frontier models
Two Million Parameters Read the Verification Signal Already Inside
HSRM ranks candidate math solutions from frozen-generator hidden states instead of re-reading their full text.
Summary
HSRM ranks candidate math solutions from frozen-generator hidden states instead of re-reading their full text.
Across four math benchmarks, the hidden-state verifier matched or beat a 55-million-parameter text energy model in 15 of 16 generator-dataset settings while using about two million parameters. It relies on access to the generator's internal representations.
Why it matters
HSRM ranks candidate math solutions from frozen-generator hidden states instead of re-reading their full text.
Limits and context
No additional limitation was separately recorded.
Key claims
HSRM ranks candidate math solutions from frozen-generator hidden states instead of re-reading their full text.
Evidence: source-2026-09-01-021
Sources
- arXiv preprint 2608.30841arXiv · primary research
Corrections
No corrections have been recorded for this story.