LLM-as-a-Verifier: A General-Purpose Verification Framework

Hugging Face Daily Papers Papers

Summary

LLM-as-a-Verifier introduces a probabilistic verification framework that computes continuous scores from LLM logits, scaling across granularity, repeated evaluation, and criteria decomposition. It achieves state-of-the-art results on multiple agentic benchmarks and provides dense feedback for RL.

Scaling pre-training, post-training, and test-time compute have become the central paradigms for improving the capabilities of LLMs. In this work, we identify verification, the ability to determine the correctness of a solution, as a new scaling axis. To unlock this and demonstrate its effectiveness, we introduce LLM-as-a-Verifier, a general-purpose verification framework that provides fine-grained feedback for agentic tasks without requiring additional training. Unlike standard LM judges that prompt LLMs to produce discrete scores for candidate solutions, LLM-as-a-Verifier computes the expectation over the distribution of scoring token logits to generate continuous scores. This probabilistic formulation enables verification to scale along multiple dimensions: (1) score granularity, (2) repeated evaluation, and (3) criteria decomposition. In particular, we show that scaling the scoring granularity leads to better separation between positive and negative solutions, resulting in more calibrated comparisons. Moreover, scaling repeated evaluation and criteria decomposition consistently lead to additional gains in verification accuracy through variance and complexity reduction. We further introduce a cost-efficient ranking algorithm for selecting the best solution among candidates using the verifier's continuous scores. LLM-as-a-Verifier achieves state-of-the-art performance on Terminal-Bench V2 (86.5%), SWE-Bench Verified (78.2%), RoboRewardBench (87.4%), and MedAgentBench (73.3%). Beyond verification, the fine-grained signals from LLM-as-a-Verifier can also serve as a proxy for estimating task progress. We build an extension for Claude Code, enabling developers to monitor and improve their own agentic systems. Finally, we show that LLM-as-a-Verifier can provide dense feedback for RL, improving the sample efficiency of SAC and GRPO on robotics and mathematical reasoning benchmarks.
Original Article
View Cached Full Text

Cached at: 07/07/26, 06:42 AM

Paper page - LLM-as-a-Verifier: A General-Purpose Verification Framework

Source: https://huggingface.co/papers/2607.05391

Abstract

LLM-as-a-Verifier introduces a probabilistic verification framework that scales across multiple dimensions to improve solution correctness assessment and agent performance across various benchmarks.

Scaling pre-training, post-training, and test-time compute have become the central paradigms for improving the capabilities of LLMs. In this work, we identifyverification, the ability to determine the correctness of a solution, as a new scaling axis. To unlock this and demonstrate its effectiveness, we introduceLLM-as-a-Verifier, a general-purposeverificationframework that provides fine-grained feedback foragentic taskswithout requiring additional training. Unlike standard LM judges that prompt LLMs to produce discrete scores for candidate solutions,LLM-as-a-Verifiercomputes the expectation over the distribution ofscoring token logitsto generatecontinuous scores. This probabilistic formulation enablesverificationto scale along multiple dimensions: (1)score granularity, (2)repeated evaluation, and (3)criteria decomposition. In particular, we show that scaling the scoring granularity leads to better separation between positive and negative solutions, resulting in more calibrated comparisons. Moreover, scalingrepeated evaluationandcriteria decompositionconsistently lead to additional gains inverificationaccuracy through variance and complexity reduction. We further introduce acost-efficient ranking algorithmfor selecting the best solution among candidates using the verifier’scontinuous scores.LLM-as-a-Verifierachieves state-of-the-art performance onTerminal-Bench V2(86.5%),SWE-Bench Verified(78.2%),RoboRewardBench(87.4%), andMedAgentBench(73.3%). Beyondverification, the fine-grained signals fromLLM-as-a-Verifiercan also serve as a proxy for estimating task progress. We build an extension for Claude Code, enabling developers to monitor and improve their own agentic systems. Finally, we show thatLLM-as-a-Verifiercan provide dense feedback for RL, improving the sample efficiency ofSACandGRPOon robotics andmathematical reasoning benchmarks.

View arXiv pageView PDFProject pageGitHub409Add to collection

Get this paper in your agent:

hf papers read 2607\.05391

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2607.05391 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2607.05391 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2607.05391 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Logic-Regularized Verifier Elicits Reasoning from LLMs

arXiv cs.CL

Introduces LoVer, an unsupervised verifier that uses logical rules (negation consistency, intra-group and inter-group consistency) to improve LLM reasoning without labeled data, achieving performance close to supervised verifiers on reasoning benchmarks.

LLM-as-a-Tutor: Policy-Aware Prompt Adaptation for Non-Verifiable RL

Hugging Face Daily Papers

LLM-as-a-Tutor introduces a framework that extends LLM's role from judge to tutor by dynamically adjusting prompt difficulty through pairwise comparison and constraint addition, improving instruction-following performance in reinforcement learning.