Agent-Editing World Model: Rethinking World Modeling for LLM Agents

What it adds beyond the summary

The training scale, the named baselines, the per-benchmark deltas, and the backbone dependence — none of which the summary carries.

How it works

The pre-execution state is s_t = h_t ⊕ (r̂_t, â_t): history plus the agent's proposed reasoning–action pair. Action Judge labels that pair CRITICAL / EXPLORATORY / NOISY before execution. Only NOISY proposals trigger State Revision, which generates a replacement pair (r̃_t, ã_t) from the same history. The selected action then runs in the real environment and its actual observation enters history — the contrast with prior world models is that AEWM edits the agent's state (s_t → h_t ⊕ (r̃_t, ã_t)) rather than predicting an observation ô_t.

Training data is synthesized, not human-labeled: a strong annotation agent decomposes verified successful trajectories turn-by-turn and assigns labels retrospectively (using the full trajectory), but the training target sees only the history visible at that turn — so a retrospective label supervises a prospective prediction. State Revision data comes from a proposal agent plus revision agent, keeping only revisions that produce "substantive progress" in the subsequent real trajectory. Two stages: mid-training on ~52B tokens across the three domains, then SFT on 120K examples (60K Action Judge + 60K State Revision).

flowchart TD
  P["Agent proposes (r̂,â)"] --> J{Action Judge}
  J -->|critical / exploratory| K["keep pair"]
  J -->|noisy| R["State Revision → (r̃,ã)"]
  K --> E["execute in real env"]
  R --> E
  E --> H["observation enters history"]
  H --> P

Evidence

• Action Judge benchmark (3,000 decisions, split evenly Search/Terminal/SWE): AEWM 70.5% macro-F1 vs. the strongest frontier baseline at 59.9. Baselines are named: Gemini-3-Pro (45.1), GLM-5.2 (53.2), Qwen3.7-Max (54.9), GPT-5.5 (55.3), DeepSeek-V4-Pro (59.9).
• EditAct (Qwen3.5-35B-A3B), agent-only → EditAct: BrowseComp 40.9→48.1 (+7.2), DeepSearchQA 69.4→78.6 (+9.2), Terminal-Bench 2.0 37.1→48.3 (+11.2), Doc2Repo 42.8→48.9 (+6.1), NL2Repo 22.4→25.3 (+2.9), SWE-Bench Pro 40.6→43.8 (+3.2).
• Backbone dependence matters: average gain over strongest baseline is +6.7, +5.2, +3.2 for Qwen3.5-4B, -9B, -35B-A3B respectively — the smaller the agent, the larger the benefit.
• AEWM-RFT (RFT on verified EditAct trajectories, no AEWM at inference) beats Self-RFT by +2.2 (BrowseComp), +2.6 (Terminal-Bench 2.0), +2.5 (Doc2Repo).

Limits

This is an abstract-plus-introduction page; the fetched text cuts off mid-Section 3.2, so filtering rules, benchmark construction, and ablations sit in appendices not shown here. Gains shrink as the agent backbone grows (+3.2 at 35B), so the value on a strong base model is the smaller end of the range. All results are the authors' own; NL2Repo and SWE-Bench Pro see only +2.9/+3.2 under EditAct. A Dataset and GitHub link are referenced but no weights, license, or run commands appear in the available text.