HExA Agents Learn from Active Experimentation, Outperforming LLMs

Abhranil Chandra, Sankaran Vaidyanathan, Utsav Dhanuka, Varun Gandhi, Scott Niekum· June 30, 2026 View original

Key takeaways

  • HExA enables LLMs to learn from active experimentation, overcoming limitations of static knowledge.
  • It iteratively designs experiments, learns reusable skills, and integrates evidence.
  • HExA significantly boosts LLM performance on novel, complex tasks like physics puzzles.
  • The framework is training-free, model-agnostic, and requires no external supervision.

Who benefits

RoboticsAutonomous SystemsScientific ResearchGamingAI Engineering

Summary

Researchers introduce Hierarchical Experimentalist Agents (HExA), a framework enabling LLMs to learn from active experimentation and acquire reusable skills without external supervision. HExA significantly improves LLM performance on complex, novel physics tasks, demonstrating its ability to discover knowledge and generalize skills.

Traditional large language model (LLM) agents often rely on pre-trained knowledge, retrieval, or search, which limits their effectiveness in novel domains or for complex queries requiring new understanding. To address this, a new framework called Hierarchical Experimentalist Agents (HExA) has been developed, allowing LLMs to learn through active experimentation. HExA iteratively designs and refines experiments, builds a library of composable skills from its experiences, and integrates experimental evidence to answer queries or perform long-horizon tasks. This training-free framework is compatible with any black-box model and requires no external supervision. Evaluated on Interphyre, a new physics-based benchmark, HExA dramatically improved a Claude Sonnet model's success rate from 2% to 77%, also outperforming other agentic baselines and demonstrating skill reusability.

Why it matters

This breakthrough enables LLMs to go beyond parametric knowledge, actively learn from interaction, and adapt to entirely new problems, opening doors for more capable and autonomous AI systems.

How to implement this in your domain

  1. 1Explore HExA's principles for developing AI agents that need to operate in dynamic or novel environments.
  2. 2Design internal simulations or sandboxes where LLM agents can actively experiment and learn new skills.
  3. 3Investigate integrating active experimentation modules into existing LLM-powered decision-making systems.
  4. 4Develop strategies for curating and reusing learned skills from experimental agents across different tasks.

Original post by Abhranil Chandra, Sankaran Vaidyanathan, Utsav Dhanuka, Varun Gandhi, Scott Niekum

"arXiv:2606.29315v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly used to take actions in the real world and support human decision-making, yet most agents rely on parametric knowledge, fixed post-training data, retrieval, or search. This paradigm brea…"

View on X

Originally posted by Abhranil Chandra, Sankaran Vaidyanathan, Utsav Dhanuka, Varun Gandhi, Scott Niekum 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 Engineering & DevToolsAI ResearchAI News & Tools

GLM-5.3 Model Demonstrates Advanced Coding and Cyber Capabilities

The GLM-5.3 model has been unveiled, showcasing advanced capabilities in frontier coding and emergent cyber operations. This development points to significant progress in AI's ability to handle complex programming tasks and potentially cybersecurity challenges.

pellaAug 14, 2026
AI Engineering & DevToolsAI ResearchAI Investing

FlowLOB Generates Realistic, Controllable Limit Order Books Efficiently

This paper introduces FlowLOB, a conditional flow-matching generator for Limit Order Book (LOB) trajectories that offers realistic market dynamics, efficient sampling, and controllable scenario generation, outperforming existing agent-based and deep generative simulators. FlowLOB achieves high fidelity with significantly fewer computational steps than diffusion models and transfers effectively to unseen instruments.

Zhuohan Wang, Andreea Bacalum, Ollie Olby, Carmine Ventre, Namid StillmanAug 14, 2026
AI Engineering & DevToolsAI Research

Auditing Reveals Bias in Neural Combinatorial Optimization Benchmarks

This paper audits test-time budget allocation in Neural Combinatorial Optimization (NCO) solvers, revealing that reported gains from non-uniform sampling often stem from "sampling luck" rather than true allocation benefits on in-distribution data. It proposes a correction procedure and demonstrates real gains under distribution shift, emphasizing the need for rigorous evaluation.

Jinhyung BaeAug 14, 2026