Hidden-state reuse across prompts partially transfers answer shifts in small LMs
Jinkun Chen · hf · 2026-11-05
This Hugging Face paper examines cross-prompt reuse of entity-replacement responses in small language models. Reusing hidden states across factual questions partially transfers relation-specific answer shifts, with effectiveness depending on prompt format, cohort, and alignment with the source relation.
More from Research
- Foresight-SWE benchmark: coding agents solve 60.7% of issues but only 14.8% without introducing bugs — OfirPress · 2026-11-05
- Paper: "bottleneck pressure" shows work AI can't do gets more valuable as AI spreads — soumitrashukla9 · 2026-11-05
- Subtracting SFT weight deltas traces model behaviors to specific datasets, OpenAI alignment work shows — corbtt · 2026-11-05
- Visual Token Routing cuts VLM agent visual tokens by up to 5.8x at same accuracy — wellecks · 2026-11-05
- Flex-RSI: a new benchmark measuring recursive self-improvement of frontier agents — _vztu · 2026-11-05
- Four generations of protein design models chart the road to fully programmable protein design — AllThingsApx · 2026-11-05