More verifier compute improves scoring, not missing evidence
svk_roy · reddit · 2026-07-22
More verifier compute improves scoring, not missing evidence
This Reddit thread argues that spending more compute on verification can produce finer scores, more consistent judgments, and clearer local reasoning — but it still cannot supply evidence the verifier never observed.
The author uses a simple numeric example to make the point: a stronger checker may be able to say that 24 is closer to the true answer 25 than 20 is, yet it still does not prove that the verifier is checking the right property in the first place. The discussion is about the limits of verification, not just the quality of the scoring function.
More from Research
- Token Collapse: Study Finds Graph Soft Prompting Tokens in LLMs Fail to Encode Structure — Sapienza · 2026-10-27
- Terence Tao's SAIR Foundation revamps Mathathon with $1.5M compute per team — _sathvikr · 2026-10-27
- Mathathon 2.0 launches with Terence Tao's SAIR, $1.5M in compute for teams — _sathvikr · 2026-10-27
- The 'sweet lesson' is real: two kinds of ML problems need opposite strategies — mike64_t · 2026-10-27
- PixGR: Pixel-space generative decoding for fast feed-forward 3D reconstruction — kwangmoo_yi · 2026-10-27
- UniAction: unified action conditioning lets one world model handle mixed action types — ZeYanjie · 2026-10-27