Aug 2026· Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2· 0 citations· 32 references
TL;DR
The proposed NumCache, which compresses SEC filings into KV caches initialized from numerically dense regions and trained directly on financial QAs, is evaluated, which highlights cache-based retrieval with number-preserving representations as an effective approach for long-context financial QA.
Abstract
Large Language Models (LLMs) are increasingly deployed in financial applications, particularly for interpreting U.S. Securities and Exchange Commission (SEC) filings. However, financial QA over these filings is challenging, as they are extremely long, numerically dense, and often require cross-document reasoning. Existing approaches struggle to scale to such settings due to long-context degradation and loss of numerical fidelity under context compression. Long-document Retrieval-Augmented Generation (RAG) improves evidence coverage through coarse-to-fine retrieval, yet semantic retrievers collapse fine-grained magnitudes and units, returning passages that lack the precise values needed for correct reasoning. Cache-Augmented Generation (CAG) projects attention states into compact KV representations and treats precomputed caches as reusable internal memory, but typically assumes caches are already well-formed and query-relevant, leaving open how to build and select number-faithful caches. To address this gap, we propose NumCache, which compresses SEC filings into KV caches initialized from numerically dense regions and trained directly on financial QAs. On top of these caches, we then train a contrastive retriever that aligns questions with cache representations, thus improving retrieval performance. We evaluate NumCache on the Fin-RATE benchmark, which comprises SEC-filing QA tasks covering single-filing reasoning, cross-firm comparison, and longitudinal trend analysis. NumCache achieves up to 4× context compression with competitive accuracy (41.4% vs.\ 44.8% for uncompressed Qwen3-4B on single-filing reasoning) and a 6.4× inference speedup, while its contrastive retriever attains 68.3% Recall@1 on single-filing retrieval; 2.3× the strongest text-based baseline (29.7%). These results highlight cache-based retrieval with number-preserving representations as an effective approach for long-context financial QA.
This study investigates the crucial deployment question of whether Cartridges and other CKVs can handle off-context queries, and proposes Cartridges++, simple modifications to cartridges that retain off-context abilities at small or negligible cost.
Sonia Laguna, João Monteiro, Marco Cuturi et al.· arXiv.org· 1 citation
This work revisits RAG compression from a data-mining perspective by aggregating historical query--document--model interactions into reusable evidence views and proposes Reusable Evidence View Aggregation (REVA), a framework that mines the target generator's historical attention traces into a document-keyed, budget-agn...
T. Nguyen, Qi-Ran Hu, Ban-Ruo Liu et al.· 0 citations
Long-context inference with Large Language Models (LLMs) is bottlenecked by the linearly growing memory of the key-value (KV) cache. Existing compression methods reduce the cache through token eviction or approximation, but degrade sharply at aggressive compression budgets. We propose PatchKV, a training-free framework...
A controlled comparison of representation-based (KV-cache based) and parametric (fine-tuning-based) adaptation methods on five knowledge-intensive benchmarks shows that in the oracle setting, Cartridges are the most accurate injection method at nearly every storage budget, outperforming parametric methods by 10 points.
P PuzzleKV is proposed, a training- and calibration-free method that treats each completed page as an independent compression unit that achieves more than 96% of Full KV performance across both evaluated models and all benchmark settings, with substantial gains over Global SVD on RULER and competitive performance on Lo...
Zi-Zhong J. Wang, Jie-Ying Wang, Zhao Zhang et al.· 0 citations
A combined approach is proposed that fine-tunes the model while taking KV cache concatenation into account and selectively recomputes a subset of the KV caches and demonstrates improved accuracy for long-context inputs.
F. Tachibana, Daisuke Miyashita, Jun Deguchi· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.