Johns Hopkins study shows LLMs fail to ignore training-time knowledge when context conflicts
mdredze · x · 2026-09-04
A Johns Hopkins (CS/CLSP) study presented at ACL 2026 shows LLMs keep relying on information learned during training even when explicitly instructed to use conflicting new information — a "context-obedience" failure.
- First author Kaiser Sun notes models are increasingly expected to combine training knowledge with at-use information (retrieved docs, report summarization, evaluating AI-generated content), so conflicts are common.
- Implications for AI search and RAG-style QA: parametric memory can silently override the provided context.
More from Research
- De novo designed antibodies protect against lethal cobra venom in vivo — BrianHie · 2026-09-04
- Roman Telescope: 100x Hubble's field of view, to catalog 1 billion galaxies in year one — PeterDiamandis · 2026-09-04
- New paper: two tales of the geometric Jensen-Shannon divergence as JSD regularizations — FrnkNlsn · 2026-09-04
- Eric Topol's Lancet essay reviews AI health models that predict disease years ahead — EricTopol · 2026-09-04
- OpenCLIP's new native ModernText encoder: more customizable than Transformers ModernBERT, decoder-capable — wightmanr · 2026-09-04
- Eureka meta-architecture grows task-specific AI brains, pushes Riemann-boundary to a≤69/200 — 新智元 · 2026-09-04