News without the noise
Benchmark
Reported in 65 papers on Zotpaper. The latest, each as the paper states it:
49 results for “ALFWorld” · page 1 of 3
Research · Agents
CORD: Conflict-Aware Optimization for Joint RL and Distillation in Agents
CORD selectively removes opposing gradient components while preserving a protected objective, boosting task success on ALFWorld and WebShop across several training recipes.
ALFWorld: 77.7% pooled success rate (as reported)
The Chinese University of Hong Kong, Tencent, Peking University
Research · Agents
SAPD: Learning from Agent Experience without Environment Interaction
SAPD decomposes recorded multi-turn histories into decision prefixes and uses relative step advantages to stabilize policy distillation, improving results on ALFWorld, WebShop and Search-based QA.
ALFWorld: 96.6% success rate (as reported)
Meituan, University of Chinese Academy of Sciences
Research · Agents
Learn Now, Use Next, Trust Later: Prequential Test-Time Learning for LLM Agents
The nonparametric framework improves LLM agent success on WebArena-Lite and ALFWorld by separating immediate guidance from cross-episode trust.
ALFWorld: 84% success rate (as reported)
Renmin University of China, Tencent
Research · Reinforcement learning
BLPO: Beyond-Local Policy Optimization for Long-Horizon LLM Agents
Across two simulated environments and two model sizes, the authors report higher task success when training reuses evidence about useful actions across tasks and training updates.
ALFWorld: 92.97% Out-Success (as reported)
Beihang University, Alibaba Group
Research · Agents
RLiD: Reinforcement Learning in the Decision Plane for Language Agents and LLMs
On a household-task benchmark, training a model to choose directly among labelled actions reached 99.3% success in 12.7 GPU-hours, where token-level reinforcement learning reached 86.6% in 17.3.
ALFWorld: 12.7 GPU-hours training and distillation compute (as reported)
University of Illinois Urbana-Champaign, National University of Singapore, Zhejiang University
Research · Agents
Verify--Backtrack Loop: Structured Precondition Recovery for Long-Horizon Language Agents
A training-free goal-stack controller lifts task success by up to 22 percentage points over ReAct across three benchmarks.
ALFWorld: 87.31% task success rate (as reported)
City University of Hong Kong, Huawei Noah's Ark Lab
Research · Evaluation & benchmarks
Retrieved but not ranked: surface-form bias in structural retrieval, from mathematics to agent trajectories
Across competition mathematics and embodied-agent trajectories, embedding models retrieved plausible candidates but often ranked lexical overlap ahead of structural similarity.
MathNet-Retrieve: 0% Hit@1 (as reported)
Independent, MIT CSAIL, Kellis Lab
Research · Agents
TalentRL: Strategy-Guided Reinforcement Learning for Hierarchical Skill Routing
A hierarchical reinforcement learning approach separates strategy selection from skill execution to improve routing credit assignment.
ALFWorld: 99.3% success rate (as reported)
Microsoft AI
Research · Reinforcement learning
Contextual Credit Policy Optimization
CCPO uses context-weighted trajectory comparisons and short-horizon progress to identify which actions contributed to eventual success.
ALFWorld: 96.9% success rate (as reported)
Stanford University, Texas A&M University, Yale University
Research · Agents
VETTA: Coordinating Turn- and Token-Level Credit Assignment for Multi-Turn LLM Agents
VETTA learns both credit types via separate value heads on a lightweight shared critic, improving success rates on two agent benchmarks.
ALFWorld: 91.9% success rate (as reported)
University of Science and Technology of China, Zhongguancun Academy, Shandong University
Research · Agents
Self-Evolving Agents through Privileged Retrospective Distillation
Retrospective Self-Distillation converts completed trajectories, including failures, into training signals without ground-truth solutions or verifier rewards for policy updates.
ALFWorld: 78.4% success rate (SR) (as reported)
Meta AI
Research · Reinforcement learning
ProAS: A Progress-Aware Advantage Shaping Framework for Multi-Turn Agentic Reinforcement Learning
ProAS uses a reference graph to mask invalid actions and penalize early deviations, boosting success and convergence.
ALFWorld: 2.44 x Spd@70 training speedup (as reported)
Zhejiang University, Westlake University, Westlake Institute for Advanced Study
Research · Reinforcement learning
TrajEvolve: Trajectory Evolution for Reinforcement Learning with Hindsight Credit Assignment
TrajEvolve revises unnecessary steps in historical trajectories and uses the resulting paths alongside fresh policy rollouts during training.
ALFWorld: 95.3% success rate (as reported)
City University of Hong Kong, Tencent, Nanjing University
Research · Agents
MemScope: Evolving Agent Memory Architectures through Localized Code Repair
MemScope uses a tool-using Patch Agent to inspect raw traces, runtime memory, and code to identify and apply bounded, in-place edits, outperforming summary-based feedback on multiple benchmarks.
ALFWorld: 71.64% success (as reported)
State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences, University of Chinese Academy of Sciences
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.
ALFWorld: 86.1% success (as reported)
University of Southern California, University of Florida, Emory University
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.
ALFWorld: 43.39% success rate (as reported)
Alibaba Group, Tsinghua University, Shanghai Jiao Tong University
Research · Agents
When Faster Decisions Do Not Make Better Agents: An Empirical Study of Jev in Agent Workflows
Across 4,500 executions, cheaper action selection produced workflow-dependent savings and sometimes reduced successful task delivery.
ALFWorld: 3.96 mUSD cost per delivered success (as reported)
Nanjing University, Technical University of Munich
Research · Evaluation & benchmarks
Harness-Delta Attribution: Dissecting Gains From Harness Evolution
Across 21 successful configurations, the authors attribute an average 54% of improvement to benchmark overfitting and 14% to increased inference-time computation.
ALFWorld: 100% success rate (as reported)
Amazon, New York University
Research · Agents
Explicit Trajectory Diversity for RL-Based Post-Training of LLM Agents
TJPO rewards task-specific variation across groups of trajectories while preserving competitive success rates.
ALFWorld: 0.85 User Constraints Success (as reported)
Microsoft, tuyoogame
Research · Agents
Reading Right, Acting Wrong: Diagnosing Decision Failures of System One Models in Agent Loops
Most decision errors in System One models occur when all probed facts about the agent's state are answered correctly, and these errors cluster on the same states across systems.
ALFWorld: 0.48 decision accuracy (as reported)
Carnegie Mellon University