SciOrch: Learning to Orchestrate Expert LLMs for Solving Frontier Multimodal Scientific Reasoning Tasks

2026-06-14

AI Summary

This work addresses the challenge posed by cutting-edge multimodal scientific reasoning tasks, which remain difficult for current large language models to solve at expert levels. The authors propose a lightweight 8B-parameter coordinator that decomposes complex problems and efficiently orchestrates multiple commercial large models through an offline training strategy guided by Monte Carlo Tree Search (MCTS). By integrating single-turn example extraction with GRPO-style reinforcement learning, the method fuses outputs from these models to generate final answers. This approach achieves an average accuracy of 56.66% on the SGI-Reasoning and Scientists' First Exam benchmarks—surpassing the strongest single-model baseline by 3.74% and the best multi-agent baseline by 3.33%—while reducing API call costs by over 50%.

Abstract

Frontier scientific reasoning remains a major challenge for large language models (LLMs), where even the strongest commercial systems fall short of expert-level performance. A closer look at model behavior reveals substantial complementarity that single-model evaluation hides: different frontier models excel on different question types, and no single model captures the full picture. We present SciOrch, a framework that trains a lightweight 8B model to orchestrate frontier LLMs for scientific reasoning. The orchestrator decomposes each question, delegates sub-problems to selected commercial models through API calls, and synthesizes a final answer. Training such an orchestrator is fundamentally harder than conventional agentic RL: each action triggers an API call that is expensive in both dollar cost and latency, making standard online rollouts infeasible. We address this with MCTS-based approach, producing diverse orchestration trajectories, extracting per-node single-turn samples, and optimizing the orchestrator with GRPO-style training. On a 240-question test set spanning SGI-Reasoning and Scientists' First Exam, SciOrch reaches 56.66% average accuracy, outperforming the strongest single commercial model by 3.74% and the strongest multi-agent baseline by 3.33%. It also attains the best accuracy on both SGI and SFE with less than half the API cost of typical multi-agent methods.

Problem

• scientific reasoning
• large language models
• multimodal
• expert-level performance
• complementarity

Innovation

• orchestration
• multimodal scientific reasoning
• MCTS-based training
• API-efficient coordination
• GRPO-style optimization