Running summarizing tools on a new article

Abstract: Retrieval-Augmented Generation (RAG) grounds large language model responses in externally retrieved evidence, but many RAG systems use a fixed generation strategy for all queries. This paper presents the Hallucination-Penalty RL-RAG Bandit, which treats generation-strategy selection as a five-armed non-contextual bandit problem. Five prompting configurations (A0–A4) differ in response style, citation emphasis, abstention behaviour, and temperature. For each training interaction, an epsilon-greedy policy selects one strategy before generation, and only the selected arm is updated using an automatic substantive-unit-level semantic-support reward computed against the retrieved context. The learned policy is compared with fixed-strategy, citation-enforced, uniform-random, and generate-all score-max baselines. Results show that the bandit learns competitive strategy preferences under the proposed proxy, although the final greedy action varies across random seeds. On held-out evaluation, the primary greedy policy using A2 achieved a mean reward of 0.667, matching the validation-selected best-fixed strategy, while the five-call generate-all comparator achieved 0.726. An independent automatic NLI analysis showed only moderate agreement with the semantic-support metric. The findings therefore support reward-driven strategy selection as an experimental framework but do not establish a universal advantage over fixed strategy selection or independently verified hallucination reduction.