News without the noise
39 results for “WebShop” · page 2 of 2
Research · Agents
RetireOPD: Self-Retiring On-Policy Distillation for Agentic Reinforcement Learning
RetireOPD uses a skill-conditioned teacher for dense supervision, then automatically retires it when the student is ready to continue with reinforcement learning alone.
Zhejiang University, Alibaba GroupCode
Research · Agents
Beyond the Teacher: Understanding and Preserving Distillation Overshoot in Agentic On-Policy Distillation
Zhong et al. show that on-policy distillation repairs weak student decisions early but erodes successful ones later, and propose SPOD to protect them.
WebShop: 50.3% success (as reported)
University of Southern California, University of Florida, Emory University
Research · Reinforcement learning
From Outcomes to Steps: Reliable Success-Aware Critic Learning for Long-Horizon Language Agents
SCALE adds an explicit success probability head to actor-critic RL and uses reliability-weighted changes in that probability to credit intermediate actions, raising success rates on five agent benchmarks.
WebShop: 4.1 pts success rate gain (as reported)
Shanghai Jiao Tong University, Huawei Noah's Ark Lab
Research · Agents
Co-Evolving Acting and Forecasting for Language Agents
CoPE trains agents to predict several subsequent actions from each interaction history, improving five benchmark results without adding inference-time computation.
WebShop: 77% success rate (as reported)
University of Illinois Urbana-Champaign, University of California San Diego, Meta AI
Research · Agents
MileGPO: Milestone Inference with Local Evidence for Graph-Based Policy Optimization of Long-Horizon LLM Agents
MileGPO extracts intermediate rewards from existing rollouts by identifying milestones and traps, weighting their reliability, and comparing alternative actions from the same state.
Beijing Jiaotong University, Peking University
Research · Agents
When Should Agents Think? Adaptive Reasoning via Cross-Turn Estimation
RACE trains language-model agents to estimate how long prior reasoning remains useful, reducing repeated reasoning across four interactive benchmarks.
WebShop: 535 tokens Avg Tok. (as reported)
Renmin University of China, Alibaba Token Hub, Alibaba Group
Research · Agents
State-OPD: State-Grounded On-Policy Distillation for Multi-Turn Agents
State-OPD separates executed interaction history from student reasoning, improving teacher supervision and broadening the actions sampled during training.
WebShop: 12.2% success rate (as reported)
Alibaba Group, Tsinghua University, Shanghai Jiao Tong University
Research · Reinforcement learning
TGRL: Temperature-Grouped Reinforcement Learning for Efficient Exploration in LLMs
By contrasting rollouts at different temperatures, TGRL allocates an explicit exploration signal without increasing sample budgets.
WebShop: 74.2% Success Rate (as reported)
University of Chinese Academy of Sciences, Meituan, MAIS&NLPR, Institute of Automation, Chinese Academy of SciencesCode
Research · Agents
RSI-Router: Evolving Subtask-Level LLM Routing and Skills for Cost-Efficient Agents
RSI-Router learns when to assign subtasks to large or small language models, along with reusable model-specific execution guidance.
ALFWorld: 75.3% inference cost reduction (as reported)
Shanghai Artificial Intelligence Laboratory
Research · Reinforcement learning
Targeting Pivotal Decisions for Credit Assignment in Agentic Reinforcement Learning
ProVer uses a model to identify consequential action segments, then validates their value through sampled outcomes before assigning extra training credit.
WebShop: 47.67% task success rate (as reported)
University of California, Davis, Microsoft, University of Washington
Research · Reinforcement learning
Cross-Rollout Bellman Closure for Long-Horizon Agentic Reinforcement Learning
CRBC recursively combines evidence from intersecting rollouts while weighting alternative actions and transitions by their observed frequencies.
WebShop: 90.11 task score (as reported)
Beihang University, Zhongguancun Academy, Communication University of China
Research · Agents
ASCENT: Online Test-Time Training of Long-Horizon Agents via Self-Distillation of Verified Experience
ASCENT enables online test-time training of LLM agents by self-distilling successful trajectories into persistent weight updates without external supervision.
WebShop: 41.9% strict success rate (as reported)
University of New South Wales (UNSW Sydney)
Research · Reinforcement learning
PivotOPD: Learning to Recover from Pivotal Mistakes in Multi-Turn Agents
PivotOPD identifies consequential mistakes in student rollouts, then distills both the preferred action and subsequent recovery behavior.
WebShop: 14.1% success rate improvement over RLSD (as reported)
Princeton University, NVIDIA, University of Maryland
Research · Agents
Retrieval-Augmented Skill Optimization via Cross-Harness Adaptation
RASO adapts procedural knowledge from an external skill corpus, using it both to initialize agent instructions and repair them after execution failures.
WebShop: 24.73 performance (as reported)
Korea University, KAIST, Meta AI
Research · Agents
UpliftMem: Learning Set-Level Uplift for Agent Memory Retrieval
UpliftMem learns which sets of stored examples improve task outcomes, then directs limited training rollouts toward alternatives most likely to change retrieval decisions.
WebShop: 44% success rate (as reported)
MIIT Key Laboratory of Data and Decision Intelligence, Beihang University
Research · Agents
Dyad: Extending Large Language Models with Native Typed Decision-Making
Dyad separates language reasoning from structured action selection, enabling efficient decisions over changing action spaces.
WebShop: 75.17% success rate (as reported)
Zhejiang University, SAP, Central South University
Research · Reinforcement learning
Mitigating the Length-Scaling Tax with Online Distillation
Length Self-Distillation reduces excess response length on easy queries without sacrificing accuracy on hard ones.
BrowseComp-Plus: 30.72% Pass@1 (as reported)
ByteDance Seed, Tongji University, Peking University
Research · Agents
Dense Is Not Enough: Hierarchical Supervision Allocation for Long-Horizon On-Policy Distillation
The LENS-OPD method outperforms dense token-level supervision by dynamically selecting which student behaviors to correct based on future impact and learnability.
WebShop: 24.5% SR (as reported)
Ant Group, Alibaba International Digital Commerce Group, University of Science and Technology of China
Research digest
Across routing, skill retrieval, harness optimization, and credit assignment, targeted structure outperformed brute-force evaluation.