Multi-Agent AI Wins Physics Challenge by Interpreting Visuals

Jiseok Kwak, Suhyeon Jo, Taewoo Kim, Yeongmin Kim, Byeonghu Na, Il-chul Moon· July 27, 2026 View original

Summary

This report details the 1st-place solution for the ICML 2026 AI4Math Track 3 Challenge, SeePhys Pro, which involved answering college-level physics questions with visual components. The approach uses a two-stage framework: visual information extraction to convert figures into text, followed by a multi-agent debate system orchestrating three solvers.

This technical report outlines the winning strategy for the ICML 2026 AI4Math Track 3 Challenge, named SeePhys Pro. The challenge required AI systems to solve college-level physics problems where critical information was often embedded within images rather than solely in text. Large Language Models (LLMs) typically struggle when visual data is paramount, creating a significant "modality gap." The winning pipeline addresses this by employing a two-stage framework. First, a visual information extraction stage translates the content of figures into solver-readable text, effectively bridging the modality gap. Second, a sophisticated reasoning stage orchestrates three distinct solvers through a multi-agent debate mechanism. Analysis revealed that the primary benefit of this orchestration comes from reliably selecting the correct answer rather than from additional debate, and that the value of visual aids increases proportionally with the amount of problem information contained within the image. This system significantly improved accuracy and secured first place.

Why it matters

This research demonstrates a powerful approach for AI to tackle complex, multi-modal problems, particularly those requiring visual reasoning, which is crucial for advancing AI in scientific and technical fields.

How to implement this in your domain

  1. 1Explore integrating visual information extraction techniques into existing LLM pipelines for multi-modal problem-solving.
  2. 2Implement a multi-agent debate or orchestration framework to improve the reliability of AI-generated answers in complex tasks.
  3. 3Apply this two-stage approach to other domains where critical information is split between text and images, such as engineering diagrams or medical imaging.
  4. 4Investigate how to fine-tune visual extraction models for specific types of diagrams or figures relevant to your industry.

Who benefits

EducationEngineeringScientific ResearchHealthcareManufacturing

Key takeaways

  • Visual information is a major challenge for LLMs in physics problems.
  • A two-stage framework extracts visual data and uses multi-agent debate.
  • Orchestration improves answer selection reliability, not just debate.
  • The value of visual aids scales with the information they contain.

Original post by Jiseok Kwak, Suhyeon Jo, Taewoo Kim, Yeongmin Kim, Byeonghu Na, Il-chul Moon

"arXiv:2607.21946v1 Announce Type: new Abstract: This technical report presents our approach to Challenge Track~3: SeePhys Pro at the 3rd AI for Math Workshop, where the task is to answer college-level physics questions whose statement and figure may be given partly or entirely as…"

View on X

Originally posted by Jiseok Kwak, Suhyeon Jo, Taewoo Kim, Yeongmin Kim, Byeonghu Na, Il-chul Moon on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI ResearchAI Engineering & DevTools

StageGuard Improves Sleep Staging by Enforcing Physiological Constraints

StageGuard is a new framework that enhances automated sleep staging by integrating physiology-informed priors, ensuring that deep learning models produce hypnograms that adhere to known biological rules. It significantly reduces physiologically implausible transitions and fragmentation while maintaining or improving accuracy.

Juntang Wang, Yihan Wang, Hao Wu, Jiayu Gao, Shixin Xu, Dongmian ZouJul 28, 2026
AI ResearchAI Engineering & DevToolsAI News & Tools

AI Model Improves Trustworthy Flood Prediction with Explainability

Researchers developed Context-Aware Concept Distillation (CACD), a framework that distills opaque Deep Learning models into interpretable, hydrology-aware surrogates for flood prediction. This method provides verifiable causal narratives required by disaster response authorities, achieving high fidelity and outperforming black-box baselines globally.

Eli Levinkopf, Efrat Morin, Claudia V. GoldmanJul 28, 2026
AI ResearchAI Engineering & DevTools

Diffusion Models' Generative Quality Gets Comprehensive Theoretical Analysis

This research provides a unified theoretical framework for understanding the generalization and convergence of score-based diffusion models. It decomposes the total generative error into four interpretable components, quantifying how training data, discretization, and optimization affect sample fidelity.

Jinshu Huang, Yiming Jiang, Chunlin WuJul 28, 2026