Meta paper: agentic meta-reasoning hits 71.5% vs Codex 58.0% on ProgramBench
CatAstro_Piyush · x · 2026-10-19
A new paper by Paras Dahal and 12 co-authors from Meta and collaborators introduces agentic meta-reasoning, an inference-time harness that turns execution control into an explicit, structured reasoning process.
- Design: workers perform task-level computation while a controller consolidates what the run has established, explores next options, evaluates each under the remaining compute budget, and dispatches work — carrying only a compact run account instead of replaying full history. The model effectively builds its own computational graph, deciding what to explore, prune, verify, and backtrack.
- Results: on ProgramBench (long-horizon program reconstruction), meta-reasoning scores 71.5% with GPT-5.5 vs 58.0% for Codex, and 67.2% with Opus 4.8 vs 65.5% for Claude Code. On abstract reasoning, multi-domain long-horizon reasoning, and proof generation benchmarks, it gains 3.6–4.2 points over direct control baselines.
- The author list includes Rob Fergus, Jason Weston, Ruslan Salakhutdinov, and Taco Cohen.
More from coding & agent
- Sentry co-founder's AI writing workflow: dictate ideas, let an agent outline, write it yourself — lmoroney · 2026-10-19
- CarryOn: object-oriented context runtime for long-horizon agents boosts OSWorld by 5.8% — McAuley-Lab · 2026-10-19
- TreeOPSD: Tencent and Tsinghua's critic-free RL for long-horizon LLM agents — Tencent-Hunyuan · 2026-10-19
- Ant International's BizBench: frontier agents often violate business constraints — ant-intl · 2026-10-19
- PRISM: privileged sibling-graph signals fix multi-agent topology drift — NationalUniversityofSingapore · 2026-10-19
- Single-prompt video editing with Opus 5.5 and Blender wows devs — _AustinCalvert_ · 2026-10-19