Training the top 0.1% coordinates by Fisher info recovers ~85% of full PG gains
YouJiacheng · x · 2026-10-15
RichardYRLi summarizes a paper's core insight: on-policy, the score has zero mean, so |gi| <= sqrt(E[Var(r|x)]) sqrt(Fi) for any reward, where Fi depends only on the model.
That means you can rank coordinates before seeing rewards: training only the top 0.1% by initial Fisher information recovers 85% of the full gain on on-policy policy-gradient methods. YouJiacheng called it "very interesting."
More from Research
- Models may detect AI-generated text by fingerprint alone, and eval sandboxes could be compromised — scaling01 · 2026-10-15
- UW Protein Design wins AI for Science grant to design plastic-degrading proteins with natural language — PangWeiKoh · 2026-10-15
- MIT simulates an AI-Symbiont in LLMs: aligned interventions help, misaligned ones hurt — patpat_mit · 2026-10-15
- JCIM paper systematically evaluates unimodal vs multimodal fusion for molecular property prediction — rbhar90 · 2026-10-15
- Mandelbrot in pure Lean: Claude renders figures with Hershey fonts, no Python — geoffreyirving · 2026-10-15
- TUM team simulates brain tumor mass effect with springs, ~100x faster than PDE inversion — maier_ak · 2026-10-15