Paper edits agents' stale reasoning instead of predicting tool outputs

single source· 1 articles · confidence: medium · first seen 2026-09-22 20:00 UTC

What this means for you

Nothing to act on today. This is a paper, not a release — no code, no weights, no API — and the scores come from the authors' own evaluation, with no harness or date given. If your agents carry stale assumptions through long traces, the decision-labelling idea is worth reading, but do not plan around the gains.

A paper proposes a different target for an agent's world model — the component that predicts what happens next in a task. Instead of predicting tool responses, which the authors argue is wasted effort when real feedback is available, AEWM labels each decision "critical", "exploratory" or "noisy" and rewrites bad reasoning out of the agent's history, so later steps are not built on assumptions it has already tested. On the authors' own Action Judge benchmark it reports 70.5% macro-F1, 10.6 points above the strongest baseline it compares against. Across six benchmarks and three agent backbones the editing loop adds 3.2–6.7 points. No code, weights or evaluation harness are named.

Key facts

  • ·AEWM reports 70.5% macro-F1 on the authors' Action Judge benchmark, 10.6 points above the strongest frontier baseline it compares against source
  • ·EditAct is reported to improve average scores by 3.2–6.7 points across six benchmarks and three agent backbones relative to the strongest baseline source
  • ·AEWM-RFT, rejection-sampling fine-tuning on verified EditAct trajectories, improves over Self-RFT by 2.2–2.6 points across three domains without online AEWM guidance source
  • ·The method was trained across Search, Terminal and Software Engineering settings using mid-training and supervised fine-tuning source
  • ·The work is posted as arXiv preprint 2609.28416 dated 22 September 2026 source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire