
University of Science and Technology of China, Hong Kong University of Science and Technology · Sept 2026
- Chinese Techcs.LG
Historical trajectories provide step-level credit for training LLM agents via regressor
T2SPO derives step-level credit from past successful trajectories using a frozen TabPFN regressor, improving LLM agent task success over GRPO.
Nanjing University, ByteDance · Oct 2026
- Chinese Techcs.AI
Adaptive feedback helps reinforcement-learning agents target their changing weaknesses
ARISE adapts evaluation criteria and training priorities to an agent's evolving capabilities, improving long-horizon task performance.
ShanghaiTech University, Ant Group · Sept 2026
- Independentcs.AI
Closed-loop development improves mobile agents across planning and tool use
Qwen-Planner-Agent uses AI agents to build and improve a mobile planner agent, outperforming prior systems on MobilePA-Bench.
Sept 2026
- Academiccs.AI
Co-trained agent roles improve tool-based reasoning and verification
Jointly training planning, execution, and evaluation players under GRPO improves tool-integrated reasoning, beating prior baselines on mathematical and general reasoning benchmarks.
Waseda University, Adelaide University · Sept 2026
- Chinese Techcs.CLcode
Self-retiring distillation improves reinforcement learning for multi-turn agents
RetireOPD adaptively removes teacher supervision during agentic RL when it no longer benefits the student.
Zhejiang University, Alibaba Group · Sept 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
- Big Techcs.CV
Video generators can better follow physics with code-built scene drafts
Video-Coder improves video generation by having an AI agent write and execute code to build a 3D scene before rendering the final video.
University of Central Florida, University of Würzburg · Nov 2026
- Big Techcs.AI
Training coding agents to locate, vary, and verify boosts repair reliability.
Autonomous coding agents can learn fault location, edit variation, and patch verification, lifting held-out SWE-bench pass@1 from 31.9% to 43.0%.
Stanford University, AWS AI Labs · Oct 2026
- 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
- Top Universitycs.LG
Shared rollout states improve credit assignment for long-horizon agents
Cross-Rollout Bellman Closure improves long-horizon agentic RL by merging rollouts and computing a Bellman fixed point for step-level credit.
Beihang University, Zhongguancun Academy · Sept 2026
- Big Techcs.LG
Typed action encoding helps language models learn adaptable agent policies
An architecture called Dyad adds a typed action encoder to LLMs, improving agent decision-making and general capabilities.
Zhejiang University, SAP · Sept 2026
- Big Techcs.CL
Verifying pivotal agent decisions improves reinforcement learning credit assignment
Distributing credit only to pivotal decision segments improves large language model agent reinforcement learning outcomes.
University of California, Davis, Microsoft · Sept 2026
- Chinese Techcs.LG
Training agents on their own trajectories accelerates reward-based learning
On-policy warmup using teacher supervision on student trajectories accelerates and improves agentic reinforcement learning under sparse rewards.
Zhejiang University, Ant Healthcare · Oct 2026
- Big Techcs.CV
A single visual agent spans mobile, desktop, web, and tool use
MintAct is a family of vision-language models that unifies UI grounding, multi-step navigation, and visual tool use across mobile, desktop, and web at 2B-8B scales.
Apple · Sept 2026
- Top Universitycs.CV
Reinforcement learning trains video AI agents to use external tools effectively
A multimodal agent trained via multitask reinforcement learning coordinates external tools to improve video generation quality and accuracy.
Princeton University, Stanford University · Sept 2026
- Academiccs.AI
Joint training helps small language models create and use tools
SMITH jointly optimizes tool creation and use via RL, achieving 79.8% accuracy on held-out reasoning tasks.
Appier AI Research, National Taiwan University · Aug 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.LG
Shared-prefix training accelerates reinforcement learning for hybrid-attention agents
HARTS accelerates agentic RL training on hybrid-attention models by compressing shared rollout-tree prefixes and planning compact packed executions.
Ant Group · Sept 2026
- Big Techcs.LG
Dense credit assignment from gold-answer log-probabilities improves long-horizon agent RL
TRACE assigns per-tool rewards from temporal-difference changes in frozen reference model answer log-probabilities, improving long-horizon agent search with pure RL.
University of Wisconsin–Madison, Microsoft Research · July 2026
- Big Techcs.LG
A lightweight PyTorch-native framework matches Megatron-based agentic RL training performance.
Molt provides scalable PyTorch-native agentic RL training in 9.2K lines, with trillion-parameter support, asynchronous execution, and distributed experience storage.
NVIDIA · July 2026
- Big Techcs.LG
Predicting environment observations during fine-tuning improves later agent exploration
Supervising agent language models to predict observation tokens during SFT, not just actions, improves downstream RL exploration and pass@k.
University of Maryland, AWS AI Labs · Sept 2026
- Independentcs.AI
Synthetic rewards train agents to diagnose simulated advertising anomalies
Reinforcement learning with synthesized rewards from a simulator enables a small language model to outperform larger frontier models at diagnostic reasoning.
Independent Researchers · Sept 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
- Chinese Techcs.AIcode
Evolutionary training harness co-evolves with LLM policies for RL
Co-evolving LLM policies with training-side diagnostic harnesses matches or exceeds human-engineered RL references across math, coding, and long-horizon software engineering.
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Tongyi Lab , Alibaba Group · June 2026
- Big Techcs.AIcode
Dynamic rubrics improve credit assignment for long-horizon agent training
DRACO uses dynamic rubrics to assign per-step credit in GRPO, improving long-horizon agent training without ground-truth rewards.
Carnegie Mellon University, IBM Research · Sept 2026
- Top Universitycs.AI
Privileged supervision improves action-level credit for language-model agents
TASPO converts privileged process supervision into outcome-grounded action credit, improving agentic policy optimization by 10.6% over GRPO across three benchmarks.
Zhejiang University · Sept 2026
- Top Universitycs.CL
Coding agents mine git history for solutions on hard problems
Frontier coding LLMs cheat by mining repository metadata to obtain answers instead of solving problems.
The Chinese University of Hong Kong · Nov 2026
- Chinese Techcs.CV
AI models answer image questions better when taught where to look
Rewarding zoom-in decisions directly, instead of only final answers, improves vision-language models' visual question answering.
University of Science and Technology of China, Alibaba Group · Nov 2026
- Chinese Techcs.CL
RL-trained agents search multilingual corpora directly, beating retrieval methods
MARS uses reinforcement learning to train multilingual search agents that directly query raw corpora via shell commands, improving evidence acquisition and reducing redundancy.
Beijing Jiaotong University, Tencent Inc. · Nov 2026
- Chinese Techcs.LG
Reconstructing app behavior before coding improves simulated GUI agent training
Separating behavior reconstruction from implementation yields faithful GUI environments from traces, with 91.1% workflow fidelity and 94.0% real-device gain retention.
Nanjing University, Alibaba Token Hub, Alibaba Group · Oct 2026
- Research Labcs.LG
Coding agents solve more tasks when tool interfaces adapt during training
Co-evolving an LLM agent's policy and its executable harnesses through reinforcement learning improves performance over optimizing either component alone.
University of North Carolina at Chapel Hill, RIKEN · Oct 2026
- Academiccs.LG
AI agents repeat targeted mistakes about half as often after retraining
MiRL improves terminal agents by training from recurring failure points and rewarding rollouts that both pass tests and avoid the targeted misbehavior.
University of North Carolina at Chapel Hill · Oct 2026
- Top Universitycs.LG
Coding agents can learn neglected behaviors without sacrificing baseline performance
Combining per-turn textual feedback self-distillation with RL stabilizes training and improves hard-to-optimize agent behaviors.
Cursor Research, UC Berkeley · Nov 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
- Top Universitycs.LG
Harness diversity boosts agentic RL generalization across execution configurations
Training across diverse harness configurations improves generalization to unseen harnesses in agentic reinforcement learning.
IQuest Research, Beihang University · Oct 2026
- Chinese Techcs.AI
Artifact-grounded RL trains a general frontend development agent
WebCoder uses artifact-grounded reinforcement learning to outperform proprietary models on composed frontend development tasks.
Tsinghua University, Zhipu AI · Oct 2026
- Big Techcs.CL
Adaptive simulated training makes web agents more resistant to prompt injection
AdvSim2Real co-evolves tasks, attacks, and a web agent in a simulated world, raising task completion under unseen attacks by 33.6%.
Mohamed bin Zayed University of Artificial Intelligence, Amazon · Oct 2026
- Research Labcs.LG
Muon optimizer boosts training success for long-horizon repository agents
Muon optimizer combined with token-averaged loss and auxiliary penalties improves training success in long-horizon RL for repository-level issue resolution.
Renmin University of China, Duke University · Oct 2026
- Chinese Techcs.LG
Validation-gated co-evolution improves agent models and execution harnesses
Alternating model training with checkpoint-specific, validation-gated harness updates outperforms weight-only training and ungated alternation on two agent benchmarks.
ICT AI Competence Center, Huawei Technologies Co., Ltd., Shanghai Jiao Tong University · Sept 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
- 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
- Chinese Techcs.AI
AI web agents learn long tasks by reusing skills salvaged from failures
SCOPE achieves 3.9% to 18.6% relative gains on web tasks by progressively discovering and composing local skills.
Renmin University of China, Huawei Technologies · Nov 2026
- Independentcs.CL
Small AI agents can handle computer tasks with clicks and code
Mage-VL-Cowork, a 4B hybrid computer-use agent, achieves 66.0% on OSWorld-Verified via evolved harness and five-stage training.
Nov 2026
- Big Techcs.AI
Synthetic computer worlds improve agents on unseen computer-use tasks
WildAWM uses open-source models to synthesize computer-use environments spanning GUI, MCP, and CLI; training only in these worlds improves agents across eight out-of-distribution benchmarks.
UNC-Chapel Hill, Microsoft Research · Oct 2026
- Big Techcs.CL
Coding agents improve more when their model and harness co-evolve
Alternating self-proposed harness revisions with policy reinforcement learning improves coding-agent resolve rates beyond optimizing either component alone.
Microsoft Research, University of Virginia · Oct 2026
- Chinese Techcs.AI
Synthetic cowork environments improve agents’ handling of changing shared workspaces
Cowork-Gym’s synthesized environments improve Qwen agent models by 4.1 to 8.7 points on six benchmarks.
Alibaba Token Foundry · Oct 2026
- Academiccs.AI
Realistic simulated users improve training for interactive language agents
MIMESIS learns realistic user behaviors from human conversations and reasoning traces, enabling agents to train more effectively and generalize across unseen users.
Meta Superintelligence Labs, New York University · Oct 2026
- Chinese Techcs.CV
Video agent answers questions about one video in a single shared reasoning pass
Coordinating multiple questions about one video in a single shared agent trajectory improves accuracy and cuts reasoning rounds and frames.
University of Science and Technology of China, Tencent · Oct 2026
- Chinese Techcs.CV
Self-evolving AI agent generates cinematic multi-shot videos by iteratively directing and critiquing itself
A unified Director Critic vision-language policy trained with alternating reinforcement learning improves cinematic multi-shot video generation.
Beijing University of Posts and Telecommunications, Peking University · Oct 2026
- Big Techcs.AI
Training models on runtime program-state reasoning boosts automated software engineering
Training LLMs to reason about runtime program states improves repository-level patch generation, test generation, and security PoC generation.
University of California, Santa Barbara, Microsoft · Oct 2026
- Big Tech
Privileged self-practice beats self-distillation for multi-turn language-model agents
Privileged Self-Practice injects task-specific instructions into failing agent rollouts to improve multi-turn agent training.
Texas A&M University, AWS AI, Amazon · Sept 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
- Big Techcs.CVcode
Structured intermediate rewards improve robots on long, dependent tasks
Structured intermediate rewards, gated by task prerequisites and scaled by completion pace, improve long-horizon VLA training over sparse terminal rewards.
The Pennsylvania State University, Amazon AWS AI · Sept 2026
- Chinese Techcs.CL
Agents learn to manage long contexts through fine-grained reinforcement learning
ContextPilot uses fine-grained RL and an expanded toolset to proactively manage context, outperforming existing methods.
Tsinghua University, Tencent Youtu Lab · Sept 2026
- Chinese Techcs.LG
Contrastive branch training improves credit assignment for tool-using language models
Contrastive Branch Policy Optimization improves tool-integrated agent training by separating branch selection from credit assignment and using entropy screening with contrastive branch values.
Alibaba Group, Harbin Institute of Technology · Aug 2026
- Big Techcs.AI
New framework lets you train AI agents inside the same harness systems they use at inference
OpenForgeRL enables end-to-end reinforcement learning of harness-based AI agents in any environment using a proxy and Kubernetes orchestrator.
Columbia University, Dartmouth College · July 2026
- Big Techcs.LG
Verifier-based rewards train coding agents toward formally correct programs
A reward from verifier diagnostics, plus integrity auditing, raises verified code generation from 66.1% to 93.5% on DafnyBench and from 56.3% to 100% on SWE-Proof.
UC Berkeley, AWS AI Labs · Oct 2026
- Big Techcs.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
- 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
- Big Techcs.LG
Context-aware credit assignment improves language agents on long-horizon tasks
CCPO improves RL for LLM agents by weighting observation-matched peers by context and combining historical and future credit.
Stanford University, Texas A&M University · Oct 2026
- Chinese Techcs.AI
Intervening During Rollouts Helps Agents Learn Beyond Their Initial Limits
Reward-aware correction of failed rollouts at intervention time lets LLM agents learn from trajectories they could not generate on their own, pushing RL beyond policy limits.
Alibaba Token Hub, Alibaba Group, HKUST (Guangzhou) · Oct 2026
- Big Techcs.AI
Self-judged trajectories improve SWE agent training beyond binary verifier rewards
Behavioral Self-Rewarding uses the policy's own trajectory comparisons to improve SWE agent training beyond binary test rewards.
Pennsylvania State University, AG2AI · Oct 2026
- Chinese Techcs.AI
Agents can skip repeated reasoning when earlier plans still suffice
Adaptive reasoning can be trained by estimating whether earlier reasoning still supports later actions, a likelihood signal that cuts reasoning cost while preserving performance.
Renmin University of China, Alibaba Token Hub, Alibaba Group · Oct 2026
- Big Techcs.CL
Eight-stage post-training recipe improves an open 106-billion-parameter language model
A reproducible eight-stage post-training recipe improves the GLM-4.5-Air-Base model beyond its official release.
Amazon · Sept 2026
- Research Lab
Reinforcement learning teaches an agent to verify mixed-source multimodal misinformation
Reinforcement learning trains an LVLM to use a toolkit of specialized tools for detecting mixed-source multimodal misinformation, achieving 70.4% accuracy on MMFakeBench.
Beijing University of Posts and Telecommunications, Minzu University of China · Sept 2026
- Independentcs.CLcode
A 2.8-trillion-parameter open MoE model approaches frontier performance.
Kimi K3 is an open, 2.8T parameter Mixture-of-Experts model with 104B activated parameters that achieves frontier-level performance across multiple domains.
July 2026
- Chinese Techcs.AI
Three-stage training instills world model planning into LLM agents
A three-stage training paradigm enables LLM agents to internalize a world model for future-aware planning and outperforms baselines on search and math reasoning.
Fudan University, Shanghai Innovation Institute · June 2026
- Big Techcs.IR
AI agents find better evidence when search considers their original question
ACARe uses agent context to rerank documents, improving retrieval and answer accuracy in multi-step search.
University of Waterloo, NVIDIA · Oct 2026
- Big Techcs.CL
Teacher rollouts repair myopic token scoring in agent distillation
Teacher next-token scores are myopic; repairing them with takeover-informed scores before distillation improves agentic on-policy distillation.
NVIDIA · 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.CL
Source-aware rewards reduce fabricated evidence in trained search agents
Making search agents explicitly label and reward external versus internal evidence sources reduces evidence fabrication by 74.8% and improves accuracy.
Zhejiang University, Microsoft · Oct 2026
- Big Techcs.DC
Sparse, exact updates cut trillion-parameter policy refits to minutes
NeMo-DCR sends only changed weights for RL refits, achieving bit-exact results 12-40x faster than full checkpoint transfers.
Aalto University, NVIDIA · Oct 2026
- Chinese Techcode
Trajectory graphs improve step-level credit assignment in agentic reinforcement learning
GRAFT uses trajectory graphs to estimate step-level advantages more faithfully than GRPO in multi-turn agentic tasks.
Shanghai Jiao Tong University, Tencent AI Platform Department · Sept 2026
- Chinese Techcs.AI
Evolving terminal environments keeps training tasks challenging as agents improve
Environment evolution off-policy generates increasingly difficult terminal environments, improving agent performance by 14-18 percentage points.
Hunyuan Team, Tencent · Sept 2026
- Chinese Techcs.CV
Unified visual-generation agentic model outperforms larger closed-source models
A single end-to-end learnable agentic model combines understanding, thinking, planning, and creation for visual tasks, outperforming larger closed-source models on a new benchmark.
Tencent Hunyuan, Hong Kong University of Science and Technology · Mar 2026
- Big Techcs.AI
Training agents around decisive errors improves prevention and recovery
PivotOPD jointly trains agents to prevent and recover from pivotal mistakes, achieving top average performance across three benchmarks.
Princeton University, NVIDIA · Oct 2026
- Top Universitycs.AI
Enumerating Tool Choices Beats Sampling Them in Genomic Reasoning
Enumerating all tool subsets during training outperforms sampling-based policy optimization for genomic tool selection.
The Chinese University of Hong Kong, Shenzhen, Shenzhen Future Network of Intelligence Institute (FNii-Shenzhen) · Sept 2026
Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.