Large Language Models (LLMs) often produce hallucinations in retrieval-augmented or long-context generation, even when relevant evidence is present. This stems from two issues: head importance is treated as input-agnostic, and raw attention weights poorly reflect each token's true contribution. We present HAVE (Head-Adaptive Gating and ValuE Calibration), a parameter-free decoding framework that directly addresses both challenges. HAVE introduces head-adaptive gating, which performs instance-level soft reweighing of attention heads, and value calibration, which augments attention with the magnitude of value vectors to approximate write-back contribution. Together, these modules construct token-level evidence aligned with model updates and fuse it with the LM distribution through a lightweight uncertainty-scaled policy. HAVE requires no finetuning and operates in a single forward pass, making it efficient and broadly applicable. Experiments across multiple QA benchmarks and LLM families demonstrate that HAVE consistently reduces hallucinations and outperforms strong baselines, including DAGCD, with modest overhead. The framework is transparent, reproducible, and readily integrates with off-the-shelf LLMs, advancing trustworthy generation in real-world settings.

I ns ti t ut Gal i lé e - U niv. P ar is - Nord 93430 V il l et ane us e - F RA NC E noui ouaf @l ipn.uni v-pa ri s 13.fr G eneral ly s peaking, form al NL s em antic s i s re ferenti al i .e. it as sum es t hat i t is pos si ble t o c reate a s tati c dis course uni verse and to equat e t he obj ect s of t his uni verse t o the (s tat ic) mea nings of w ords . The me aning of a sent ence is then buil t from t he me anings of the w ords in a c ompos iti onal proces s and the se mant ic inte rpretat ion of a s entenc e i s reduce d to it s logic al i nterpret ati on bas ed on t he t ruth condit ions . The very diffic ult tas k of ada pting the mea ning of a s ent ence to its c ontext is often left to the pragm ati c l evel, and this tas k re quires t o us e a huge a mount of com mon s ens e know ledge a bout the domai n. It has bee n s howe d t hat the above tri-pa rtit ion i s very arti fici al becaus e l inguis ti c a s we ll as e xtra-li nguis tic know ledge i nterac t i n t he s am e gl obal proces s to provide the ne ces sa ry elem ents for unders ta nding. But what kind of rea soni ng is needed for na tural language se manti cs? T he ans we r to thi s que st ion is bas ed on the remark t hat t exts s eldom provide norma l det ail s t hat are a ss umed to be known to the reader.