News without the noise
57 results for “ALFWorld” · page 2 of 3
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: 98% overfitting share (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 Disruption Success (as reported)
Microsoft, tuyoogame
Research · Reinforcement learning
GraphHCA: Closed-Form Hindsight Credit Assignment for Long-Horizon LLM Agents
GraphHCA derives step-level training signals from pooled rollout graphs without a critic, hindsight model, or additional forward pass.
ALFWorld: 96.9% success rate (as reported)
Beihang University, Zhongguancun Academy, Communication University of China
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
Research · Reinforcement learning
My FAULT: Self-Diagnosis as Credit Assignment in Self-Evolving Agentic Reinforcement Learning
FAULT converts verified natural-language error diagnoses into step-level training signals while preserving the trajectory’s terminal outcome.
ALFWorld: 68.7% success rate (as reported)
Alibaba Group, Kyoto University, NII LLMC
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.
ALFWorld: 90% success rate (as reported)
University of New South Wales (UNSW Sydney)
Research · Agents
Graph-Conditioned On-Policy Agent Distillation from Off-the-Shelf Teachers
Graph-conditioned on-policy distillation scores student actions against indexed teacher executions, improving success on three agent benchmarks.
ALFWorld: 96.07% success rate (as reported)
Yuanbao Team, Tencent, Tsinghua University, Huazhong University of Science and Technology
Research · Agents
Learning from Interaction Experience via Localized Context Distillation
The authors propose a method that identifies and revises only critical spans in a student's trajectory, outperforming both off-policy and on-policy context distillation across math, code, and agent tasks.
ALFWorld: 33.08% task success rate (as reported)
Tencent, The Chinese University of Hong Kong, Zhejiang University
Research · Robotics & embodied AI
SAGE: Symbolic Action-Gating and Editing for LLM Task Planners
SAGE blocks actions that violate environmental preconditions and regenerates only the affected portion of an LLM-generated plan.
AI2-THOR: 0.71 step_success (as reported)
Korea Electronics Technology Institute (KETI)
Research · Agents
Do LLM Agents Execute the Plans They Declare? From Planning-Mode Declaration to Pattern-Specific Execution
Routing tasks to executors designed for specific planning patterns substantially improved structural adherence and task success.
ALFWorld: 0.92 TSR (as reported)
ADIA Lab, University of Granada, Luxembourg Institute of Science and Technology
Research · Reinforcement learning
T2SPO: Trajectory-to-Step Policy Optimization for Agentic Reinforcement Learning
T2SPO uses a frozen TabPFN to estimate remaining distance to success at each state, deriving auxiliary step rewards from past successful rollouts to improve policy learning.
ALFWorld: 74.7% overall success rate (as reported)
Nanjing University, ByteDance
Research · Agents
EvoBelief: Falsifiable World Models for Self-Improving LLM Agents
EvoBelief introduces a lifecycle of hypothesis formation, validation, and revision that improves task performance across static and changing environments.
ALFWorld: 66.42% success rate (as reported)
Tulane University, Amazon
Research · Agents
Plan-and-Patch: Diffusion Language Models for Agentic Planning
The Plan-and-Patch framework uses parallel unmasking for planning and infilling for repairs, outperforming autoregressive models in repair success and latency.
ALFWorld: 44% task success rate (as reported)
University of Texas at Austin, Amazon
Research · Agents
Sibyl: An Efficient Small-large Model Collaboration Framework for Long-horizon Tasks
Sibyl trains a 0.6-billion-parameter model to request step-level assistance only when useful, then absorb that guidance into later decisions.
ALFWorld: 0.8 calls/episode calls per episode (as reported)
Hangzhou Dianzi University, Fudan University, The Hong Kong University of Science and Technology
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.
ALFWorld: 5.59 pts overall success rate improvement over strongest baseline (as reported)
Beihang University, Zhongguancun Academy, Communication University of China
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.
ALFWorld: 73.7% success rate (as reported)
Princeton University, NVIDIA, University of Maryland
Research · Agents
Does This Action Still Explain the Task? Reverse Scoring for Diffusion Language Model Agents
Reflect Reverse reranks candidate actions by how well they explain the task, reducing retry loops without additional training.
ALFWorld: 29.33% progress rate (as reported)
Huawei Noah’s Ark Lab, Technical University of Darmstadt, UCL Centre for AI
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.
ALFWorld: 90.9% success rate (as reported)
Zhejiang University, SAP, Central South University
Research · Reinforcement learning
From Imitation to Reward Discovery: On-Policy Warmup for Agentic RL
A teacher-guided warmup improved early reward discovery and subsequent reinforcement learning in interactive agent tasks.
ALFWorld: 96.7% success (pass@1) (as reported)
Zhejiang University, Ant Healthcare, Ant Group
Research · Agents
Procedural Exposure Bias in Skill-Augmented Language Model Agents
The same procedural content, shown as text versus as an image, leads to measurable differences in action preference and downstream agent success.
ALFWorld, ScienceWorld, WebShop: 44.06% macro-average success rate (as reported)
Northeastern University, Alibaba Group