{"slug": "beyond-truncation-entailment-licensed-elision-and-the-limits-of-semantic-in-llm", "title": "Beyond Truncation? Entailment-Licensed Elision and the Limits of Semantic Preservation in LLM Decoding", "summary": "A survey of 23 training-free decoding-time mechanism families found none that selects content better than simply truncating a model's own answer, according to an arXiv paper (2611.08975v1) that also introduces Entailment-Licensed In-Decoding Elision (ELIDE). The paper reports that an entropy-gated logits processor removes 25.5-34.9% of tokens on MT-Bench development questions and 19.9-34.2% on a sealed test split within pre-registered BERTScore non-inferiority margins, but proves and confirms by audit that its output is always a prefix of the unmodified answer. ELIDE beats matched-length truncation on three model families, yet a manual reading of 936 of its elisions across 670 public prompts found 59% harm the answer, a failure no automatic measure detects; scoping the licence to its discourse unit cuts the harmful share to about one elision in ten on two fresh tests.", "body_md": "arXiv:2611.08975v1 Announce Type: new \nAbstract: Instruction-tuned large language models (LLMs) answer at greater length than a request needs. We ask whether a training-free controller inside the decoding loop can remove that surplus while preserving content, judged against the simplest possible baseline: truncating the model's own answer to the same length. An entropy-gated logits processor removes 25.5-34.9% of tokens on MT-Bench development questions and 19.9-34.2% on a sealed test split within pre-registered BERTScore non-inferiority margins, yet we prove and confirm by audit that its output is always a prefix of the unmodified answer: the saving is truncation. A survey of 23 training-free mechanism families finds none that selects content better than truncation without oracle labels. We therefore introduce Entailment-Licensed In-Decoding Elision (ELIDE), which, at each sentence boundary, withholds a span from the reader only when a natural-language inference model judges it entailed by the text already shown, while leaving the model's generation unchanged. ELIDE beats matched-length truncation on three model families, but reading 936 of its elisions on 670 public prompts shows that 59% harm the answer, a failure that no automatic measure detects. Scoping the licence to its discourse unit reduces the harmful share to about one elision in ten on two fresh tests.", "url": "https://wpnews.pro/news/beyond-truncation-entailment-licensed-elision-and-the-limits-of-semantic-in-llm", "canonical_source": "https://arxiv.org/abs/2611.08975", "published_at": "2026-11-09 05:00:00+00:00", "updated_at": "2026-11-09 05:21:06.768203+00:00", "lang": "en", "topics": ["large-language-models", "ai-research", "natural-language-processing", "artificial-intelligence"], "entities": ["ELIDE", "MT-Bench", "BERTScore", "arXiv"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/beyond-truncation-entailment-licensed-elision-and-the-limits-of-semantic-in-llm", "markdown": "https://wpnews.pro/news/beyond-truncation-entailment-licensed-elision-and-the-limits-of-semantic-in-llm.md", "text": "https://wpnews.pro/news/beyond-truncation-entailment-licensed-elision-and-the-limits-of-semantic-in-llm.txt", "jsonld": "https://wpnews.pro/news/beyond-truncation-entailment-licensed-elision-and-the-limits-of-semantic-in-llm.jsonld"}}