ANTE's belief-anchored RL lifts Qwen3-32B to 57.2% on SWE-bench Verified
Yi Ge · hf · 2026-10-16
Researchers from the University of Maryland and AWS AI Labs (including Rein Houthooft and Aritra Sengupta) present ANTE (Anticipate, Then Execute), an agentic RL framework that gives coding agents an internal world model.
- Motivation: most coding agents never look ahead — their policies don't anticipate action consequences, and terminal rewards barely attribute failures to earlier steps, so a bad edit can irreversibly corrupt the repo before any signal arrives
- Method: the policy states pre-action predictions that are verified by real execution and turned into per-step credit and trust-or-abstain decisions; Belief-Anchored Policy Optimization (BAPO) anchors credit on the policy's own forecasts, while ANTE-TTO enables test-time scaling by selecting among candidate actions without an external scorer
- Results: raises Qwen3-32B from 25.0% to 57.2% pass@1 on SWE-bench Verified, 6.2% above a matched GRPO baseline; the world model's accuracy climbs from 44% to 69%
More from coding & agent
- Alibaba open-sources OpenGCUA, a full-stack generalist computer-use agent; +19.6 points on OSWorld-v2 across 8 models — Tongyi-MAI · 2026-10-16
- BotBeat: a social network where all users are AI agents, exposed via 32 MCP tools — Dismal_Piccolo4973 · 2026-10-16
- DKGI paper: agents write their own tools to search knowledge graphs, beating baselines by 6.7 Hit@1 — yuz9yuz · 2026-10-16
- Dev open-sources lgtmaybe: running PR code reviews on a local 27B model via hunk splitting — mattjcoles · 2026-10-16
- GLM 5.3 Flash is now free in OpenCode coding tool — DanielLockyer · 2026-10-16
- 12 React performance concepts in one infographic, from code splitting to Web Vitals — _jaydeepkarale · 2026-10-16