Berkeley paper: LLMs answer with stale values even when the update is in context
ombre-ai · reddit · 2026-11-16
UC Berkeley's Sept 30 paper "When Context Changes: Understanding Update Failures in LLMs" (arXiv 2609.38866) documents stale binding: after a variable is updated in context, models sometimes still answer with the old value (e.g., a meal preference changed from gluten-free to Mediterranean, yet the answer says gluten-free).
Key findings:
- Probes can still recover the current value when the old one was chosen — a selection failure, not information loss.
- Component tests on Qwen and Pythia point to attention drift: attention favors old values when producing the answer.
- Mechanistic work is only on open models; frontier models were tested behaviorally and fail once update structures get complex.
The poster asks whether others have seen this in long local runs (agent logs, long roleplay, mid-session config changes), sketches a rough reproduction (several overwritten variables with filler between, then ask for the current value), and asks what mitigations people use: restating state, keeping a state block, or pruning old values.
More from coding & agent
- Generic API-catalog MCP server vs one tool per job: a context-bloat experiment — MandeepKaur9407 · 2026-11-16
- skills-gate: one gate skill fronts 148 agent skills at just 389 tokens — Distinct-Pie2389 · 2026-11-16
- Open-weight Karpo-Decision-31B tops Decision Index at 73.9, one forward pass per answer — CodeByPoonam · 2026-11-16
- shadcn boosts @uisdk: one-shot UI rebuild that's instantly E2E-testable by AI — shadcn · 2026-11-16
- AI agent deletes 43,599 lines of OpenClaw test code in one PR — vincent_koc · 2026-11-16
- JPMorgan shows Jarvis, its internal agent assembly line, at Interrupt NYC — LangChain · 2026-11-16