Weaver: Shrinking the Generation-Verification Gap with Weak Verifiers
summary
Weaver aggregates many small, noisy verifiers (reward models and LM judges) without labels to close the generation‑verification gap, matching o3‑mini on MATH500, GPQA Diamond and MMLU Pro and boosting best‑of‑n success up to 27.8% above majority. The ensemble weights are learned via weak supervision, then distilled into a compact 400M model that preserves up to 98.7% of ensemble accuracy while cutting verification FLOPs by ~99.97%; Weaver distilled enables cheap online verification.
AI-written summary · about
chapters (10)
- 0:00Inference compute scaling
- 1:45Generation-verification gap
- 3:52Weak supervision for verifiers
- 5:37Aggregating weak verifiers
- 7:01Performance gains and comparisons
- 8:45Verification compute strategies
- 10:50Reducing class performance gaps
- 12:14Distilled verification with Weaver
- 14:18Future directions
- 15:41Q&A
related talks
27mThe Judge Problem: Agentic Verifier is Rotting in Production
31mMaking the most of test-time compute in LLMs
30mUnfamiliar Terrain
10mWeaver: Efficient Multi-LLM Serving with Attention Offloading
15mA benchmark for vericoding: formally verified program synthesis
20mIncremental Verification of Neural Networks
46mInteractive Proofs for Verifying Machine Learning
01h 08mLean: Machine-Checked Mathematics and Verified Programming, Past and Future