AI

Unsatz: Evaluating Scientific Discovery Without a Given Ansatz

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 04/11/2026, 05:18:24 UTC.

Briefings

The paper introduces 'Unsatz', a benchmark where LLM agents must construct the mathematical ansatz (assumption) for simulated systems before testing. Unlike traditional benchmarks that provide equations, agents discover the structure themselves. Across 80 systems in 20 domains, nine language models solved between 21% and 60% of tasks. Classical equation-discovery methods rarely succeed here, highlighting that finding the underlying structure is the primary challenge. Agents also benefit from sharing discovered programs with related systems, and stronger agents reduce scientific errors.

Key points

• Unsatz is a benchmark where LLM agents must construct the mathematical ansatz (assumption) for simulated systems before testing, rather than being given equations.
• Across 80 systems in 20 domains, nine language models solved between 21% and 60% of tasks, with classical equation-discovery methods rarely succeeding without prior knowledge of the true equations.

Context

This research addresses the challenge of automated equation discovery by introducing a new benchmark that requires agents to construct the ansatz themselves. It matters because it shifts focus from estimating coefficients to discovering the underlying structure of physical systems, which is where most difficulty lies.

Limitations

The study covers 80 systems across 20 domains using nine language models. Results are limited to the specific tasks where agents successfully constructed the ansatz, and the benchmark does not account for all possible scientific errors or unknown measurement channels.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 04/11/2026, 07:32:15 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeid5xgevzkzplwebz6kklx6irmnoxz6zgbzoslsgaud73xhqs7wb6m