research: verifier head-to-head v2 — cross-model (Qwen + Llama), Wilson CIs - #110
Merged
Conversation
…on CIs Upgrades the inference-receipt evidence to paper grade: two model families, 150 gsm8k prompts (~27k tokens each), greedy decode, a real argmax-verifier protocol measuring BOTH error directions, and Wilson 95% CIs on every low-margin rate (kills the small-N + one-model caveats from the first run). 1x H100, vLLM 0.23.0. Low-margin = top1-top2 gap < 0.3 nats. true-KV verifier is 0 in every cell by construction. - gap masking (re-feed drift >= the top1-top2 gap, i.e. a swap hides in the verifier's own noise): Qwen 27.6% (CI 24.5-30.9), Llama 22.8% (CI 20.6-25.3) — ~1 in 4 decision tokens, both families, tight intervals. - argmax forgery-accept (re-feed argmax == decode runner-up): Qwen 4.6% (CI 3.3-6.4), Llama 3.3% (CI 2.5-4.5). argmax false-reject ~3.5-4.6%. - logit false-reject at tau=0.01: Qwen 8.6%, Llama 11.5%. So replay-from-re-feed verification pays a false-reject cost on genuine output AND has a forgery-acceptance hole at the tokens that matter, cross-family with tight CIs; thaw's true-KV inference receipt removes both. Honest scope: a deployed-verifier (DiFR/TOPLOC) head-to-head is the remaining step.
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Paper-grade upgrade of the inference-receipt evidence: two model families, 150 gsm8k prompts (~27k tokens each), greedy decode, a real argmax-verifier protocol measuring both error directions, and Wilson 95% CIs on every low-margin rate — directly killing the small-N and one-model caveats from the first run. 1× H100, vLLM 0.23.0.
true_kvverifier is 0 in every cell by construction.~1 in 4 decision tokens have re-feed drift exceeding the gap that distinguishes the real token from the runner-up — cross-family, tight CIs. So re-feed verification pays a false-reject cost on genuine output and has a forgery-acceptance hole at the tokens that matter; thaw's true-KV inference receipt removes both.
benchmarks/verifier_headtohead.py(v2, multi-model + CIs) +site/receipts/2026-06-25_h100_verifier_headtohead_v2.json. Honest scope: a deployed-verifier (DiFR/TOPLOC codebase) head-to-head is the remaining step toward a full paper.