# Can LLMs Uncover Hidden Physical Laws through Free Exploration?

Source: [arXiv](https://arxiv.org/abs/2610.37649v1)  
Feed7 permalink: https://feed7.dev/p/2610-37649v1-1q38qod  
Published: 2026-10-27T17:57:00.000Z  
Trust: Needs Review (needs_review)

## Why Included

LongChang tests whether models can discover altered physics through unguided exploration; 1 of 12 models found any law, and its findings were partial.

## Source Summary

LongChang creates simulated worlds with altered physical laws and lets models explore without hints or assigned research tasks. **GPT-6-Astra found a law in 16.7% of cases**; the **other 11 models scored 0%**.

## Practical Implication

For builders evaluating autonomous agents, this setup probes whether agents can notice anomalies, form questions, and investigate without a predefined task. The framework also uses separate agents to review scenes, tools, and reports.

## Agent-Ready Context

LongChang creates simulated worlds with altered physical laws and lets models explore without hints or assigned research tasks. **GPT-6-Astra found a law in 16.7% of cases**; the **other 11 models scored 0%**.

For builders evaluating autonomous agents, this setup probes whether agents can notice anomalies, form questions, and investigate without a predefined task. The framework also uses separate agents to review scenes, tools, and reports.

The exploration records showed uninvestigated anomalies and only partial identification of altered laws, so discovery remained limited even for the top-scoring model.

## Connected Context

Feed7 judgment across 944 accumulated Signals:

This benchmark tests whether agents can discover physical laws without assigned research tasks or hints, extending agent evaluation to open-ended exploration. The low discovery rates and uninvestigated anomalies show that success at autonomous research remains limited even when exploration is possible.

- [First Steps Toward Automated AI Research — Richard Socher, CEO Recursive AI](https://feed7.dev/p/first-steps-toward-automated-ai-research-richard-socher-ceo-recursive-ai-17qblkw) — The benchmark isolates unguided discovery, while Socher’s design adds orchestration and experimental validation to automated research; together they distinguish exploration from a fuller research loop.
- [The Future of Evals: From LLM as a Judge to Agent as a Judge — Aparna Dhinakaran, Arize AI](https://feed7.dev/p/the-future-of-evals-from-llm-as-a-judge-to-agent-as-a-judge-aparna-dhina-1fu560o) — The benchmark’s exploratory trajectories can contain missed anomalies, which supports evaluating agent behavior across long trajectories rather than relying only on a fixed rubric or final answer.

## Context Map

- Layer: benchmark
- Domains: research
- Topics: agent-evals, agent-reliability, multi-agent

## Uncertainty

- The exploration records showed uninvestigated anomalies and only partial identification of altered laws, so discovery remained limited even for the top-scoring model.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.