ALFWorld

2026-10-23

RESEARCH · 1 SOURCE · arXiv cs.AI

D-Skill: distilling black-box teacher agents into natural-language skills

D-Skill is a framework that analyzes paired teacher-student rollouts to extract interpretable natural-language skill documents that are provided to the student at inference time. Built entirely offline from fixed trajectories with no further environment interaction or weight updates, D-Skill closes 75%–95% of the teacher-student performance gap across SpreadsheetBench-Verified, ALFWorld, and chess puzzles, outperforms one-epoch SFT baselines after a single offline pass, remains competitive with multi-epoch SFT, and yields 5.7x–8.3x lower inference cost per task for the student under the same model family.

7.0

2026-10-23

RESEARCH · 1 SOURCE · arXiv cs.AI

SPEM: structure-aware procedural memory for efficient skill use in LLM agents (arXiv preprint)

The paper introduces SPEM, a lightweight, structure-aware procedural memory framework that stores Atomic Skill Memory for individual skills and Skill-DAG Memory for reusable skill compositions; it uses Skill-GNN for state- and structure-aware selection and writes successful executions back as new Skill-DAGs. On ALFWorld and SkillsBench, SPEM achieves an 87.1% ALFWorld success rate and 38.38% SkillsBench reward while reducing token and runtime costs compared to strong baselines.

7.0

2026-10-21

RESEARCH · 1 SOURCE · arXiv cs.AI

M-Plan: closed-loop multi-agent framework integrating long- and short-term memory for LLM agents

M-Plan is a closed-loop multi-agent system (MAS) framework that combines long-term and short-term memory via global-local and past-current task-specific guidance. The paper (arXiv:2610.27032v1) describes guidance generation during execution, local hints when agents are confused, post-task memory optimization for future retrieval, and reports improved accuracy and efficiency on multi-turn benchmarks including ALFWorld, WebShop, and PDDL compared to several baselines.

6.0

2026-10-16

RESEARCH · 1 SOURCE · InfoQ AI, ML & Data Engineering

Google open-sources EnvHarness to adapt training environments for self-evolving agents

Google researchers open-sourced EnvHarness, a framework that wraps static environments via the standard reset()/step() interface and applies composable setup, rule, and link mutations so environments adapt to an agent's weaknesses without changing the underlying environment or evaluators. The release also includes EnvRigger, an automated loop that analyzes agent rollouts to generate and validate targeted environment edits; Google reports consistent performance gains (≈2.9–3.7 percentage points) across benchmarks such as WebArena, SWE-bench Verified, and ALFWorld on models including Gemini, Qwen, and Claude Sonnet.

7.0

2026-10-12

RESEARCH · 1 SOURCE · arXiv cs.AI

EvoBelief: falsifiable world-model paper for self-improving frozen LLM agents (arXiv v1)

EvoBelief is a self-evolving world model that represents environment knowledge as falsifiable hypotheses: unexpected outcomes trigger hypothesis formation and controlled validation before rules are used for prediction or planning while the language model weights remain frozen. Evaluated on ALFWorld, MiniGrid, and DiscoveryWorld Reactor Lab, EvoBelief improves ALFWorld success over a matched Frozen World Model by 23.9% with Claude Haiku 4.5 and 31.3% with GPT-5.4, adapts to hidden dynamics on MiniGrid, and discovers validated quantitative laws in DiscoveryWorld (83.3% Normal-tier, 58.3% Challenge-tier vs 0% for baselines).

7.0

2026-10-09

RESEARCH · 1 SOURCE · arXiv cs.AI

Plan-and-Patch: diffusion LLMs for generating and repairing long-horizon plans

The paper introduces Plan-and-Patch, a plan-and-act framework where a diffusion language model (dLLM) generates structured, program-like plans via parallel unmasking and repairs plans by filling selected regions while preserving surrounding steps. The authors compare DreamReasoner-8B (diffusion) and Qwen3-8B (autoregressive): on Natural Plan without task-specific training diffusion attains a 53.7% plan-repair success rate versus 27.0% for AR, and after task-specific training on ALFWorld and TextCraft diffusion and AR reach similar observed success while diffusion cuts mean plan-generation latency by 39–46%.

6.0

2026-10-07

RESEARCH · 1 SOURCE · arXiv cs.AI

SAGA: experience-grounded knowledge abstraction for self-evolving LLM agents (arXiv:2610.06964v1)

This paper introduces SAGA, a framework that progressively abstracts LLM agent interaction trajectories into hierarchical memory entries (episodic descriptions, reusable procedures, and principles with applicability conditions) linked to execution evidence. Retrieved principles are instantiated as task-specific guidance to refine actions, creating a feedback loop that improves performance in interactive benchmarks (ScienceWorld, ALFWorld) and where ablations show the importance of contextual instantiation and action regulation.

6.0

2026-10-02

RESEARCH · 1 SOURCE · arXiv cs.AI

Heavy-tailed memory traces in long-horizon language agents (arXiv:2610.00010v1)

This arXiv preprint studies how external memory use in long-horizon language agents concentrates on a small core of frequently retrieved states while leaving rare states in a long tail where prediction errors accumulate, with tail shape depending on policy (log-normal-like for random-walk agents vs truncated-power-law-like for semantic LLM policies). The authors introduce the Core--Tail World Model (CTWM), a rank-based memory controller that allocates prompt budget with a single exponent τ and preserves a summarized tail; on Synthetic Graph World CTWM preserves state/transition coverage, reduces prompt tokens by 5.9% and lowers bottom-half tail prediction error by 13.6% versus a graph-memory baseline, and also shows token savings on ALFWorld and a 24.48% token reduction on LongMemEval with aggregate accuracy parity.

6.0

2026-10-02

RESEARCH · 1 SOURCE · arXiv cs.AI

GAVA: Grounded arbitration of user corrections in text-based embodied agents (arXiv:2610.00282v1)

The paper formulates 'grounded correction arbitration' — a decision framework for agents to accept, reject, inspect, or ask about potentially incorrect user corrections — and introduces GAVA, which uses observation-bounded evidence, legal probes, and a one-step expected-loss rule. Evaluated in a text-only ALFWorld benchmark across many checkpoints and scenarios, GAVA achieves near-perfect correction accuracy while lowering interaction cost via selective inspection informed by an object-location prior, but the study is limited to symbolic observations, controlled speakers, and no human, visual, or robotic evaluation.

6.0

2026-10-02

RESEARCH · 1 SOURCE · arXiv cs.AI

When Harnesses Lose the Signal: causal evaluation of recovery in LLM agents and the Causal Intervention Router (CIR)

The paper frames recovery in LLM agents as a causal decision problem, comparing outcomes from the same execution state with and without recovery to separate rescue from harm and to track how recovery value changes over time. It introduces the Causal Intervention Router (CIR), a lightweight policy that decides whether to intervene using pre-recovery information; on long-horizon ALFWorld tasks with Qwen3-14B CIR raised success from 70.33% to 73.33% and left trajectories with correct observations untouched, with controls indicating benefits are not solely due to the new observation returned by the environment.

6.0

2026-09-18

RESEARCH · 1 SOURCE · arXiv cs.AI

Paper proposes 'checkpoint handoff' to separate REACH and SOLVE effects in agentic RL

The arXiv paper (2609.19636v1) introduces 'checkpoint handoff', an evaluation protocol that clones states reached by one checkpoint and runs another checkpoint from those identical states to decompose endpoint success into REACH (arriving at useful states) and SOLVE (finishing from identical states). Applied across two benchmarks and two released pipelines (including experiments on ALFWorld), the method shows positive interaction: RL histories help RL solvers more than SFT solvers, and independent REACH and SOLVE gaps predict aggregate gains.

7.0