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.

PaperBig Techcs.AIarXiv:2607.05391v2 ↗

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

Stanford University · UC Berkeley · NVIDIA Research

Research Digest·7 July 2026·3 min read

Read the paper →

Thread:Agent Rule Compliance

What they did

The authors propose a verification framework that prompts an LLM to generate a single scoring token (e.g., "A" through "F") and then computes the expected score as the weighted sum over the logit distribution for that token. This yields continuous scores instead of discrete grades. They scale verification along three dimensions: (1) score granularity—by increasing the number of score levels (e.g., from 5 to 15); (2) repeated evaluation—averaging scores over multiple independent runs; and (3) criteria decomposition—breaking a rubric into subcriteria and aggregating sub-scores.

The framework is evaluated on four agentic benchmarks: Terminal-Bench V2 (command-line tasks), SWE-Bench Verified (software engineering), RoboRewardBench (robotics reward modeling), and MedAgentBench (medical agent). They also build a Claude Code extension for real-time monitoring and test the verifier's signals as dense rewards for reinforcement learning algorithms (SAC, GRPO) in robotics and math domains.

Key findings

LLM-as-a-Verifier achieves state-of-the-art on all four benchmarks: Terminal-Bench V2 86.5%, SWE-Bench Verified 78.2%, RoboRewardBench 87.4%, MedAgentBench 73.3%—surpassing prior discrete-scoring judges by 3–8 absolute points.
Increasing scoring granularity from 5 to 15 levels reduces tie rates between candidate solutions by 40% and improves rank correlation with ground truth by 12%.
Repeated evaluation (e.g., 10 runs) reduces variance and yields consistent accuracy gains of 1–3%, while criteria decomposition (3–5 subcriteria) adds another 2–4% improvement.
The continuous scores serve as effective dense rewards: SAC with verifier rewards achieves 23% higher success on manipulation tasks, and GRPO with verifier feedback improves math reasoning accuracy by 15% over baseline.

Why it matters

This work identifies verification as a distinct scaling axis for LLM agents, separate from pre-training, post-training, or test-time compute. The probabilistic scoring method is training-free and general-purpose, meaning it can be applied to any LLM-based agent system to improve selection accuracy, provide richer feedback for RL, and offer interpretable progress signals. This could reduce the need for expensive human verification or hand-crafted reward functions in complex agentic domains.

Caveats

The framework requires access to the LLM's logits for the scoring token, which may not be available through all API endpoints (e.g., closed-source models). The approach is evaluated only on benchmarks and may not generalize to highly open-ended tasks with no predefined rubric. The repeated evaluation and criteria decomposition increase inference cost linearly, which could be a practical limitation. Finally, while the verifier provides dense feedback, it is still noisy and may not replace environment-based rewards where those exist.

The work fits into a growing line of research on using LLMs as judges or verifiers, but departs from prior work by leveraging continuous scores derived from logit distributions rather than discrete Likert-scale outputs. This probabilistic treatment is reminiscent of calibration techniques in classification and could be extended to other scoring scenarios. The demonstration that verifier signals can bootstrap RL is particularly interesting, as it suggests a path to self-improving agents without hand-crafted reward models. However, the paper does not explore how the verifier itself could be improved through training or whether scaling to larger LLMs yields diminishing returns. Future work might investigate combining multiple verifiers or using the framework for active learning.

ZOTPAPER