{"paper":{"title":"Training LLMs with Reinforcement Learning for Intent-Aware Personalized Question Answering","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Reinforcement learning trains LLMs to infer implicit user intent from single-turn questions and generate better-aligned personalized answers.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Benjamin Charles Germain Lee, Jevin West, Maryam Amirizaniani, Nicholas Weber","submitted_at":"2026-05-12T18:38:05Z","abstract_excerpt":"Effective personalized question answering (PQA) in language models requires grounding responses in the user's underlying intent, where intent refers to the implicit ``why'' behind a query beyond its explicit wording. However, existing approaches to intent-aware personalization rely on multi-turn conversational context or rich user profiles, and do not explicitly model user intent during the reasoning process. This limits their effectiveness in single-turn settings, where the user's latent goal must be inferred from minimal input and integrated into the thinking and reasoning process. To bridge"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Through experiments on the LaMP-QA benchmark across six models, IAP consistently outperforms all baselines, achieving an average macro-score gain of around 7.5% over the strongest competitor.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That a tag-based schema combined with a personalized reward function can reliably infer implicit intent from single-turn questions and optimize generation paths to produce better-aligned answers without multi-turn context or user profiles.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"IAP uses RL to train LLMs to explicitly infer and apply implicit user intent in single-turn personalized QA, achieving ~7.5% average macro-score gains over baselines on LaMP-QA.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Reinforcement learning trains LLMs to infer implicit user intent from single-turn questions and generate better-aligned personalized answers.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"dfc034d118c321f692f23d3343f39fac88d2f93d01aab786fa0beab25b22160d"},"source":{"id":"2605.12645","kind":"arxiv","version":1},"verdict":{"id":"19f2dc69-3520-4bfa-98fe-6c6c88e005c6","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:54:25.568151Z","strongest_claim":"Through experiments on the LaMP-QA benchmark across six models, IAP consistently outperforms all baselines, achieving an average macro-score gain of around 7.5% over the strongest competitor.","one_line_summary":"IAP uses RL to train LLMs to explicitly infer and apply implicit user intent in single-turn personalized QA, achieving ~7.5% average macro-score gains over baselines on LaMP-QA.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That a tag-based schema combined with a personalized reward function can reliably infer implicit intent from single-turn questions and optimize generation paths to produce better-aligned answers without multi-turn context or user profiles.","pith_extraction_headline":"Reinforcement learning trains LLMs to infer implicit user intent from single-turn questions and generate better-aligned personalized answers."},"references":{"count":91,"sample":[{"doi":"","year":null,"title":"Second Conference on Language Modeling , year=","work_id":"ffcdbe1d-35f0-4dc4-a785-0b468dbbc059","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Yixin Ye and Zhen Huang and Yang Xiao and Ethan Chern and Shijie Xia and Pengfei Liu , booktitle=. 2025 , url=","work_id":"ad06628e-c7a4-4e8f-9893-bb30e26e388d","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Search-R1: Training","work_id":"ff26bde5-9a7f-4f44-93bd-690d4d4352ba","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.18653/v1/2025.findings-emnlp.84","year":2025,"title":"Legal Mathematical Reasoning with LLM s: Procedural Alignment through Two-Stage Reinforcement Learning","work_id":"dea66ccf-46fa-4f16-b72e-15198d93a238","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.18653/v1/2025.findings-emnlp.95","year":2025,"title":"Beyond Guilt: Legal Judgment Prediction with Trichotomous Reasoning","work_id":"ed49b423-0235-4879-9953-3463dda20901","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":91,"snapshot_sha256":"28bb72d7dae523175e4b1f8838f492128c21e20d99b7f191f51f26b576ad8af6","internal_anchors":9},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}