MIRA-AC trains research agents on proxy feedback, lifting symbolic regression from 50 to 63 solved
anirudhg9119 · x · 2026-10-14
The author presents MIRA-AC: an autoresearch policy trained purely on proxy feedback that improves gold outcomes over untrained MIRA:
- Symbolic regression: 50 → 63 solved
- Physics discovery: 8 → 10 solved
- Causal discovery: 76 → 84 solved
Key design: the critic sees causal data; the actor does not. For RL, full research episodes are reconstructed and decision-level advantages computed from downstream outcomes; only the meta-reasoner's reasoning and work orders are updated while executor, tools and context curation stay fixed — "training WHAT, not HOW." To fix credit-assignment bias from auto-compaction splitting long episodes, MIRA resets context at every decision and stitches meta-reasoning decisions back into one logical episode.
More from coding & agent
- OpenClaw UI adds sidebar rail, multiplayer features and smarter model picker — heyneighbor · 2026-10-14
- Merged Changes Per Day Jump 10x to 30,000 in Six Months — mcraddock · 2026-10-14
- Effect ships zero-dependency native test runner, agent-first — samgoodwin89 · 2026-10-14
- Codebases will look a lot like databases in the AI coding era — hargup13 · 2026-10-14
- Agent Argon reads Spot's SDK docs itself, builds trash-sorting skill on Boston Dynamics robot — chris_j_paxton · 2026-10-14
- Handling live web search in MCP agents without flooding the context window — Few_Training_8950 · 2026-10-14