從工具狀態到科學行動:稽核生物醫學研究智能體的決策溯源
From Tool State to Scientific Action: Auditing Decision Provenance in Biomedical Research Agents
研究在健康硬件上運行 Biomni A1,並重播故障 GPU 工具輸出進行 hERG 心臟安全篩查,結果有兩個基準標記為 hERG 阻斷劑的化合物失去原有測試安排。
A faulty GPU changed our agent’s safety-testing decisions while the shortlist’s blocker count stayed the same.
👋 Author here. We study silent GPU numerical errors: computation can return incorrect values without a crash or NaN.
We replayed captured faulty tool outputs into a hERG cardiac-safety screening task with the native Biomni A1 agent, while the agent itself ran on healthy hardware:
- Two benchmark-labeled hERG blockers lost the testing assignments they received under the healthy reference.
- The shortlist still contained 351 labeled blockers out of 402 compounds. The count stayed the same; the members changed.
- Finite-value and data-integrity checks flagged 0/16 fault sources. Linear-layer checksums flagged 16/16.
- After the agent had read the faulty values, replacing the prediction file recovered 0/39 changed decisions. Restoring its pre-read execution state and supplying trusted values recovered 39/39.
SciFact shows a related separation: agents can return the same verdict while citing different evidence.
Takeaway: correcting a tool result may not undo decisions already formed from it. Recovery also needs to address the agent’s dependent execution state.
We’re looking for independent GPU testers. We plan to publicly release Aritheia, our GPU numerical reliability testing tool. If you’d like to test your NVIDIA GPU or cluster when it is available, comment with your GPU model and driver/CUDA/PyTorch versions. We welcome both runs with no detected anomalies and reproducible numerical mismatches.
Project page: https://intosyn.github.io/BioAgent-DF/
Happy to discuss the results and help with testing.
來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co