{"id":106,"identifier":"LOR-00106","url":"https://lifeofresearch.com/articles/106","title":"Lossless at What Length? Near-Lossless KV Cache Compression Is a Claim Indexed by Decode Exposure, Scoring Statistic and Budget Accounting, the Decode-Aware Line Has Answered the First Index Mostly on Final-Answer Accuracy, and Divergence From the Full Cache Is Not by Itself the Defect","abstract":"(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0.\nKey-value cache compression is routinely reported as near-lossless at aggressive budgets: eviction, quantisation and merging methods keep a fifth or a tenth of the cache and match the full-cache score. A 2025-2026 line of work argues that the result was manufactured by evaluation, because the benchmarks that established it are dominated by long prompts and short answers. In this analysis the author reviews and synthesises published results to show that the critique is right about what the headline claims omit and too strong about what follows. A losslessness claim carries three indices the headline rarely states: how many tokens are decoded under the compressed state, which statistic scores the output, and how the budget is accounted. On the first, the same eviction method that scores 100 percent on a 7-digit passkey scores 35 percent on a 64-digit one, and sequence-level divergence from the full cache can only accumulate with length. But decode-aware methods evaluated on reasoning traces of thousands of tokens report near-lossless accuracy, so long decoding does not by itself defeat compression. On the second, aggregate accuracy hides instruction-specific eviction, per-sample failures and rationales that no longer support correct answers; yet divergence from the full cache is not the defect either, since the uncompressed reference itself moves under numerical nondeterminism. On the third, one method's \"10 percent\" and \"34 percent\" lossless ratios are fixed budgets of 1,536 and 1,024 tokens divided by different mean lengths. The paper states a reporting procedure and the measurements that would settle what remains open.","authors":[{"name":"Pranay Mahendrakar","affiliation":"SONYTECH","orcid":"0009-0003-7224-029X"}],"type":"article","license":"CC-BY-4.0","keywords":["KV cache compression","token eviction","KV cache quantization","long-context inference","reasoning models","evaluation methodology","instruction following","LLM serving"],"publication_date":"2026-10-13","version":1,"views":6,"downloads":6,"created_at":"2026-10-14 08:49:41","doi":"10.5281/zenodo.23343116","source":"Zenodo","source_url":"https://zenodo.org/records/23343116","files":[{"id":108,"name":"lossless-at-what-length.pdf","size":392947,"mime":"application/pdf","downloads":6,"download_url":"https://lifeofresearch.com/articles/106/files/108/lossless-at-what-length.pdf"}]}