[ { "id": "length-regularized-dpo", "title": "Disentangling Length from Quality in Direct Preference Optimization", "authors": [ "Ryan Park", "Rafael Rafailov", "Stefano Ermon", "Chelsea Finn" ], "date": "2024-08", "url": "https://aclanthology.org/2024.findings-acl.297/", "kind": "strategy-evidence", "method": "Pythia 2.8B was trained with standard and length-regularized DPO on dialogue and summarization preferences. Evaluation compared 256 generated answers per setting using GPT-4 judgments.", "finding": "Regularization reduced answer length and improved preference results against settings producing similarly long answers. Uncontrolled preference optimization amplified verbosity.", "limits": "One model size, two datasets and a model judge. This training intervention does not establish that a shortness prompt improves documentation.", "application": "Voice implication: compare alternatives at similar lengths and assess task completion separately. Treat length as a diagnostic, preserving explanations and qualifications the reader needs.", "readingDepth": "Methods and results read in the ACL proceedings PDF: sections 3–4, Figure 3, section 7 and Appendix C. Reviewed 2026-09-12.", "relatedRuleIds": [ "repetition", "task-relevance", "exhaustive-checklist", "format-fit" ], "apiUse": { "availableNow": [ { "symbol": "composeWritingReviewPrompt", "purpose": "State the reader task and select relevance or repetition guidance." } ], "currentUse": "A host can request alternatives without imposing a blanket shortness rule. The API does not optimize a model, judge completion or compare answer lengths.", "missingCapability": "Compare alternatives using explicit task completion and retained qualifications alongside length, with fixtures that distinguish useful explanation from unnecessary expansion.", "releaseFunctionIds": [ "alternative-comparison", "prompt-benchmarks" ] } }, { "id": "self-refine", "title": "Self-Refine: Iterative Refinement with Self-Feedback", "authors": [ "Aman Madaan", "Niket Tandon", "Prakhar Gupta", "Skyler Hallinan", "Luyu Gao", "Sarah Wiegreffe", "Uri Alon", "Nouha Dziri", "Shrimai Prabhumoye", "Yiming Yang", "Shashank Gupta", "Bodhisattwa Prasad Majumder", "Katherine Hermann", "Sean Welleck", "Amir Yazdanbakhsh", "Peter Clark" ], "date": "2023", "url": "https://proceedings.neurips.cc/paper_files/paper/2023/file/91edff07232fb1b55a505a9e9f6c0ff3-Paper-Conference.pdf", "kind": "strategy-evidence", "method": "The same model generated, critiqued and revised outputs using task-specific prompts across seven tasks. Evaluation combined task metrics, model judgments and blinded author judgments.", "finding": "Refinement improved the tested task results; human judges preferred refined outputs in dialogue, sentiment reversal, acronym generation and code readability.", "limits": "English tasks and mainly proprietary 2023 models. Human evaluation usually used one author judgment per example. Better dialogue can also mean more elaborate text.", "application": "Voice implication: request a specific, evidence-linked critique before alternatives. Keep revision bounded, compare against the original and allow the reviewer to retain it.", "readingDepth": "NeurIPS 2023 full-text sections 2–3 and 6, Appendix C human evaluation and Appendix P dialogue results read. Reviewed 2026-09-12.", "relatedRuleIds": [ "reader-goal", "repetition", "claims-evidence", "task-relevance" ], "apiUse": { "availableNow": [ { "symbol": "composeWritingReviewPrompt", "purpose": "Prepare a bounded critique request with grounded alternatives." }, { "symbol": "selectWritingRules", "purpose": "Narrow a subsequent review to the relevant rule IDs." } ], "currentUse": "Hosts must submit prompts and manage any revision loop themselves. maxFindings requests an output limit; it neither enforces compliance nor bounds inference spending.", "missingCapability": "Validate each critique and compose scoped review batches with explicit usage budgets before a host runs another round.", "releaseFunctionIds": [ "result-validation", "review-batching", "usage-budget" ] } }, { "id": "self-correction-counterevidence", "title": "Large Language Models Cannot Self-Correct Reasoning Yet", "authors": [ "Jie Huang", "Xinyun Chen", "Swaroop Mishra", "Huaixiu Steven Zheng", "Adams Wei Yu", "Xinying Song", "Denny Zhou" ], "date": "2024-03-14", "url": "https://arxiv.org/html/2310.01798v2", "kind": "strategy-evidence", "method": "Experiments compared initial answers and up to two self-correction rounds on reasoning benchmarks, separating feedback with known correct answers from model-only feedback.", "finding": "Without correctness labels, the tested models often changed correct answers into errors. Stronger initial prompts also reduced apparent benefits attributed to refinement.", "limits": "The ICLR 2024 study tests older models and reasoning tasks. Its authors explicitly distinguish style preferences, where self-correction may help.", "application": "Voice implication: a model's critique is a proposal, not verification. Check revised claims against source evidence and retain the original when the critique lacks support.", "readingDepth": "Version 2, dated 2024-03-14: experimental setup, Tables 2–6, prompt comparison in section 5 and section 7 limits read. Reviewed 2026-09-12.", "relatedRuleIds": [ "claims-evidence", "reported-evidence", "unsupported-causality", "calibrated-uncertainty" ], "apiUse": { "availableNow": [ { "symbol": "composeWritingReviewPrompt", "purpose": "Ask for missing evidence and justified keep decisions." }, { "symbol": "getWritingRule", "purpose": "Retrieve claims-evidence and reported-evidence requirements." } ], "currentUse": "These instructions can guide a second opinion, but the library has no fact checker and cannot establish that a revised answer is better.", "missingCapability": "Verify each proposed claim change against supplied evidence and compare the revision with the original before accepting any correction.", "releaseFunctionIds": [ "claim-evidence", "alternative-comparison", "result-validation" ] } }, { "id": "coauthor-writer-control", "title": "CoAuthor: Designing a Human-AI Collaborative Writing Dataset for Exploring Language Model Capabilities", "authors": [ "Mina Lee", "Percy Liang", "Qian Yang" ], "date": "2022-01-25", "url": "https://arxiv.org/html/2201.06796v2", "kind": "strategy-evidence", "method": "Sixty-three writers completed 1,445 English writing sessions with GPT-3 suggestions. The interface logged requests, selections, dismissals and edits; surveys measured ownership and satisfaction.", "finding": "Perceived ownership correlated with the fraction written by the writer. Editing activity alone showed little relationship with ownership; collaboration varied considerably between writers.", "limits": "This CHI 2022 dataset offers correlations, not a randomized test of intent preservation. Participants were qualified crowd workers writing short assigned stories and essays.", "application": "Voice implication: make alternatives optional and record keep, accept and edit decisions. Evaluate author intent directly instead of treating acceptance rate as success.", "readingDepth": "Version 2, dated 2022-01-25: interface and recruitment in section 4, ownership analysis in 5.2 and discussion in section 6 read. Reviewed 2026-09-12.", "relatedRuleIds": [ "reader-goal", "claims-evidence" ], "apiUse": { "availableNow": [ { "symbol": "mountVoiceReview", "purpose": "Expose selection and finding callbacks to the host's review interface." }, { "symbol": "composeWritingReviewPrompt", "purpose": "Request optional alternatives or a justified decision to keep." } ], "currentUse": "Hosts provide the choice panel and persist its actions; mounting the library alone does not record acceptance, dismissal or editing history.", "missingCapability": "Preserve author decisions with source versions and compare alternatives against stated intent, without treating acceptance counts as a quality measure.", "releaseFunctionIds": [ "decision-preservation", "alternative-comparison" ] } } ]