BioAgent-DF audit finds citation shifts without verdict changes
Original title: From Tool State to Scientific Action: Auditing Decision Provenance in Biomedical Research Agents
On eight SciFact batches, Qwen3.5-4B and Gemma 3 4B IT showed repeatable evidence changes across seven claim combinations covering two claims.
The method saves paired reference and candidate values at the same computation site, then restores or transplants them before rerunning the agent on healthy hardware. Full restoration recovered the reference evidence sets while verdicts stayed the same. In an hERG assay-allocation task, silent numerical faults changed 39 testing decisions; restoring prior execution state and certified scores recovered all of them, while replacing the prediction file recovered none. A streaming arithmetic check cut forward execution and verification time by 42.4% versus FP64 with identical flags.
- Qwen3.5-4B had repeatable citation changes in three capture-claim combinations.
- Gemma 3 4B IT had repeatable read and citation changes in four combinations.
- Baseline-prompt verdict changes did not recur under the explicit-example prompt.
- The fixed-budget shortlist kept the same count of compounds labeled blockers.
- No incorrect action was committed after faulty values were used.
From the original article
1 min read at huggingface.co
A scientific agent can give the same verdict while reading and citing different evidence. BioAgent-DF tests this separation using paired reference and candidate values saved at the same scientific computation site. It restores or transplants those values at their original coordinates, then runs the agent on healthy hardware.
By Xingyu Yi, Xinya Zhang, Yunshen Zhou · huggingface.co
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.