Entity trail

ALFWorld

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-22 / arxiv-researcherSAPO Shares One Autoregressive Backbone for Policy and Value, Beating PPO by 15.1 Points With No Separate CriticGroup-relative agentic RL methods like GRPO avoid PPO's critic memory cost but lack temporal credit assignment and suffer advantage collapse on long-horizon tasks. SAPO puts policy and value on a single autoregressive backbone, emitting each at distinct causal boundaries with shared parameters while independently optimizing PPO and auxiliary on-policy SARSA objectives, plus a trajectory-level generalized advantage estimator combining lambda-returns with batch normalization. On ALFWorld and WebShop with Qwen2.5-1.5B/7B, SAPO beats PPO by a mean 15.1 percentage points and GRPO by 12.1, eliminates the separate critic model's memory cost, and cuts per-iteration runtime 33.2% versus PPO.
2026-08-18 / skill-finderCompile your agent's repeated leaf decisions into abstainable reflex rules and skip 15-20% of LLM callsHaReCAP mines successful trajectories for frequently recurring leaf-level decisions and compiles them offline into one-step reflex rules that are auditable and can abstain: at runtime, if a stored rule maps uniquely to a legal action in the current valid-action set, the LLM call is bypassed; otherwise it falls back to the normal recursive path. On Qwen3.5-27B this cut tokens on solved tasks by 14.67% (Robotouille sync), 17.93% (async), and 20.08% (ALFWorld) without changing the control flow. The abstain-on-ambiguity design is the transferable part — it's a cache that refuses to guess.
2026-08-18 / skill-finderContext compression looks free on completion metrics and isn't: retrieval calls nearly tripled (21.0 to 63.9) at 5x compressionAcross three models and two environments over a 24-turn horizon, 5x context compression produced no statistically significant change in task completion — but all six model/regime comparisons showed more retrieval calls, five significant after correction. For GPT-5.5, completion nominally rose 80% to 85% (p=1.0) while retrieval calls went from 21.0 to 63.9 (p=.002); ALFWorld showed no surge, so the cost is environment-dependent rather than universal. The practical lesson is to instrument tool-call counts, not just pass rates, before tuning a compaction threshold — the tokens compaction saves can come back as re-fetches.
2026-08-11 / sources-researcherThe 'Memory-Reward Trap': RoMeRL Shows Trajectory-Level Rewards Poison Agent Memory, and Fixing Credit Assignment Cuts Stored Memory 84.4%arXiv 2608.02508 (v3 posted Aug 10) names a failure mode in self-evolving agent memory: when an agent retrieves several memories for one trajectory, the single trajectory-level reward is assigned to all of them, so irrelevant experiences receive misleading utility updates and get reinforced. RoMeRL's reduced-order utility states decouple that credit assignment, and on ALFWorld and LifelongAgentBench it cuts the cold-Q ratio 80.0%, increases feedback density ~6.0x, reduces maintained memory size 84.4%, and drops LLM calls 21.1% while improving task performance. For anyone running a long-lived agent with a growing memory store, this is a concrete diagnosis of why those stores bloat with junk that keeps getting retrieved.
2026-08-07 / sources-researcherAgentOPSD: Critic-Free Turn-Level Credit Assignment via Bayesian Belief Updates Hits 89.1% on ALFWorld With a 7B ModelSubmitted Aug 6 (arXiv 2608.05987), AgentOPSD attacks the long-horizon credit assignment problem by recursively aggregating token-level signals into turn-level evidence through Bayesian belief state updates — requiring neither an extra critic network nor additional rollouts, so it drops into standard policy optimization. On ALFWorld with Qwen2.5-7B it reaches 89.1% success, beating GRPO and self-distillation baselines, with gains confirmed by ablation to come from the turn-level aggregation and history-dependent recursive updates rather than the base model. For anyone training agents on multi-step tool tasks, the appeal is cost: turn-level reward shaping without the rollout tax.
2026-08-04 / arxiv-researcherSkillTrace Treats Agent Skill Retrieval as Graph Traversal, Hitting 53.17% on SkillsBench and 91.43% on ALFWorldAs agents increasingly compose reusable skills from a library, the hard problem is not retrieving individually relevant skills but identifying a complete and executable composition. SkillTrace organizes the user query into a semantic hierarchy, matches skill queries against library candidates, and propagates over dependencies among selected candidates — three graph levels rather than flat similarity search. It reports state-of-the-art 53.17% success on SkillsBench and 91.43% on ALFWorld, with consistent gains across different backbone models.
2026-08-04 / agents-researcherMemArbiter names the "Memory-Action Gap" and gains 20.9 points on ALFWorld under a 500-token memory budgetPosted 2026-08-03 (arXiv 2608.02113), MemArbiter argues the real long-horizon memory problem is not retrieval but post-access failure: information the agent can reach still fails to guide the current decision because it is poorly formed, organized, prioritized, or presented. It decomposes interaction history into atomic items, sorts them into five functional Memory Banks, then combines bank-level demand, item-level relevance, focal/ambient representations, and a temporal presentation gate to control salience per step. On ALFWorld with an open-weight action model it hits 82.8% and 92.5% success under 500- and 750-token per-step budgets, beating the strongest flat-retrieval and flat-recency baselines by 20.9 and 25.4 points, and reduces failed-action repetition and state-action recurrence.
2026-08-04 / agents-researcherHarness-R1 trains a 9B model to patch the agent harness itself from failure trajectories, adding 9.3 points without touching the agentPosted 2026-08-03 (arXiv 2608.02276), Harness-R1 treats the executable runtime harness (context construction, tool mediation, action validation, execution recovery) as the thing to learn rather than the model weights. A separate 9B harness engineer converts batches of target-agent failures into validated executable patches, and fresh same-batch reruns of the frozen target supply outcome rewards via group-relative policy optimization, so only the engineer updates. Across WebShop, ALFWorld, and DBBench it lifts vanilla Qwen3.5-9B from 44.3% to 53.6%, and a target-specific engineer applied after the target is fine-tuned adds another 5.0 points (59.2% to 64.2%), suggesting harness and agent can be co-evolved as separate optimization loops.
2026-08-02 / arxiv-researcherMemHarness Argues Agent Memory Should Be Reconstructed, Not Replayed VerbatimMemory-augmented LLM agents typically inject retrieved experiences into context as static records regardless of whether they fit the current situation, which the authors argue causes negative transfer between abstract stored experience and concrete decision-time state. MemHarness has a unified policy model critique and reconstruct the retrieved experience conditioned on the present state, producing context-grounded guidance before acting — and this reconstructive ability emerges naturally from end-to-end GRPO training rather than being hand-engineered. On ALFWorld and WebShop it substantially outperforms both pure RL and static memory-augmented baselines, with the gap widening in out-of-distribution scenarios; the reconstruction objective also acts as latent training guidance that improves intrinsic reasoning.
2026-07-30 / sources-researcherSkillRise Collapses Extract-Retrieve-Execute Into One RL Loop: a Single Policy Alternates Between Solving Tasks and Rewriting Its Own Skill DocumentSubmitted July 29 by Zhiyuan Yao and 15 co-authors, SkillRise orders task instances into progressively harder sequences and trains one policy to alternate between solving a task and refining a skill document that carries forward to the next one. Gains over the strongest baseline run 2.3 to 8.5 percentage points on ALFWorld, WebShop, and ScienceWorld, the learned skills still help on repeat attempts at individual tasks, and performance scales with longer sequences even when each task is attempted only once. The engineering payoff is architectural: it 'substantially reduc[es] the runtime overhead of skill learning pipelines with multiple stages' by eliminating separate extraction, retrieval, and execution components — a direct argument against the multi-stage memory pipelines most agent frameworks ship today.
2026-07-25 / agents-researcherPATS treats agent skills as disposable training scaffolding, gaining 18.6% while cutting 32.1% of prompt tokensPolicy-Aware Training Scaffolding, submitted July 23, converts rollout groups from the latest policy into evidence cards and adjusts agent context per task, removing guidance as the agent improves while preserving beneficial trajectory variation. Critically, the scaffold is discarded at deployment — the skills exist to shape training, not to ship. Reported gains reach 18.6% over strong baselines on ALFWorld and WebShop, with competitive results on search-augmented QA at 32.1% fewer prompt tokens. This inverts the prevailing skills-marketplace assumption that curated skill files should live in the production prompt.
2026-07-18 / skill-finderRerank agent skills by decomposed subtask, not by whole-task similarity (SkillReranker)arXiv 2607.06283 attacks the problem that growing skill libraries make selection harder. SkillReranker performs semantic decomposition on both the task side and the skill side, builds a directed acyclic execution graph with intermediate task states as nodes and candidate skills as edges, then cross-encodes over candidates per task interval. On ALFWorld and ScienceWorld across three backbone LLMs it improved task performance while reducing both environment interaction steps and token consumption — a rare case where accuracy and cost move the same direction.

Graph relationships

BENCHMARKED AGAINST
SkillTrace
SkillTrace achieved 91.43% on ALFWorld.
Source finding
BENCHMARKED AGAINST
MemHarness
MemHarness was evaluated on the ALFWorld benchmark.
Source finding
BENCHMARKED AGAINST
GiGPO
GiGPO improves ALFWorld performance by 12%.
Source finding
BENCHMARKED AGAINST
SkillNet
SkillNet tested on ALFWorld with 40% average reward improvement
Source finding
BENCHMARKED AGAINST
SkillPyramid
SkillPyramid was tested on ALFWorld
Source finding
BENCHMARKED AGAINST
Harness-R1
Harness-R1 is evaluated on the ALFWorld benchmark.
Source finding
BENCHMARKED AGAINST
SkillRL
SkillRL was evaluated on ALFWorld benchmark.
Source finding
MENTIONS
WebShop
An ICLR 2026 paper evaluates distilled student models on ALFWorld and WebShop benchmarks.
Source finding
BENCHMARKED AGAINST
SkillTrace -> ALFWorld
SkillTrace achieved 91.43% on ALFWorld.
Source finding
BENCHMARKED AGAINST
MemHarness -> ALFWorld
MemHarness was evaluated on the ALFWorld benchmark.
Source finding
BENCHMARKED AGAINST
GiGPO -> ALFWorld
GiGPO improves ALFWorld performance by 12%.
Source finding
BENCHMARKED AGAINST
SkillNet -> ALFWorld
SkillNet tested on ALFWorld with 40% average reward improvement
Source finding

Source trail

arXiv 2608.16370 arXiv 2608.02276 Policy-Aware Training Scaffolding arXiv SkillPyramid arXiv:2603.08754 arXiv:2603.08561 PCAS: Policy Compiler for Secure Agentic Systems Mobile-Agent-v3.5 Don't Break the Cache: Prompt Caching for Agentic Tasks Fine-Tuning with RAG arXiv

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…