MIRA-AC initializes a shared actor–critic from the pretrained value model
anirudhg9119 · x · 2026-10-14
Training design: predicting value alone isn't enough. MIRA-AC initializes a shared actor–critic from the pretrained value model — the critic predicts remaining return, the actor chooses the next investigation, so the model learns both to value progress and to decide what to study. The critic is trained with a KL loss toward realized episode return; step-level generative critics reportedly beat token-level scalar value heads and every other critic design tested.
More from Research
- China's GigaAI Opens First Provincial-Level General World Model Lab in Beijing — pstAsiatech · 2026-10-14
- Mosaic model engineers cell states to expose hidden disease variants, opening chromatin at 28% of AMD loci vs 5-7% at rest — anshulkundaje · 2026-10-14
- NeurIPS paper: visual in-context learning lifts VLM accuracy by up to 27% — cindy_x_wu · 2026-10-14
- Pythia team releases protein-RNA binding atlas preprint probing why the model predicts — anshulkundaje · 2026-10-14
- pqnld: A Rust Service That Turns Your Existing LLM Into a Closed-Set Decision Model via Logprobs — Difficult_Ad_2778 · 2026-10-14
- Mixed-precision, mixed-accelerator samplers cut RL training cost with off-policy correction — xiuyu_l · 2026-10-14