Language model audits and repairs flaws in small clinical decision networks
Original title: Weight-space programming for auditing clinical decision networks
Claude Opus 5.5 audited and repaired flaws in radiotherapy networks with 433-4801 parameters
The study tests frontier language models on weight-space programming for auditing neural networks in radiotherapy planning. Networks trained on a clinical protocol had planted flaws like confounds and threshold shifts. Claude Opus 5.5 identified broken rules in 16 of 18 cases and repaired them by altering a median of 3 parameters.
- Networks had 433 to 4,801 parameters
- Claude Opus 5.5 succeeded in 16 of 18 flawed networks
- Fine-tuning on protocol-labeled requests also repaired all networks
- Each flaw was visible in single units
- Tags
- weight-space-programmingclinical-decision-networksradiotherapyml-auditlanguage-modelneural-networks
- AI models
- Claude Opus 5.5
From the original article
1 min read at arxiv.orgMachine learning models in radiotherapy are commissioned by testing them on cases with known answers, which may not reveal a shortcut that the test cases share with the training data. Frontier language models have recently been reported to read and write the raw weights of small neural networks, an ability called weight-space programming.
By Humza Nusrat · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.