{"id":"arxiv:2312.13558","title":"The Truth is in There: Improving Reasoning in Language Models with Layer-Selective Rank Reduction","date":"2023-12-21","section":"Reasoning and the \"physics\" of language models","quality":0.025953,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/","source_url":"https://arxiv.org/abs/2312.13558","text_available":true,"authors":["Pratyusha Sharma","Jordan T. Ash","Dipendra Misra"],"outgoing":[{"source":"arxiv:2312.13558","target":"arxiv:1803.03635","evidence":"title","id":"arxiv:1803.03635","title":"The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a9d7f0235205afcc709e3e74f28cc7707831bcb1d46239c418cc7e4b624d23bd/"}],"incoming":[],"text_parts":[{"start":0,"end":15285,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-1.html"},{"start":15285,"end":31015,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-2.html"},{"start":31015,"end":46502,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html"},{"start":46502,"end":52331,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-4.html"}],"contents":[{"title":"The Truth is in There: Improving Reasoning in Language Models with Layer-Selective Rank Reduction","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-1.html#text"},{"title":"Abstract","offset":101,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-1.html#text"},{"title":"1 Introduction","offset":1220,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-1.html#text"},{"title":"2 Related work","offset":5540,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-1.html#text"},{"title":"How facts are stored.","offset":5838,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-1.html#text"},{"title":"Model compression.","offset":7561,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-1.html#text"},{"title":"Low-rank approximations of weight matrices.","offset":9128,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-1.html#text"},{"title":"Model distillation and low-rank training.","offset":10107,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-1.html#text"},{"title":"3 Preliminaries","offset":10913,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-1.html#text"},{"title":"Maths Notation.","offset":11011,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-1.html#text"},{"title":"Transformer Architecture.","offset":11699,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-1.html#text"},{"title":"Rank- r r Approximation and SVD.","offset":15245,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-1.html#text"},{"title":"4 LA yer SE lective R ank Reduction ( 𝙻𝙰𝚂𝙴𝚁 {\\tt LASER} )","offset":17121,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-2.html#text"},{"title":"5 Experiments","offset":19322,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-2.html#text"},{"title":"GPT-J, CounterFact and PILE.","offset":19861,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-2.html#text"},{"title":"5.1 A Thorough Analysis with GPT-J on the CounterFact Dataset","offset":20333,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-2.html#text"},{"title":"Improved accuracy and robustness to paraphrases.","offset":21845,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-2.html#text"},{"title":"Effect on language modeling and fluency.","offset":23416,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-2.html#text"},{"title":"Composing reductions across layers.","offset":24074,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-2.html#text"},{"title":"5.1.1 Which facts in the dataset are recovered by rank reduction?","offset":25089,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-2.html#text"},{"title":"5.1.2 What are higher-order components storing?","offset":25940,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-2.html#text"},{"title":"5.2 How generally does this hold?","offset":28551,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-2.html#text"},{"title":"Natural Language Understanding Tasks.","offset":28696,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-2.html#text"},{"title":"Evaluation metrics.","offset":30293,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-2.html#text"},{"title":"5.3 Non-text domains","offset":31346,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"Policy learning.","offset":31548,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"6 Conclusion and Discussion","offset":32435,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"Acknowledgements","offset":34376,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"Appendix","offset":34683,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"Appendix A Dataset Details","offset":34696,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"CounterFact.","offset":34727,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"PILE.","offset":35293,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"HotpotQA.","offset":35916,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"FEVER.","offset":36428,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"Bias in Bios.","offset":36787,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"BigBench-Epistemic Reasoning:","offset":37209,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"TruthfulQA:","offset":37846,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"QA Wiki Data:","offset":38691,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"Appendix B Details of the Decision Transformer Domain","offset":39054,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"Sokoban Details.","offset":39112,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"Appendix C Extended Analysis","offset":39605,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"C.1 Are weight matrices already low-rank?","offset":39638,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"C.2 How much reduction is too much?","offset":40265,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"C.3 Does 𝙻𝙰𝚂𝙴𝚁 {\\tt LASER} select the same layer for different tasks?","offset":40926,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"C.4 Measuring Perplexity on PILE.","offset":41254,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"C.5 Final 𝙻𝙰𝚂𝙴𝚁 {\\tt LASER} search results","offset":41838,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"Appendix D Alternate Pruning Methods","offset":42561,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"Appendix E Implementation Details","offset":43060,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"E.1 Dataset Processing Details","offset":43098,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"CounterFact.","offset":43678,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"Hotpot.","offset":44042,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"Fever.","offset":44639,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"Bios Gender.","offset":45156,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"Bios Profession.","offset":45546,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"BigBench Epistemic Reasoning.","offset":46184,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"Truthful QA.","offset":46482,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-3.html#text"},{"title":"BigBench Wikidata QA.","offset":47296,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-4.html#text"},{"title":"E.2 Details for Computing Accuracy and Log Loss","offset":47564,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-4.html#text"},{"title":"CounterFact.","offset":48993,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-4.html#text"},{"title":"HotPotQA.","offset":49196,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-4.html#text"},{"title":"Fever.","offset":49627,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-4.html#text"},{"title":"Bios Gender.","offset":49786,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-4.html#text"},{"title":"Bios Profession.","offset":49952,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-4.html#text"},{"title":"BigBench Epistemic Reasoning.","offset":50137,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-4.html#text"},{"title":"TruthfulQA.","offset":50334,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-4.html#text"},{"title":"BigBench WikidataQA.","offset":50498,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-4.html#text"},{"title":"E.3 Code","offset":50721,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-4.html#text"},{"title":"E.4 Compute Details","offset":51026,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-4.html#text"},{"title":"Instructions for reporting errors","offset":51367,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/part-4.html#text"}],"text_code_points":52331,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e6ddb16378e3684d190823b8f271c8c3273b47f7bada2fc2c692b828d40e3384/source.md"}