Metraipolis › archive › dc_eb5ed34099ed
Compression-Aware Abstention: Teaching LLMs to Refuse When KV-Compression Masks Remove Answer Evidence — arXiv:2608.29934
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2608.29934
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-17T08:07:15.985Z
- SHA-256 of the archived text
14566dd11e3263e7c8b1fd695c8314e68105092980498feee2d30ac77523d4e4- Stable URL of this page
https://metraipolis.com/documento/dc_eb5ed34099ed
How to cite
Compression-Aware Abstention: Teaching LLMs to Refuse When KV-Compression Masks Remove Answer Evidence — arXiv:2608.29934 — arXiv (en), https://arxiv.org/abs/2608.29934 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_eb5ed34099ed (Metraipolis archive dc_eb5ed34099ed, sha256 14566dd11e3263e7c8b1fd695c8314e68105092980498feee2d30ac77523d4e4)
BibTeX
@misc{dceb5ed34099ed,
title = {Compression-Aware Abstention: Teaching LLMs to Refuse When KV-Compression Masks Remove Answer Evidence — arXiv:2608.29934},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2608.29934},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_eb5ed34099ed; sha256 14566dd11e3263e7c8b1fd695c8314e68105092980498feee2d30ac77523d4e4},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.29934. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
KV-cache compression reduces LLM inference memory by evicting context tokens, but when the evicted tokens contain answer-bearing evidence, the model may hallucinate instead of recognizing that the compressed context is insufficient. We address this failure from a behavioral perspective: to our knowledge, this is the first work to formulate compression-aware abstention as a learning problem, in which a model learns to answer when supporting evidence survives compression and abstain when it does not. We construct supervision from compressor survival masks and tight answer-bearing spans, labeling examples as Confident when evidence survives and Abstain when it is removed. A 10.1M-parameter LoRA adapter trained on ~2.6K MuSiQue 2-hop QA examples reduces base-model hallucinations by 97% under prompt-style truncation while preserving correct answering on evidence-retaining examples. Unlike prompt-only abstention baselines, which over-abstain on many answerable high-retention examples, the trained adapter learns a conditional policy. We also evaluate the method under actual compressed-cache decoding, where multi-compressor training yields a 6-22x relative lift over the unaided base on evidence-retaining examples. Controlled-deletion experiments show that the learned behavior is driven by evidence content rather than input length alone. Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-01 in cs.CL). Categorie arXiv: Computation and Language (cs.CL) ; Machine Learning (cs.LG). Pagina del lavoro: https://arxiv.org/abs/2608.29934 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito). Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina. Letto il 2026-10-17T07:47:52.757Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-09-01?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).