Autonomous Agents for Scientific Tasks - Sina Shahandeh, Radicait

2026-07-18 · Source: AI Engineer · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Health & Medical Research · Depth: Advanced, long

Summary

Sina Shahandeh of Radicait presented a method to enhance autonomous agents for scientific tasks, addressing their tendency to saturate due to a lack of novel hypothesis generation. Traditional coding agents optimize existing code well but struggle with open-ended scientific problems requiring "research taste." The proposed solution explicitly decomposes complex scientific problems, such as in-silico PET generation from CT scans. This creates a hierarchical structure of subcomponents. This documented hierarchy enables LLM-based agents to systematically explore and propose radical improvements. An example is transitioning from 2.5D to 3D convolutions in image translation models. The approach integrates multimodal models, such as Gemini, for qualitative image review in tasks like image registration. It also uses reasoning models, like GPT-4.5 Pro via Oracle CLI, for hypothesis generation. These models also critique within the scientific discovery loop. This structured decomposition acts as a "trick" to improve LLM reasoning when models lack scientific image training.

Key takeaway

For AI Scientists or Machine Learning Engineers tackling complex, open-ended scientific research, you should implement explicit problem decomposition. This hierarchical approach, documented and traversable by LLM agents, will significantly improve hypothesis generation beyond simple hill-climbing optimization. Integrate multimodal models for qualitative data assessment and reasoning models for robust hypothesis critique. This strategy helps overcome agent saturation, accelerating scientific discovery and enabling more radical model improvements.

Key insights

Decomposing scientific problems hierarchically enables LLM agents to generate more effective hypotheses and overcome saturation.

Principles

• Scientific tasks require continuous hypothesis generation.
• LLMs benefit from structured problem decomposition.
• Integrate multimodal and reasoning models for critique.

Method

Explicitly decompose a scientific problem into a hierarchy of subcomponents, documenting each. Use an LLM to traverse this hierarchy, generating and refining hypotheses for each component.

In practice

• Create a linked document hierarchy for your codebase.
• Use multimodal models for qualitative data review.
• Employ reasoning models for hypothesis generation and critique.

Topics

• Autonomous Agents
• Hypothesis Generation
• Problem Decomposition
• Multimodal AI
• In-silico PET Generation
• Scientific Research Automation

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Articulating Assumptions in AI-Generated Scientific Analyses through Task Decomposition — Multi-agent task decomposition improves LLM-generated scientific code transparency and enables smaller models.
• Toward Generalist Autonomous Research via Hypothesis-Tree Refinement — Autonomous research can be a cumulative process driven by structured hypothesis refinement and persistent knowledge integration.
• From Data to Theory: Autonomous Large Language Model Agents for Materials Science — Autonomous LLM agents can perform end-to-end scientific theory development, from equation selection to code execution and validation.
• AI systems devise hypotheses and ways to test them — Multi-agent AI systems can extend human scientific discovery by generating and testing hypotheses.
• Battery-Sim-Agent: Leveraging LLM-Agent for Inverse Battery Parameter Estimation — LLM agents can reframe inverse scientific problems as reasoning tasks, outperforming traditional black-box optimization.
• Towards Generalist Agents for Accelerating Scientific Discovery171 — LLMs can accelerate scientific discovery by generating and refining hypotheses within agentic, iterative workflows.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Engineer.