The Truth is in There: Improving Reasoning in Language Models with Layer-Selective Rank Reduction

arxiv:2312.13558

Original source · Paper JSON

Reasoning and the "physics" of language models · 2023-12-21

Cites (1) · Cited by (0)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3
4. Part 4

Contents

• The Truth is in There: Improving Reasoning in Language Models with Layer-Selective Rank Reduction
• Abstract
• 1 Introduction
• 2 Related work
• How facts are stored.
• Model compression.
• Low-rank approximations of weight matrices.
• Model distillation and low-rank training.
• 3 Preliminaries
• Maths Notation.
• Transformer Architecture.
• Rank- r r Approximation and SVD.
• 4 LA yer SE lective R ank Reduction ( 𝙻𝙰𝚂𝙴𝚁 {\tt LASER} )
• 5 Experiments
• GPT-J, CounterFact and PILE.
• 5.1 A Thorough Analysis with GPT-J on the CounterFact Dataset
• Improved accuracy and robustness to paraphrases.
• Effect on language modeling and fluency.
• Composing reductions across layers.
• 5.1.1 Which facts in the dataset are recovered by rank reduction?
• 5.1.2 What are higher-order components storing?
• 5.2 How generally does this hold?
• Natural Language Understanding Tasks.
• Evaluation metrics.
• 5.3 Non-text domains
• Policy learning.
• 6 Conclusion and Discussion
• Acknowledgements
• Appendix
• Appendix A Dataset Details
• CounterFact.
• PILE.
• HotpotQA.
• FEVER.
• Bias in Bios.
• BigBench-Epistemic Reasoning:
• TruthfulQA:
• QA Wiki Data:
• Appendix B Details of the Decision Transformer Domain
• Sokoban Details.
• Appendix C Extended Analysis
• C.1 Are weight matrices already low-rank?
• C.2 How much reduction is too much?
• C.3 Does 𝙻𝙰𝚂𝙴𝚁 {\tt LASER} select the same layer for different tasks?
• C.4 Measuring Perplexity on PILE.
• C.5 Final 𝙻𝙰𝚂𝙴𝚁 {\tt LASER} search results
• Appendix D Alternate Pruning Methods
• Appendix E Implementation Details
• E.1 Dataset Processing Details
• CounterFact.
• Hotpot.
• Fever.
• Bios Gender.
• Bios Profession.
• BigBench Epistemic Reasoning.
• Truthful QA.
• BigBench Wikidata QA.
• E.2 Details for Computing Accuracy and Log Loss
• CounterFact.
• HotPotQA.
• Fever.
• Bios Gender.
• Bios Profession.
• BigBench Epistemic Reasoning.
• TruthfulQA.
• BigBench WikidataQA.
• E.3 Code
• E.4 Compute Details
• Instructions for reporting errors