Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
MammoClaw establishes a training-free agent harness that equips frozen multimodal LLMs with specialized mammography tools—such as paired-view and contralateral-breast inspection—for BI-RADS assessment and breast density estimation. To prevent brittle decision-making, the framework implements an episodic skill-evolution mechanism that converts failed reasoning trajectories into non-parametric, reusable prompt guidance for future runs. Crucially, the authors discover that augmenting MLLMs with specialized vision tools alone yields negligible gains, whereas compounding tool invocation with evolved procedural skills reliably rectifies tool-misuse failure modes.
Specialized medical vision tools fail to boost multimodal LLM diagnostic accuracy on their own—performance gains only emerge when the agent autonomously converts prior trajectory failures into reusable procedural skills.
In this work, we explore MammoClaw, a training-free agent framework that leverages frozen MLLMs for mammography analysis. To support agentic investigation, we equip the agent with lightweight mammography-specific tools for targeted image analysis, including ROI, paired-view, and contralateral-breast examination. MammoClaw iteratively gathers evidence through these tools, while skill evolution enables non-parametric adaptation by transforming failed trajectories into reusable guidance for later runs. We evaluate the framework on BI-RADS assessment and breast density estimation tasks. In our experiments, we find that tools alone do not reliably improve performance, whereas evolved skills can improve tool-use behavior and performance in some settings. Beyond these results, MammoClaw enables transparent inspection of evidence acquisition, tool interactions, and failure modes, facilitating the analysis and auditing of agent behavior. We view this work as an exploratory study of training-free, self-evolving agentic approaches for mammography and hope it provides a concrete starting point for future work on mammography-specific tools and self-evolution mechanisms. We release our code at https://krishnakanthnakka.github.io/mammoclaw.