Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TIM-PRM: Verifying multimodal reasoning with Tool-Integrated PRM

Peng Kuang, Xiangxiang Wang, Wentao Liu, Jian Dong, Kaidi Xu

TL;DR

This paper tackles the reliability of multimodal reasoning by addressing visual grounding failures in verification. It introduces TIM-PRM, an agentic, tool-integrated verifier that uses Independent Question Asking to decouple perception from hypothesis and ground verification in external evidence. Through a curated dataset and extensive experiments on VisualProcessBench, TIM-PRM 8BSET achieves state-of-the-art open-weight performance and superior error localization (FISI) compared to larger baselines, demonstrating improved interpretability and robustness. The approach offers a scalable path toward more trustworthy multimodal reasoning in high-stakes settings.

Abstract

Multimodal Large Language Models (MLLMs) have achieved impressive performances in mathematical reasoning, yet they remain vulnerable to visual hallucinations and logical inconsistencies that standard outcome-based supervision fails to mitigate. While Process Reward Models (PRMs) promise step-by-step verification, current approaches typically operate as scalar scorers or generative critics that suffer from sycophancy, blindly validating the flawed hypotheses rather than grounding them in visual reality. To bridge this gap, we introduce TIM-PRM (Tool-Integrated Multimodal PRM), a novel agentic framework that transforms verification from a passive classification task into an active, tool-augmented investigation. TIM-PRM is trained to explicitly plan verification strategies and utilizes a mechanism of Independent Question Asking to query evidence via external tools, effectively decoupling verification from the reasoning context to eliminate confirmation bias. We instantiate this method by curating a high-quality dataset of tool-integrated verification trajectories. Extensive experiments on VisualProcessBench demonstrate that our 8B parameter model surpasses existing open-source multimodal PRMs, significantly outperforming much larger models like Qwen2.5-72B and InternVL-78B, while offering interpretable insights into the verification process.

TIM-PRM: Verifying multimodal reasoning with Tool-Integrated PRM

TL;DR

This paper tackles the reliability of multimodal reasoning by addressing visual grounding failures in verification. It introduces TIM-PRM, an agentic, tool-integrated verifier that uses Independent Question Asking to decouple perception from hypothesis and ground verification in external evidence. Through a curated dataset and extensive experiments on VisualProcessBench, TIM-PRM 8BSET achieves state-of-the-art open-weight performance and superior error localization (FISI) compared to larger baselines, demonstrating improved interpretability and robustness. The approach offers a scalable path toward more trustworthy multimodal reasoning in high-stakes settings.

Abstract

Multimodal Large Language Models (MLLMs) have achieved impressive performances in mathematical reasoning, yet they remain vulnerable to visual hallucinations and logical inconsistencies that standard outcome-based supervision fails to mitigate. While Process Reward Models (PRMs) promise step-by-step verification, current approaches typically operate as scalar scorers or generative critics that suffer from sycophancy, blindly validating the flawed hypotheses rather than grounding them in visual reality. To bridge this gap, we introduce TIM-PRM (Tool-Integrated Multimodal PRM), a novel agentic framework that transforms verification from a passive classification task into an active, tool-augmented investigation. TIM-PRM is trained to explicitly plan verification strategies and utilizes a mechanism of Independent Question Asking to query evidence via external tools, effectively decoupling verification from the reasoning context to eliminate confirmation bias. We instantiate this method by curating a high-quality dataset of tool-integrated verification trajectories. Extensive experiments on VisualProcessBench demonstrate that our 8B parameter model surpasses existing open-source multimodal PRMs, significantly outperforming much larger models like Qwen2.5-72B and InternVL-78B, while offering interpretable insights into the verification process.

Paper Structure

This paper contains 23 sections, 2 equations, 1 figure, 6 tables.

Table of Contents

  1. Introduction
  2. Related works
  3. Revisiting Verification of Multimodal Reasoning
  4. Failure Modes in Visual Reasoning Verification
  5. Methodology
  6. Preliminary
  7. Step-wise Verification with Tool Integration
  8. Debiasing Sycophantic Verification via Independent Question Asking
  9. Sample Upweighting
  10. Experiments
  11. Experiment Setup
  12. Benchmarks.
  13. Baselines.
  14. Training data.
  15. Implementation Details.
  16. ...and 8 more sections

Figures (1)

  • Figure 1: Analysis of Data Filtering Consistency.