Can Computation from Earlier Problems Help LLMs Solve New Ones?
Summary
This paper introduces STAIR (Stale-Token Attention for Inter-query Reuse), a method that stores keys and values from earlier response generations in a fixed bank and trains only 12,288 parameters to direct current queries, improving later-turn accuracy by up to 11.67 percentage points on Qwen models across four benchmarks.
View Cached Full Text
Cached at: 10/05/26, 12:47 PM
Paper page - Can Computation from Earlier Problems Help LLMs Solve New Ones?
Source: https://huggingface.co/papers/2609.39394
Abstract
Largelanguagemodelsoftensolveindependentproblemsinthesameconversation.Cancomputationfromearlierproblemshelpthemsolvenewones?Toanswerthisquestion,wefirstconductpreliminaryexperimentsshowingthatretainedhistorycanraiseorlowerlater-turnaccuracy,evenwithinthesamedomain.Tounderstandtheseeffects,weusecontrolledreplaytoisolateinternalstatechangesspecifictoeachproblem-historypairing.Acrossdifferenthistories,thesechangespreservesimilarrelationshipsamongcurrentproblems.Toimprovereasoningunderretainedhistory,weintroduceSTAIR(Stale-TokenAttentionforInter-queryReuse).STAIRcaptureskeysandvaluesfromearlierresponsegenerationinafixedbank.Itlearnstoredirectcurrentquerieswhentheyreadthisbankduringpromptprocessing.Thebasemodelremainsfrozen;only12,288parametersaretrained.AcrossthreeQwenmodelsandfourbenchmarks,STAIRimprovesaveragelater-turnaccuracybyupto11.67percentagepointsovertheunmodifiedmodelwithhistory.
View arXiv pageView PDFAdd to collection
Get this paper in your agent:
hf papers read 2609\.39394
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2609.39394 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2609.39394 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2609.39394 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
Recent Developments in LLM Architectures: KV Sharing, mHC, and Compressed Attention [P]
Sebastian Raschka reviews recent innovations in LLM architectures focused on long-context efficiency, including KV sharing, compressed convolutional attention, and layer-wise attention budgeting from models like Gemma 4, ZAYA1, Laguna XS.2, and DeepSeek V4.
Nous Research Releases Token Superposition Training to Speed Up LLM Pre-Training by Up to 2.5x Across 270M to 10B Parameter Models
Nous Research releases Token Superposition Training (TST), a method that speeds up LLM pre-training by up to 2.5x across models from 270M to 10B parameters, reducing wall-clock time without altering architecture or data.
Researchers Discover a Way to Reduce LLM Training Compute by 15% Without Sacrificing Quality
At EMNLP 2026, T-Tech researchers presented LIMe, a lightweight attention extension that reuses key-value buffers from earlier layers to mitigate representation collapse, reducing LLM training compute by ~15% at equal quality versus baseline LLaMA on ~1B-parameter models.
Continual LLM Upcycling: A Predictor-Gated Bank-Wise Sparsity Training Recipe for Dense-to-Sparse LLMs
This paper proposes a dense-to-sparse continual training method for LLMs, using a predictor-gated bank-wise sparsity to achieve 4x FFN sparsity, and demonstrates it on Qwen2.5-8B with long-context training.
Asking For An Old Friend: Diagnosing and Mitigating Temporal Failure Modes in LLM-based Statutory Question Answering
This paper studies temporal failure modes in LLM-based statutory question answering, including post-cutoff staleness and recency bias. It introduces a benchmark of 312 expert-validated German statutory QA pairs and evaluates LLMs under various inference settings.