acceptodds
Desk rejected at ICLR 2027

Tracing Biomedical RAG Update Windows with Joint Evidence Admission

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

A newly indexed biomedical passage can meet the same query at 0, 6, 12, 24, or 48 hours, yet snapshot evaluation assigns these encounters one security state. We introduce a paired-delay replay that advances query time while fixing bytes, query, payload, wave, candidate content, and credibility inputs, together with Temper (Temporal Evidence Management with Provenance Enforcement for RAG), which jointly admits committed use, age-indexed exposure, and claim support. Under its \(4\!\rightarrow\!3\!\rightarrow\!2\!\rightarrow\!0\!\rightarrow\!0\) pending-slot cap, success on the same 7,500 injection trials traces 4.4%, 3.8%, 3.0%, 2.1%, and 2.3%, a 2.2-point [1.3, 3.1] contrast absent from no-quarantine and SecAlign+RAGuard controls. Against the evaluated RAGuard–SecAlign composition, Temper reduces poisoning from 6.1%0.9% to 3.2%0.7% and accepted evidence modifications from 688/1,000 to 2/1,000. Under adaptive attacks, independent labels of emitted citations give pooled F1 of 80.7% versus 68.6%, while high-severity clinical-harm events fall from 8.8%1.4% to 5.9%1.1%. Temper retains 5.2%/4.1% injection/poisoning success with Qwen2.5-72B-Instruct and benchmark accuracy within 0.9 points of Vanilla; warm-cache p50 overhead is 11.4%. Evidence age is therefore a measurable security coordinate from newly committed bytes to cited biomedical claims.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
trustworthy evidence security and provenance, Tracing Biomedical RAG Update Windows with Joint Evidence Admission

Related papers

Discussion (0)

Sign in to comment.