Self-Improving VLM Judges Without Human Annotations

• URL: http://arxiv.org/abs/2512.05145v1
• Date: Tue, 02 Dec 2025 20:52:19 GMT
• Title: Self-Improving VLM Judges Without Human Annotations
• Authors: Inna Wanyin Lin, Yushi Hu, Shuyue Stella Li, Scott Geng, Pang Wei Koh, Luke Zettlemoyer, Tim Althoff, Marjan Ghazvininejad,
• Abstract summary: We present a framework to self-train a VLM judge model without any human preference annotations, using only self-synthesized data.<n>Our method improves a Llama-3.2-11B multimodal judge from 0.38 to 0.51 in overall accuracy on Multimodal RewardBench.<n>The overall strength of these human-annotation-free results suggest the potential for a future self-judge that evolves alongside rapidly improving VLM capabilities.
• Score: 74.29324865147838
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Effective judges of Vision-Language Models (VLMs) are crucial for model development. Current methods for training VLM judges mainly rely on large-scale human preference annotations. However, such an approach is costly, and the annotations easily become obsolete as models rapidly improve. In this work, we present a framework to self-train a VLM judge model without any human preference annotations, using only self-synthesized data. Our method is iterative and has three stages: (1) generate diverse multimodal instruction-response pairs at varying quality levels, (2) generate reasoning traces and judgments for each pair, removing the ones that do not match our expected quality levels, and (3) training on correct judge answers and their reasoning traces. We evaluate the resulting judge on Multimodal RewardBench and VL-RewardBench across domains: correctness, preference, reasoning, safety, and visual question-answering. Our method improves a Llama-3.2-11B multimodal judge from 0.38 to 0.51 in overall accuracy on VL-RewardBench, often outperforming much larger models including Llama-3.2-90B, GPT-4o, and Claude 3.5 Sonnet, with particularly strong gains in general, hallucination, and reasoning dimensions. The overall strength of these human-annotation-free results suggest the potential for a future self-judge that evolves alongside rapidly improving VLM capabilities.

Related papers

• JudgeRLVR: Judge First, Generate Second for Efficient Reasoning [20.448286296459344]
  Reinforcement Learning with Verifiable Rewards (RLVR) has become a standard paradigm for reasoning in Large Language Models.<n>In this paper, we argue that discriminative capability is a prerequisite for efficient generation.<n>We propose JudgeRLVR, a two-stage judge-then-generate paradigm.
  arXiv  Detail & Related papers  (2026-01-13T11:47:42Z)
• DatBench: Discriminative, Faithful, and Efficient VLM Evaluations [17.234602646114997]
  Empirical evaluation serves as the primary compass guiding research progress in foundation models.<n>We propose three desiderata that evaluations should satisfy: faithfulness to the modality and application, discriminability between models of varying quality, and efficiency in compute.<n>We release DatBench-Full, a cleaned evaluation suite of 33 datasets spanning nine VLM capabilities, and DatBench, a discriminative subset that achieves 13x average speedup.
  arXiv  Detail & Related papers  (2026-01-05T18:07:51Z)
• Self-Consistency as a Free Lunch: Reducing Hallucinations in Vision-Language Models via Self-Reflection [71.8243083897721]
  Vision-language models often hallucinate details, generating non-existent objects or inaccurate attributes that compromise output reliability.<n>We present a novel framework that leverages the model's self-consistency between long responses and short answers to generate preference pairs for training.
  arXiv  Detail & Related papers  (2025-09-27T10:37:11Z)
• Multimodal RewardBench: Holistic Evaluation of Reward Models for Vision Language Models [82.92771279118888]
  We introduce Multimodal RewardBench, an expert-annotated benchmark for evaluating multimodal reward models.<n>Our dataset comprises 5,211 annotated (prompt, chosen response, rejected response) triplets collected from various vision-language models.<n>We find that even the top-performing models, Gemini 1.5 Pro and Claude 3.5 Sonnet, achieve only 72% overall accuracy.
  arXiv  Detail & Related papers  (2025-02-20T01:48:13Z)
• VHELM: A Holistic Evaluation of Vision Language Models [75.88987277686914]
  We present the Holistic Evaluation of Vision Language Models (VHELM) VHELM aggregates various datasets to cover one or more of the 9 aspects: visual perception, knowledge, reasoning, bias, fairness, multilinguality, robustness, toxicity, and safety. Our framework is designed to be lightweight and automatic so that evaluation runs are cheap and fast.
  arXiv  Detail & Related papers  (2024-10-09T17:46:34Z)
• Direct Judgement Preference Optimization [79.54459973726405]
  We train large language models (LLMs) as generative judges to evaluate and critique other models' outputs.<n>We employ three approaches to collect the preference pairs for different use cases, each aimed at improving our generative judge from a different perspective.<n>Our model robustly counters inherent biases such as position and length bias, flexibly adapts to any evaluation protocol specified by practitioners, and provides helpful language feedback for improving downstream generator models.
  arXiv  Detail & Related papers  (2024-09-23T02:08:20Z)
• MJ-Bench: Is Your Multimodal Reward Model Really a Good Judge for Text-to-Image Generation? [59.7772329962047]
  We introduce MJ-Bench, a novel benchmark which incorporates a comprehensive preference dataset to evaluate multimodal judges. Specifically, we evaluate a large variety of multimodal judges including smaller-sized CLIP-based scoring models, open-source VLMs, and close-source VLMs. Experiments reveal that close-source VLMs generally provide better feedback, with GPT-4o outperforming other judges in average.
  arXiv  Detail & Related papers  (2024-07-05T20:03:16Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.