MIRA-AC: decomposition plus learned value priors fix delayed credit assignment for research agents
anirudhg9119 · x · 2026-10-14
A long thread introducing MIRA-AC, a training recipe for autoresearch agents tackling delayed credit assignment: an experiment's immediate outcome often diverges from the long-term value of the decision to run it.
Three key ideas:
- Decomposition: split WHAT to investigate (meta-reasoner) from HOW to execute (fresh executor + persistent repository), so one research decision = one action and its full investigation = one transition, avoiding optimization over thousands of execution tokens.
- Value learning: a generative critic predicts remaining return at each decision boundary, becoming a transferable prior for partial progress.
- RL: an actor–critic (MIRA-AC) initialized from the pretrained value model assigns delayed credit and improves the research policy.
The author reports gold-metric gains across multiple autoresearch environments and transfer to a held-out environment excluded from policy training.
More from coding & agent
- Anthropic's internal VCS 'Scion' revealed: built for huge monorepos and agent swarms — nlarusstone · 2026-10-14
- PeerBrush: open-source image editor lets you and an AI co-edit via MCP — Deftroyy · 2026-10-14
- TV noise crashes voice agents from 38.7% to 8.6%; full duplex is the fix — SIGKITTEN · 2026-10-14
- Six DNS checks that filter junk emails for agents — without sending a single message — S_B_B_B_K · 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