2026-10-06T16:46:39Z https://api.figshare.com/v2/oai

oai:figshare.com:article/34021311 2026-09-29T05:07:32Z category_35549 category_35903 portal_1005 item_type_14 month_year_09_2026

Retrieval-Augmented Decoding for Improving Truthfulness in Open-Ended Generation Huu Nguyen (19957053) Sunil Gupta (309664) Hung Le (3029535) Information and computing sciences Machine learning Minority Health Health Disparities and Racial or Ethnic Minority Health Research Generic health relevance Ensuring truthfulness in large language models (LLMs) remains a critical challenge for reliable text generation. While supervised fine-tuning and reinforcement learning with human feedback have shown promise, they require a substantial amount of annotated data and computational resources, limiting scalability. In contrast, decoding-time interventions offer lightweight alternatives without model retraining. However, existing decoding strategies often face issues like prompt sensitivity, limited generalization, or dependence on internal model states. We propose Retrieval-Augmented Decoding (RAD), a context-aware adaptive decoding method that leverages a compact reference grounding space built from as few as 10 annotated examples and comprising pairs of context embeddings and next-token logits from truthful responses, to enable retrieval-based logit shaping during inference. At each decoding step, RAD retrieves high-quality semantically similar contexts from this grounding space and aggregates their associated next token logits to modify the model’s current logits. Across four open-ended generation benchmarks and four LLMs, our method consistently outperforms strong baselines and shows robust cross-task generalization, underscoring the promise of context-aware decoding for enhancing factual reliability.<p></p> 2027-01-01T00:00:00Z Text Conference contribution 10.26187/deakin.34021311 https://figshare.com/articles/conference_contribution/Retrieval-Augmented_Decoding_for_Improving_Truthfulness_in_Open-Ended_Generation/34021311 All Rights Reserved Restricted Access