LLM-as-a-Verifier: A General-Purpose Verification Framework
TL;DR Highlight
LLM의 토큰 확률 분포를 활용해 discrete 점수 대신 continuous 점수를 뽑아내면, 추가 학습 없이 코딩·로봇·의료 에이전트 평가 정확도를 SOTA로 끌어올릴 수 있다.
Who Should Read
여러 에이전트 출력 중 최선을 고르는 Best-of-N 선택 로직을 구현하는 백엔드/ML 개발자. LLM Judge의 동점(tie) 문제나 coarse한 점수에 불만을 느끼는 AI 에이전트 개발자.
Core Mechanics
- 기존 LLM Judge는 모델이 생성한 가장 높은 확률의 토큰 하나만 점수로 쓰는데, 이 방식은 복잡한 솔루션 비교 시 27%나 동점이 발생해 변별력이 없다.
- LLM-as-a-Verifier는 점수 토큰의 logit 분포 전체에 대한 기댓값(expectation)을 계산해 continuous 점수를 만든다 — 이 방식만으로도 동점률이 0%로 떨어진다.
- Verification을 3가지 축으로 스케일할 수 있다: (1) score granularity(점수 해상도를 1단계→20단계로), (2) repeated evaluation(같은 쌍을 K번 반복 채점해 분산 감소), (3) criteria decomposition(평가 기준을 세부 항목으로 분해해 프롬프트 편향 감소).
- 후보 N개 중 최선을 고를 때 O(N²) 비교가 필요한 round-robin 대신, 'Probabilistic Pivot Tournament(PPT)'를 써서 O(Nk²)로 줄인다 — ring pass로 positional bias도 자동 제거.
- 추가 학습 없이 Gemini 2.5 Flash를 verifier로 쓰면 Terminal-Bench V2 86.5%, SWE-Bench Verified 78.2%, RoboRewardBench 87.4%, MedAgentBench 73.3%로 각 벤치마크 SOTA 달성.
- Verifier 점수가 에이전트의 태스크 진행도와 강하게 상관(Spearman VOC 0.848)되어, 실시간 진행 상황 모니터링 및 RL의 dense reward로도 활용 가능하다.
Evidence
- Terminal-Bench V2에서 discrete judge는 동점률 26.7%(k=1)인 반면, continuous verifier는 k=1에서도 동점률 0% — k=1 verifier(74.7%)가 k=16 judge(74.7%)와 동등한 성능.
- Granularity G=1→20으로 키우자 verification accuracy 73.1%→77.5%, SNR 0.775→0.799로 상승. query-optimize 사례에서 discrete judge는 100회 중 88회 동점, verifier G=20은 77회 정답 정확히 구분.
- RoboRewardBench에서 훈련된 로봇 전용 reward model(RoboReward-8B 81.4%, Robometer-4B 78.8%)을 제로샷으로 87.4%로 초과 달성.
- RL에 dense reward로 사용 시: LIBERO 로봇 태스크에서 DSRL-SAC 기준 sparse reward 대비 1.8× 샘플 효율 향상 + 최종 성공률 0.69→0.76. Qwen3-8B + GRPO로 MATH 학습 시 약 1.1× 샘플 효율 향상.
How to Apply
- Best-of-N 코딩 에이전트를 만들 때: GPT-5.5나 Claude로 N개 후보 생성 후, Gemini 2.5 Flash의 logprobs API로 각 후보를 G=20 granularity로 채점하고 PPT 알고리즘으로 최선 선택 — SWE-Bench 기준 Pass@1 76.1%→78.2%로 올라간다.
- logprobs를 제공하지 않는 GPT-5.5·Claude 같은 closed API를 verifier로 쓰고 싶다면: 먼저 closed 모델로 reasoning trace를 생성하고, 그 reasoning을 Gemini 2.5 Flash에 넘겨 logprobs를 추출하는 2단계 파이프라인을 쓰면 된다 — discrete 방식 대비 +5.2p 정확도, 동점 0%.
- RL 학습에서 보상 설계가 어렵거나 sparse reward로 학습이 안 될 때: 매 rollout 끝에 verifier로 단계별 점수를 계산해 r_t = r_env + λ·ρ_t 형태로 shaped reward를 추가하면 추가 reward model 학습 없이 샘플 효율이 올라간다.
Code Example
Terminology
Related Papers
Show HN: Mindwalk – Replay coding-agent sessions on a 3D map of your codebase
Claude Code나 Codex 같은 AI 코딩 에이전트가 세션 중 코드베이스의 어떤 파일을 탐색하고 수정했는지를 3D 지도 형태로 시각화해서 재생해주는 로컬 도구다. 에이전트가 작업을 어떻게 이해했는지 한눈에 파악할 수 있다.
Ghost Font: A font that humans can read but AI cannot
움직임(모션)을 이용해 글자를 표현해서 AI 모델이 정적 이미지 분석으로는 메시지를 해독하지 못하게 막는 실험적 프로젝트인데, 커뮤니티에서는 이미 GPT-5.6, Claude Opus 등으로 해독에 성공한 사례가 속출해 실효성 논쟁이 뜨겁다.
GPT-5.6, Grok 4.5, Claude, and Muse Spark build the same 4 apps
12개 LLM 모델에게 레이캐스터 미로, 루빅스 큐브, 계산기, Game of Life 앱을 각각 5번씩 만들게 해서 성공률·비용·속도를 비교한 실전 벤치마크다. GPT-5.6 Sol이 전반적으로 가장 일관된 결과를 냈고, Grok 4.5는 가성비 면에서 눈에 띄었다.
Benchmarking coding agents on Databricks' multi-million line codebase
Databricks가 자사 실제 코드베이스를 기반으로 여러 AI 코딩 에이전트의 성능과 비용을 직접 측정했고, 모델 토큰 가격과 실제 태스크 비용이 전혀 다르다는 점, 그리고 오픈소스 모델이 이제 최상위 수준에 도달했다는 점을 확인했다.
Estimating Uncertainty from Reasoning: A Large-Scale Study of Multi- and Crosslingual MCQA Performance in LLMs
LLM이 저자원 언어 질문을 받을 때 영어로 추론하게 하면 불확실성 추정 성능이 고자원 언어 수준으로 올라온다.
Does code cleanliness affect coding agents? A controlled minimal-pair study
SonarSource 연구팀이 코드 품질(cleanliness)이 AI 코딩 에이전트의 작업 성공률에는 영향을 주지 않지만, 토큰 사용량을 7~8% 줄이고 파일 재방문을 34% 감소시킨다는 사실을 통제 실험으로 밝혔다.
Related Resources
Original Abstract (Expand)
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.