Research map / Multi-agent systems and games / LLM agents

Simple Agents Outperform Experts in Biomedical Imaging Workflow Optimization

CVPR 2026 · Multi-agent systems and games › LLM agents · approach: Software engineering agent benchmarks · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Systematically studies design choices of minimal LLM code-writing agents that adapt production biomedical imaging tools (Polaris, Cellpose, MedSAM) via preprocessing/postprocessing code, showing simple agents beat expert-written baselines.

Problem

Adapting pretrained production computer vision tools to bespoke scientific datasets with small labeled validation sets requires weeks to months of manual coding; it is unclear which agent design components are needed for this targeted tool-adaptation task.

Approach

A minimal code-edit-and-execute agent (Base Agent with data prompt and API list) generates pre/postprocessing function pairs scored on a small validation set; an ablation over LLM type, expert functions, function bank, AutoML, and tree search (AIDE) shows component effects are task-dependent and complexity is not universally beneficial.

Key results

• Base Agent outperforms expert baseline on all three tasks, e.g., MedSAM 0.971 vs 0.820 and Polaris 0.867 vs 0.841 (Cellpose 0.409 vs 0.402).
• Expert functions raise Polaris to 0.929 but lower MedSAM to 0.888; Reasoning LLM (o3) gives MedSAM 1.020 but Polaris 0.844.
• Ablating the API list improves all three tasks (Polaris 0.868, Cellpose 0.417, MedSAM 1.037); minimal Base Agent matches AIDE tree search (Polaris 0.872, Cellpose 0.414, MedSAM 0.971).

Surprising findings

• Providing the API list consistently hurt performance and biased API usage toward specific functions.
• AIDE tree-search agent showed no advantage over the minimal Base Agent, and Function Bank hurt MedSAM despite helping other tasks.

Limitations

• Focused exclusively on biomedical imaging tasks (three case studies).
• Agents overfit small validation sets; AutoML worsened Polaris test performance.
• Only simple minimal agent designs studied; more complex designs (RAG, multi-agent) not explored.

Experimental setup

Datasets: Polaris validation set (95 images; test 94) Cellpose3 test sets, TissueNet 1.1, Omnipose bacterial sets (807 test images) Codabench MedSAM dermoscopy validation set (25 val / 25 test)

Benchmarks: Polaris spot detection Cellpose segmentation (807 test images) MedSAM dermoscopy segmentation

Metrics: F1 score average precision at IoU 0.5 NSD+DSC sum

Baselines: Expert Baseline (official tool-author functions) Non-Agentic AutoML AIDE agent Small LLM (Llama 3.3-70B) Reasoning LLM (o3)

Builds on: AIDE Optuna Cellpose MedSAM Polaris

Closest papers by meaning

• From Reproduction to Replication: Evaluating Research Agents with Progressive Code Masking (ICLR 2026 · similarity 0.86)
• AstaBench: Rigorous Benchmarking of AI Agents with a Scientific Research Suite (ICLR 2026 · similarity 0.86)
• AutoScientists: Self-Organizing Agent Teams for Long-Running Scientific Experimentation (NeurIPS 2026 · similarity 0.85)
• An Intelligent Agentic System for Complex Image Restoration Problems (ICLR 2025 · similarity 0.85)
• BioAgent Bench: An AI Agent Evaluation Suite for Bioinformatics (ICML 2026 · similarity 0.85)
• AgenticVBench: Can AI Agents Complete Real-World Video Production Tasks? (NeurIPS 2026 · similarity 0.85)