Mila's Recursive Self-Aggregation pushes Gemini 3 Flash near top of ARC-AGI-2
morgymcg · x · 2026-10-20
- Mila's Recursive Self-Aggregation (RSA) paper introduces a test-time scaling method combining parallel and sequential scaling: sample 8 reasoning chains per question, randomly form 4-member subgroups, feed each back to the model with an aggregation prompt to pick the best solution or try a new approach, and repeat K times (5 in the paper).
- RSA beats purely parallel and sequential scaling on AIME-25, HMMT-25, Reasoning Gym, LiveCodeBench-v6 and SuperGQA; Gemini 3 Flash + RSA approaches the top of the ARC-AGI-2 public leaderboard, and Qwen3-4B becomes competitive with larger reasoners like DeepSeek-R1 and o3-mini (high).
- The authors also propose an aggregation-aware RL approach that trains models to combine candidate solutions for further gains.
- Relatedly, Harness2 borrows RSA for agent harness optimization: per task it proposes domain-specific and domain-agnostic harness edits in parallel, feeding the harness and its execution traces to the model, with an outer harness deciding how to combine them before re-running.
More from coding & agent
- Developer turns son's PC into a GitHub runner, then realizes the security risk — tristanbob · 2026-10-20
- ArkenOS launches Optimus-1, a swarm harness for autonomous research — _AustinCalvert_ · 2026-10-20
- Zero code, seven minutes: AI built a Bitcoin-prize hacker puzzle game for a 30th birthday — jamesbrooksco · 2026-10-20
- 10-step workflow to set up Claude Cowork as a personal AI workspace in 2 hours — nikola_mr64990 · 2026-10-20
- OpenAI DevDay Bengaluru: staff share lessons on long-running agents — paw_lean · 2026-10-20
- Proof of Agent: an experimental network where AI agents mine, grade and self-govern — adamamcbride · 2026-10-20