MIRA-AC results: proxy-signal training transfers across environments using 25% of tokens
anirudhg9119 · x · 2026-10-14
Experimental takeaways (with results table):
- MIRA-AC shows models can learn from their own proxy hill-climbing signals, improving gold outcomes across several autoresearch environments (e.g., Symbolic Regression Solved from 50 to 60–65) and transferring to the held-out BNLearn environment (Solved 84, SHD gain 0.9869).
- Decision-level credit assignment: policy optimization focuses only on decisions steering research direction — roughly 25% of decision and execution tokens per episode, i.e., training WHAT not HOW.
- Full research episodes are reconstructed and decision-level advantages computed from downstream outcomes; the policy loss updates only the meta-reasoner's reasoning and work orders, keeping executor, tools, and context curation fixed.
More from coding & agent
- TV noise crashes voice agents from 38.7% to 8.6%; full duplex is the fix — SIGKITTEN · 2026-10-14
- TypeSafe AI's Jev Claims 100x Faster Decisions, Can Replace 9 LLM Call Types — Arindam_1729 · 2026-10-14
- Gemini 4 Argon High tested with 'Horse prompt' via OpenCode V2 and Antigravity agent — ZealousidealTown1974 · 2026-10-14
- Buoy benchmark: Claude runs apps 6x faster and cheaper with MCP tools than none — LovesWorkin · 2026-10-14
- LMArena Launches Code Arena for WebDev Model Benchmarking — arena · 2026-10-14
- msgblast: Open-Source Mac App to Prompt Multiple AI Agents at Once — msg · 2026-10-14