MammoClaw: Towards Skill-Evolving Agent Harness for Breast Cancer Mammography Analysis

Krishna Kanth Nakka

eess.IVSep 25, 2026

MammoClaw: Towards Skill-Evolving Agent Harness for Breast Cancer Mammography Analysis

Authors: Krishna Kanth Nakka

Organizations: Munich, Bavaria, Germany

Abstract

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.

Figures & tables

Figure 1: Overview of the MammoClaw framework. The top panel illustrates the agent orchestration, while the bottom panel shows the skill-evolution workflow.

Table 1: Effect of tools and evolved skills on mammography assessment. Tools alone do not consistently improve macro-F1, whereas skill evolution leads to improved performance.

Figure 2: Skill evolution increases evidence gathering per case. After skill evolution, the agent makes more tool calls on the BI-RADS assessment task, suggesting a more deliberate multi-step inspection process before final prediction.

Figure 3: Comparison of tool-call frequency per case before and after skill evolution on the BI-RADS assessment task.

Figure 4: Skill-guided agent trajectory for BI-RADS prediction. The agent first localizes the suspicious region, verifies the finding using the paired MLO view and the contralateral breast, and incrementally refines its reasoning before producing the final assessment. The trajectory illustrates how tool interactions and evolved skills can support evidence-driven mammography assessment.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Task	Comparison	$\Delta$ Macro-F1	Boot. $p$ (Holm)	McNemar (acc.) $p$ (Holm)
BI-RADS	No tools $\rightarrow$ Tools	$-0.002$	$0.70$	$0.0017$
BI-RADS	No tools $\rightarrow$ Skills	$+0.040$	$<0.001$	$<0.001$
BI-RADS	Tools $\rightarrow$ Skills	$+0.043$	$<0.001$	$<0.001$
Density	No tools $\rightarrow$ Tools	$+0.019$	$0.70$	$0.062$
Density	No tools $\rightarrow$ Skills	$+0.093$	$0.18$	$<0.001$
Density	Tools $\rightarrow$ Skills	$+0.074$	$0.017$	$0.062$

Appendix

Table 2: Statistical significance of macro-F1 differences across tool and skill configurations.

Explore similar work

EvoClawBench: Can Agents Learn Reusable Skills from Their Own Runs?
Jun 23, 2026Zhiyuan Peng, Xin Yin, Chenhao Ying +5Computer-Use Agent Benchmarks LLM Agent Self-Improvement
1Shanghai Jiao Tong University · 2Zhejiang University · 3Hithink Research
RS-Claw: Progressive Active Tool Exploration via Hierarchical Skill Trees for Remote Sensing Agents
May 13, 2026Liangtian Liu, Zeyuan Wang, Ziyu Li +8Remote Sensing Image Understanding Tool-Using Agents
School of Geosciences and Info-Physics, Central South University, Changsha 410083, China · School of Resources and Environment, University of Electronic Science and Technology of China, Xian 710071, China · School of Earth Sciences and Spatial Information Engineering, Hunan University of Science and Technology, Xiangtan 411201, China +1
SkillMentor: LLM Agent Self-Evolution via Learning Blind-Spot Diagnosis
Jul 29, 2026Xiaoyi Bao, Yuanzhen Xie, Yunzhi Tan +5Agent Failure Analysis LLM Agent Self-Improvement
The Hong Kong Polytechnic University · Work done during an internship at Tencent. · Platform and Content Group, Tencent +1

Jun 23, 2026J/K move · Enter open · S save

EvoClawBench: Can Agents Learn Reusable Skills from Their Own Runs?

Zhiyuan Peng, Xin Yin, Chenhao Ying, Zhe Cui+4

1Shanghai Jiao Tong University · 2Zhejiang University · 3Hithink Research

Computer-Use Agent Benchmarks LLM Agent Self-Improvement LLM Agent Skill Learning AI Agent Benchmarks

Open paper AskPDF arXiv

Existing agent benchmarks primarily test task completion, tool use, or skill utility, but do not isolate whether a runtime can convert evidence from its own runs into reusable skills that improve fresh executions after authoring overhead. We introduce EvoClawBench, a benchmark for this closed-loop skill-learning question on repeated, fixture-backed tasks. EvoClawBench compares direct execution without skills, PreSkill authoring before execution, and PostSkill summarization from first-run evidence followed by a fresh second execution. The suite contains 100 tasks and 502 sub-problems across coding, data, office, security, operations, and domain-document workflows, with support for multiple agent runtimes. Experiments with OpenClaw and nanobot under local execution show that direct baseline performance is strongly runtime-dependent: OpenClaw remains below 20% across models, while nanobot ranges from 56.45% to 96.13%. Self-authored skills have mixed effects. nanobot GPT-5.4 stays above 96% in all modes and MiniMax-M2.7 improves from 90.97% to 94.50% under PostSkill, but nanobot DeepSeek-V4-Pro drops from 77.77% to 4.80% with PreSkill and 0.99% with PostSkill. OpenClaw shows similarly non-monotonic behavior, with some skill runs near baseline and others collapsing. These results indicate that learning reusable skills from an agent's own runs is selective and cost-sensitive, rather than an automatic benefit of adding skill authoring to an agent loop.

Abstract

Explore similar work

EvoClawBench: Can Agents Learn Reusable Skills from Their Own Runs?

RS-Claw: Progressive Active Tool Exploration via Hierarchical Skill Trees for Remote Sensing Agents

SkillMentor: LLM Agent Self-Evolution via Learning Blind-Spot Diagnosis