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

A new framework called SciOrch trains a lightweight 8-billion-parameter model to coordinate multiple frontier large language models for scientific reasoning tasks, achieving higher accuracy than any single commercial system while cutting API costs by more than half, according to research posted to arXiv on June 14 [1][2]. Frontier scientific reasoning remains a major challenge for large language models, where even the strongest commercial systems fall short of expert-level performance [2]. The researchers behind SciOrch found that different frontier models excel on different question types, and no single model captures the full picture [2]. Their orchestrator decomposes each question, delegates sub-problems to selected commercial models through API calls, and synthesizes a final answer [2]. Training such an orchestrator is fundamentally harder than conventional agentic reinforcement learning because each action triggers an API call that is expensive in both dollar cost and latency, making standard online rollouts infeasible [2]. The team addressed this with a Monte Carlo tree search-based approach, producing diverse orchestration trajectories, extracting per-node single-turn samples, and optimizing the orchestrator with GRPO-style training [2]. On a 240-question test set spanning SGI-Reasoning and Scientists' First Exam, SciOrch reached 56.66% average accuracy, outperforming the strongest single commercial model by 3.74% and the strongest multi-agent baseline by 3.33% [2]. It also attained the best accuracy on both benchmarks with less than half the API cost of typical multi-agent methods [2]. The paper appears on arXiv, a preprint repository that has integrated with Hugging Face Spaces since 2022 to make machine learning research more accessible through interactive demos [3][4]. Through that integration, users can find open-source demos linked directly on a paper's arXiv abstract page, allowing anyone to try models without writing code [5]. Hugging Face Spaces has been used to build and share over 12,000 open-source machine learning demos crafted by the community [3]. The broader landscape of large language models has seen rapid evolution. Chinese firm DeepSeek, founded in July 2023 by High-Flyer co-founder Liang Wenfeng, launched its DeepSeek-R1 model in January 2025 with performance comparable to OpenAI's GPT-4 and o1, while claiming training costs of $6 million for its V3 model — far less than the $100 million reported for GPT-4 in 2023 [6]. DeepSeek's models are described as open-weight, meaning the exact parameters are openly shared, though training data is not openly licensed [6]. LLMs are language models with many parameters, trained with self-supervised learning on vast amounts of text [7]. The field has attracted researchers such as Douwe Kiela, a research scientist director at Google DeepMind who previously led the team at Meta AI that introduced the retrieval-augmented generation approach in 2020 and later served as Head of Research at Hugging Face [8].