Can Computation from Earlier Problems Help LLMs Solve New Ones?
STAIR learns to read attention states saved from earlier answers, keeping the backbone and the saved states fixed.
On AIME 2025 with Qwen3.5-4B, STAIR improves mean later-turn accuracy (Avg@4) by 11.67 percentage points over the unmodified model with conversation history, using just 12,288 trainable parameters.