Jungseob Lee, Seungyoon Lee, Hyeonseok Moon, Seongtae Hong, Suhyune Son, Sugyeong Eo, Minhyuk Kim, Chanjun Park · 2026 · arXiv
DOI: 10.48550/arXiv.2610.29727 · open-access PDF
Les grands modèles linguistiques répondent souvent à partir des connaissances acquises au cours de la formation, même quand un passage fourni contredit ce savoir. Pour résoudre ce problème, le décodage contrastant contraste les prédictions sous le plein et une prompt modifiée à chaque token, mais il modifie également les réponses correctes que le passage soutient. Nous montrons que vérifier les réponses contre le passage supprime ce compromis. Sur la base de cette observation, nous proposons Accept-or-Repair (AoR), une procédure sans formation. AoR conserve une réponse échantillonnée si le passage le soutient.
Machine translation (Meta M2M100); the English original follows.
Large language models often answer from knowledge learned during training, even when a provided passage contradicts that knowledge. To address this problem, contrastive decoding contrasts the predictions under the full and a modified prompt at every token, but it also alters correct answers that the passage supports. We show that checking answers against the passage removes this trade-off. Based on this observation, we propose Accept-or-Repair (AoR), a training-free procedure. AoR keeps a sampled answer if the passage supports it. Otherwise, it tries up to two replacements with contrastive dec
Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese
GET https://store.agentexchange.work/gw/papers/verify?title=...&year=... (0.003 USD), fetch this record as JSON with /gw/papers/paper?id=2610.29727, or list what it cites with /gw/papers/references?id=.... Pay per call in USDC (x402) or with a 19 USD card key for 1,000 calls. Catalogue · llms.txtSources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange