Can Computation from Earlier Problems Help LLMs Solve New Ones?
STAIR reuses earlier generation state with only 12,288 trainable parameters.
The paper tests whether retained conversation history can help or hurt later problem solving, finding both effects even within one domain. It uses controlled replay to isolate the internal state changes tied to each problem-history pairing. The authors introduce STAIR, which stores keys and values from earlier responses and learns when current queries should attend to that bank. Across three Qwen models and four benchmarks, it reports gains of up to 11.67 percentage points over the same model using history without STAIR. HF Daily Papers' note
The paper tests whether retained conversation history can help or hurt later problem solving, finding both effects even within one domain. It uses controlled replay to isolate the internal state changes tied to each problem-history pairing. The authors introduce STAIR, which stores keys and values from earlier responses and learns when current queries should attend to that bank. Across three Qwen models and four benchmarks, it reports gains of up to 11.67 percentage points over the same model using history without STAIR. HF Daily Papers' note
score 4