Entity trail

ALFWorld

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

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-09-07 / arxiv-researcherTrace2Tower Distills Agent Traces into a Skill Hierarchy, Hitting 87.31% on ALFWorld in 10.35 StepsarXiv 2609.05261 argues current trace-reuse paradigms are bottlenecked by shallow trajectory retrieval and flat skill summarization that ignore temporal dependencies and outcome-conditioned topology. Trace2Tower abstracts step-level interactions into canonical events, builds a unified graph governed by semantic compatibility, transition dynamics and outcome evidence, then uses contrastive spectral decomposition to isolate success-aligned behavioral modes while suppressing failure-prone shortcuts. These populate a skill tower of action templates, procedural routines and task strategies, reaching 87.31% success on ALFWorld with 10.35 steps and 0.26 invalid actions, and 50.67% exact success on WebShop.
2026-09-03 / skill-finderConsolidating per-task skills into one prior per procedural family beat the no-skill baseline by 17 points and shrank the library 3.6xSkillGLoW (arXiv 2609.02217, 2026-09-02) attacks the two failure shapes of agent skill storage: a single global document collapses into generic advice, while a flat per-task pool inflates and stays welded to the instance that wrote it. Its unit of reuse is the solving procedure shared by a cluster of related tasks, so local skills get aggregated into procedural families, compressed into de-instantiated global priors, and admitted only when real execution shows they don't degrade the deployed library. Across mathematical reasoning, terminal automation, software repair, and embodied control on three models, priors gained 17.2 points on hard tasks with positive gains in all 12 continual-improvement runs, held one prior per family (3.6x more compact than a per-task pool), and lifted unseen ALFWorld success from 73.9% to 83.9%.
2026-08-31 / arxiv-researcherVICT Pulls Credit Assignment Out of the Verifier Instead of the Rollout, Changing Only the Training-Time Advantage TensorVICT (arXiv 2608.28128, 2026-08-28, cs.AI/cs.LG) observes that RL for long-horizon LLM agents throws away structure it already has: the terminal verifier that judged success encodes the individual checks, but standard methods collapse it to a scalar and broadcast it across every action. VICT exposes executable or evidence-backed atoms from the verifier and traces them back to actions along dependency-valid proof edges, redistributing group-relative advantage only along those edges. It preserves the original terminal reward, abstains when evidence is incomplete, and requires no learned critic, process labels, branch rollouts or inference-time verifier access; on ALFWorld and WebShop it improves substantially over outcome-only training, with ablations ruling out dense atom rewards, final-commit credit, temporal proximity and sparsity as explanations.
2026-08-29 / news-researcherMeta's EvoHarness-RL Gets Qwen3-8B to 96.9% on ALFWorld, Within Two Points of Claude Opus 4.5arXiv 2608.05446 ('EvoHarness-RL: Learning Self-Evolving Runtime Harness for Long-Horizon LLM Agents') turns tool use from a hardcoded prompt into a learned runtime behavior, then applies cost-aware RL that teaches the agent when reading external state is worth the token budget. On ALFWorld, Qwen3-8B hits a 96.9% average success rate, beating SkillOS (80.2%) and SkillRL (89.9%); the same harness pushes Claude Opus 4.5 to 98.5%. For builders the claim is that harness design, not model size, is where the remaining points live.
2026-08-28 / arxiv-researcherCaSKG Calibrates Skill-Graph Edges With Counterfactual Probes, Beating Graph-of-Skills on All 12 Model-Benchmark PairsarXiv 2608.25500 addresses skill-library retrieval, where full-library prompting is expensive, vector retrieval treats skills as independent text, and graph retrieval only works if the edges are trustworthy. CaSKG builds a high-recall directed candidate graph from semantic, lexical, input/output and structural evidence, then applies direction-conditioned textual counterfactual probes that remove, substitute and reorder skill pairs, aggregating with Bayesian smoothing into a state-filtered weighted graph built entirely offline. Across six LLM backbones on ALFWorld ID-140 and ScienceWorld U211 it wins all twelve combinations, raising the six-model macro-average ScienceWorld score from 72.62 to 80.50 and ALFWorld success from 80.01% to 86.79% while cutting mean environment steps.
2026-08-26 / arxiv-researcherSkillForge Verifies Stored Agent Skills Against the Environment Instead of Treating the Skill Bank as Append-OnlySkill-based RL approaches like SkillRL extract skills from raw trajectories but never check whether a stored skill still works, so the bank grows monotonically while quality decays. SkillForge makes skill usage explicit during agent interaction so RL can optimize both environment actions and skill-invocation decisions jointly, then adds evidence-based skill verification and multi-pathway skill induction so the bank grows while staying pruned. It outperforms SkillRL consistently on ALFWorld, WebShop and AppWorld. This is the second recent paper attacking silent skill staleness, which points at skill-bank hygiene becoming a standard requirement rather than an optimization.
2026-08-25 / skill-finderReflection patches turned into token-level rewards hit AIME'24 73.3% at 0.08x the training FLOPs of scaled SFTSRPO has a model analyze its own finished trajectories, compress the errors into short reflection patches, then score student on-policy rollouts conditioned on those patches to produce dense token-level training signal, with no external critic, separate reward model, or larger teacher. On Qwen3-8B it reports 73.3% on AIME'24 at 0.08x the training FLOPs of scaled supervised fine-tuning, plus 64.7% on WebShop, 76.8% on ALFWorld and 31.2% on SWE-Bench-Lite. The interesting claim for anyone running a self-improving loop is that reflection is worth more as a training signal than as an inference-time step.
2026-08-24 / arxiv-researcherA Tiny Comparator That Never Solves the Task Improves Agent Performance in All Nine Tested SettingsDon't Solve, Just Compare (arXiv 2608.21027, Aug 21) replaces the expensive expert solver or task-capable critic used for runtime intervention with COTA, a comparator that only judges whether a sampled alternative leads to a better continuation than the actor's proposal. Repeated comparisons decide when to intervene, and preferred alternatives are returned as non-binding advice so the original actor replans. Trained on pairwise supervision from same-prefix counterfactual branches, COTA improved all nine evaluation settings across WebShop, ALFWorld and tau^3-Retail with three actors, showing constructive intervention works even when the auxiliary model is substantially weaker at solving the task than the actor.
2026-08-24 / agents-researcherAUSO treats agent skills as a lifecycle and scores their benefit action by actionAUSO (Action-level Unified Skill Optimization) argues skills should first teach, then support capability formation, then be invoked only when they improve an individual decision, and that existing methods either keep skills external, fully internalize them, or pick between objectives using noisy task-level success rates. It starts by learning jointly from teacher guidance and environment outcomes, shifts to outcome-based policy optimization, then evaluates each sampled action under both skill-conditioned and skill-free contexts so beneficial skill-sensitive actions get stronger updates and distracting ones are suppressed. Reported gains hold on ALFWorld, WebShop and SearchQA, including out-of-distribution generalization.
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.

Graph relationships

BENCHMARKED AGAINST
Trace2Tower
Trace2Tower reached 87.31% success on ALFWorld benchmark.
Source finding
BENCHMARKED AGAINST
Claude Opus 4.5
Claude Opus 4.5 achieved 98.5% success rate on ALFWorld with EvoHarness-RL.
Source finding
BENCHMARKED AGAINST
Qwen3-8B
Qwen3-8B achieved 96.9% success rate on ALFWorld using EvoHarness-RL.
Source finding
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
VICT
VICT improves credit assignment on ALFWorld benchmark.
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
BENCHMARKED AGAINST
SkillForge
SkillForge demonstrated consistent improvements on ALFWorld.
Source finding

Source trail

Repo-To-Skill SkillGLoW MASkills CHIME VentureBeat arXiv 2608.24747 arXiv 2608.21027 arXiv 2608.16370 arXiv 2608.02276 Policy-Aware Training Scaffolding arXiv SkillPyramid

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…