{"paper":{"title":"Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Self-RAG trains a single language model to adaptively retrieve passages on demand and critique its own outputs using special reflection tokens.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Akari Asai, Avirup Sil, Hannaneh Hajishirzi, Yizhong Wang, Zeqiu Wu","submitted_at":"2023-10-17T18:18:32Z","abstract_excerpt":"Despite their remarkable capabilities, large language models (LLMs) often produce responses containing factual inaccuracies due to their sole reliance on the parametric knowledge they encapsulate. Retrieval-Augmented Generation (RAG), an ad hoc approach that augments LMs with retrieval of relevant knowledge, decreases such issues. However, indiscriminately retrieving and incorporating a fixed number of retrieved passages, regardless of whether retrieval is necessary, or passages are relevant, diminishes LM versatility or can lead to unhelpful response generation. We introduce a new framework c"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Self-RAG (7B and 13B parameters) significantly outperforms state-of-the-art LLMs and retrieval-augmented models on a diverse set of tasks. Specifically, Self-RAG outperforms ChatGPT and retrieval-augmented Llama2-chat on Open-domain QA, reasoning and fact verification tasks, and it shows significant gains in improving factuality and citation accuracy for long-form generations relative to these models.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That an arbitrary LM can be trained to generate and act on reflection tokens in a way that improves rather than degrades performance across tasks without introducing new failure modes from the added tokens.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Self-RAG trains LLMs to adaptively retrieve passages on demand and self-critique using reflection tokens, outperforming ChatGPT and retrieval-augmented Llama2 on QA, reasoning, and fact verification.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Self-RAG trains a single language model to adaptively retrieve passages on demand and critique its own outputs using special reflection tokens.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"23d566505dfd6411625c27004b8d998868e212164cf9d9736afc9c8f84a533f2"},"source":{"id":"2310.11511","kind":"arxiv","version":1},"verdict":{"id":"baa194fe-ec6e-45a6-bbd4-8745d6e57266","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-12T14:09:42.760870Z","strongest_claim":"Self-RAG (7B and 13B parameters) significantly outperforms state-of-the-art LLMs and retrieval-augmented models on a diverse set of tasks. Specifically, Self-RAG outperforms ChatGPT and retrieval-augmented Llama2-chat on Open-domain QA, reasoning and fact verification tasks, and it shows significant gains in improving factuality and citation accuracy for long-form generations relative to these models.","one_line_summary":"Self-RAG trains LLMs to adaptively retrieve passages on demand and self-critique using reflection tokens, outperforming ChatGPT and retrieval-augmented Llama2 on QA, reasoning, and fact verification.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That an arbitrary LM can be trained to generate and act on reflection tokens in a way that improves rather than degrades performance across tasks without introducing new failure modes from the added tokens.","pith_extraction_headline":"Self-RAG trains a single language model to adaptively retrieve passages on demand and critique its own outputs using special reflection tokens."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2310.11511/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":154,"sample":[{"doi":"","year":2020,"title":"Learning to retrieve reasoning paths over wikipedia graph for question answering","work_id":"2f4e6df6-c1e2-4f5e-80b4-e86b8e29dadd","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Retrieval-based language models and applications","work_id":"8539632b-8089-43ed-b215-005ead5faa6e","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Task-aware retrieval with instructions","work_id":"31b1a479-fc74-4ee5-8290-a68ea7acc584","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Flashattention: Fast and memory-efficient exact attention with io-awareness","work_id":"0a762bb5-5795-45ee-a82d-71bfb1e1f9a4","ref_index":7,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Chain-of-Verification Reduces Hallucination in Large Language Models","work_id":"e9e9983c-a72a-4864-9005-cd09bbd7876c","ref_index":8,"cited_arxiv_id":"2309.11495","is_internal_anchor":true}],"resolved_work":154,"snapshot_sha256":"57d13d5f94e8cf59559f8f5e742c799af4ccd15be3917592ec7e1650e834e232","internal_anchors":23},"formal_canon":{"evidence_count":3,"snapshot_sha256":"45cd1e37665bd876ce638c287082c7a51af5ea5aa951b2e3a37e679baf017c19"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}