{"paper":{"title":"Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Standard DPO can lower the absolute likelihood of preferred responses while still raising their relative odds; a modified loss called DPOP prevents the drop and improves results.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Arka Pal, Colin White, Deep Karkhanis, Manley Roberts, Samuel Dooley, Siddartha Naidu","submitted_at":"2024-02-20T18:42:34Z","abstract_excerpt":"Direct Preference Optimisation (DPO) is effective at significantly improving the performance of large language models (LLMs) on downstream tasks such as reasoning, summarisation, and alignment. Using pairs of preferred and dispreferred data, DPO models the relative probability of picking one response over another. In this work, first we show theoretically that the standard DPO loss can lead to a reduction of the model's likelihood of the preferred examples, as long as the relative probability between the preferred and dispreferred classes increases. We then show empirically that this phenomeno"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"we find that DPOP outperforms DPO and other fine-tuning procedures across a wide variety of datasets and downstream tasks, including datasets with high edit distances between completions. Furthermore, we find that the DPOP-tuned model outperforms the DPO-tuned model (all else equal) on benchmarks independent of the fine-tuning data, such as MT-Bench. Finally, using DPOP, we create and open-source Smaug-34B and Smaug-72B, with the latter becoming the first open-source LLM to surpass an average accuracy of 80% on the HuggingFace Open LLM Leaderboard.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the identified DPO failure mode is the primary limiter of performance and that the DPOP modification will generalize without introducing new unintended effects on model behavior or other tasks.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"DPOP is a new loss function that prevents DPO from lowering preferred response likelihoods and outperforms standard DPO on diverse datasets, MT-Bench, and enables Smaug-72B to exceed 80% on the Open LLM Leaderboard.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Standard DPO can lower the absolute likelihood of preferred responses while still raising their relative odds; a modified loss called DPOP prevents the drop and improves results.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"ec60584343cd8d40203e874fc411a386cb546186d0de056ee89dbb35e871b04b"},"source":{"id":"2402.13228","kind":"arxiv","version":2},"verdict":{"id":"75e7dd77-7a29-47ef-9ca9-013f52e8463f","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T22:59:45.344560Z","strongest_claim":"we find that DPOP outperforms DPO and other fine-tuning procedures across a wide variety of datasets and downstream tasks, including datasets with high edit distances between completions. Furthermore, we find that the DPOP-tuned model outperforms the DPO-tuned model (all else equal) on benchmarks independent of the fine-tuning data, such as MT-Bench. Finally, using DPOP, we create and open-source Smaug-34B and Smaug-72B, with the latter becoming the first open-source LLM to surpass an average accuracy of 80% on the HuggingFace Open LLM Leaderboard.","one_line_summary":"DPOP is a new loss function that prevents DPO from lowering preferred response likelihoods and outperforms standard DPO on diverse datasets, MT-Bench, and enables Smaug-72B to exceed 80% on the Open LLM Leaderboard.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the identified DPO failure mode is the primary limiter of performance and that the DPOP modification will generalize without introducing new unintended effects on model behavior or other tasks.","pith_extraction_headline":"Standard DPO can lower the absolute likelihood of preferred responses while still raising their relative odds; a modified loss called DPOP prevents the drop and improves results."},"references":{"count":288,"sample":[{"doi":"","year":2024,"title":"01.AI. Yi-34b-200k, 2024. URL https://huggingface.co/01-ai/Yi-34B-200K","work_id":"9b96eb44-1e33-47dc-9c92-3c378b36d521","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Ultrafeedback binarized clean, 2024","work_id":"3e5fab3c-daf5-4574-b2d5-77651d6cc25a","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Learning from mistakes makes llm better reasoner","work_id":"59b539e9-2013-4930-92f1-8b4ea134cb51","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"A general theoretical paradigm to understand learning from human preferences","work_id":"1566c4f1-846f-4750-8161-dc5cd585eee0","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Training a helpful and harmless assistant with reinforcement learning from human feedback, 2022","work_id":"6b9cf002-ab59-4c61-ae20-2d2f7b0eecaf","ref_index":6,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":288,"snapshot_sha256":"22f81e3a140139e64d421cbf8ab57f148d21567d1b5546fb3bf230dcbd3ab15c","internal_anchors":31},"formal_canon":{"evidence_count":3,"snapshot_sha256":"1121ff1a48e4fc4b4dac903e6a183d08718a8e713ccb1a42806682ef4c22a2bf"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}