Probabilistic scoring unlocks verification as a new scaling axis for LLMs

By computing continuous scores from token logit distributions, the framework improves verification accuracy across coding, robotics, and medical benchmarks without additional training.

Big Tech

Jacky Kwok · Shulu Li · Pranav Atreya · Yuejiang Liu · Yixing Jiang · Chelsea Finn · +3 more

Stanford University · UC Berkeley · NVIDIA Research

Research Digest·6 July 2026·3 min read

Thread:Agent Rule Compliance

In one line

Scoring token logit expectations, repeated evaluations, and decomposed criteria improve LLM verification accuracy and achieve state-of-the-art agentic benchmark results.

Before you read this

Two ideas this paper builds on

verificationChecking outputs for correctness before accepting them, often to catch errors or reward hacking.277 papers →reinforcement learningA machine learning method where an agent learns to make sequences of decisions by receiving rewards or penalties for its actions.239 papers →

Read it as

LLM-as-a-Verifier replaces discrete judge scores with continuous scores computed from the expectation over scoring token logits, requiring no additional training. The work shows scales along three dimensions: score granularity, repeated evaluation, and criteria decomposition. Continuous scoring reduces ties and improves separation between correct and incorrect solutions, while repeated evaluation and decomposed criteria further raise verification accuracy. It reports state-of-the-art results on Terminal-Bench V2 (86.5%), SWE-Bench Verified (78.2%), RoboRewardBench (87.4%), and MedAgentBench (73.3%).

Beyond ranking candidate solutions, the fine-grained signals act as a progress proxy in Claude Code and Codex extensions and as a dense reward for , improving sample efficiency for SAC and GRPO on robotics and mathematical reasoning. This matters because verification is positioned as an independent scaling axis, not just a component of generation.

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

verification continuous scoring verification scaling fine-grained feedback reinforcement learning