Review Rarely Beats the Stronger Model: Separating Verification from Capability Transfer in LLM Pipelines
We show that this confuses verification, which adds accuracy beyond either model alone, with capability transfer, where a stronger reviewer supplies answers it could have given on its own.
ProofPaper ↗
Key points
- Qwen3-8B reviewing Llama-3.1-8B in math adds 33.2 points, yet Qwen3-8B alone is 6.4 points more accurate than the result.
- Reviewing another model's answer is usually judged by its gain over the generator.
- Measured against the stronger model alone, across 48 pairs of 12 open-weight models (3B to 72B) in math, knowledge, code and reasoning, review beats it in only 9 of 96 review directions, and does worse the larger the capability gap.
- Studies of verification should report the stronger model alone, not only the gain over the generator.
Sources (1)
- [1]Review Rarely Beats the Stronger Model: Separating Verification from Capability Transfer in LLM PipelinesarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 11, 08:26 PM
We show that this confuses verification, which adds accuracy beyond either model alone, with capability transfer, where a stronger reviewer supplies answers it could have given on its own.
Qwen3-8B reviewing Llama-3.1-8B in math adds 33.2 points, yet Qwen3-8B alone is 6.4 points more accurate than the result.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 11, 2026openai/openai-python v3.29.0
- Oct 11, 2026Agent Speedrun — Elizabeth Fuentes Leone & Sandhya Subramani, AWS
- Oct 10, 2026KV4-RL: Unlocking Faster Asynchronous Rollouts for RLVR with INT4 KV Cache Quantization
- Oct 10, 2026EviScene: Beyond "Is It AI?" for Agentic Image Factuality Assessment
- Oct 10, 2026LightVAE: Towards Compact and Efficient Video Autoencoders
- Oct 9, 2026Work2Skill: Capturing Human GUI Experience for Reusable Agent Skill Induction
