Can Computation from Earlier Problems Help LLMs Solve New Ones?
STAIR reuses earlier response keys and values, lifting later-turn accuracy by up to 11.67 points on Qwen.
The paper asks whether computation from earlier independent problems helps later ones, finding that retained history can raise or lower accuracy even within one domain. STAIR stores keys and values from earlier response generation in a fixed bank and learns to redirect current queries while the base model stays frozen, training only 12,288 parameters. Across three Qwen models and four benchmarks, average later-turn accuracy improves by up to 11.67 percentage points over the unmodified model with history.
- Retained history can raise or lower later-turn accuracy.
- STAIR stores earlier keys and values in a fixed bank.
- Only 12,288 parameters are trained; the base model stays frozen.
- Later-turn accuracy improves by up to 11.67 points.
Full article149 words · extracted from huggingface.co · click to collapse
Large language models often solve independent problems in the same conversation. Can computation from earlier problems help them solve new ones? To answer this question, we first conduct preliminary experiments showing that retained history can raise or lower later-turn accuracy, even within the same domain. To understand these effects, we use controlled replay to isolate internal state changes specific to each problem-history pairing. Across different histories, these changes preserve similar relationships among current problems. To improve reasoning under retained history, we introduce STAIR (Stale-Token Attention for Inter-query Reuse). STAIR captures keys and values from earlier response generation in a fixed bank. It learns to redirect current queries when they read this bank during prompt processing. The base model remains frozen; only 12,288 parameters are trained. Across three Qwen models and four benchmarks, STAIR improves average later-turn accuracy by up to 11.67 percentage points over the unmodified model with history.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.39394