{"id":"arxiv:2606.03982","title":"Language Models Compare Quantities Using Number-specific and Unit-specific Heuristics","date":"2026-06-02","section":"Reasoning and the \"physics\" of language models","quality":5.9e-05,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/","source_url":"https://arxiv.org/abs/2606.03982","text_available":true,"authors":["Mutsumi Sasaki","Go kamoda","Ryosuke Takahashi","Kosuke Sato"],"outgoing":[{"source":"arxiv:2606.03982","target":"arxiv:2301.05217","evidence":"title","id":"arxiv:2301.05217","title":"Progress measures for grokking via mechanistic interpretability","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f71e3b65a8b3f35beb171d17f6c65878e0914227a0042ccba59b88952554baef/"},{"source":"arxiv:2606.03982","target":"arxiv:2406.03445","evidence":"title","id":"arxiv:2406.03445","title":"Pre-trained Large Language Models Use Fourier Features to Compute Addition","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/169a3e64467f1e0862508b86b2198187fc8eeb3ab114a47d2bd4caaea4a9ee2b/"},{"source":"arxiv:2606.03982","target":"arxiv:2410.21272","evidence":"title","id":"arxiv:2410.21272","title":"Arithmetic Without Algorithms: Language Models Solve Math With a Bag of Heuristics","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/"},{"source":"arxiv:2606.03982","target":"arxiv:2502.00873","evidence":"title","id":"arxiv:2502.00873","title":"Language Models Use Trigonometry to Do Addition","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/7766b2640c6b5ab53a295e8d60920cb2419bfbbd812810d9b35d4b7c323922ed/"},{"source":"arxiv:2606.03982","target":"arxiv:2512.13961","evidence":"id","id":"arxiv:2512.13961","title":"Olmo 3","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/"}],"incoming":[],"text_parts":[{"start":0,"end":15700,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-1.html"},{"start":15700,"end":31617,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-2.html"},{"start":31617,"end":47530,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-3.html"},{"start":47530,"end":63390,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-4.html"}],"contents":[{"title":"Language Models Compare Quantities Using Number-specific and Unit-specific Heuristics","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-1.html#text"},{"title":"Abstract","offset":89,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-1.html#text"},{"title":"1 Introduction","offset":928,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-1.html#text"},{"title":"2 Task Formulation and Overall Experimental Setup","offset":5372,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-1.html#text"},{"title":"2.1 Task Formulation","offset":5748,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-1.html#text"},{"title":"2.2 Overall Experimental Setup","offset":8424,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-1.html#text"},{"title":"3 Behavioral Observation: LMs Become Less Accurate Near the Boundary","offset":9636,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-1.html#text"},{"title":"3.1 Experimental Setup","offset":9852,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-1.html#text"},{"title":"3.2 Results","offset":10511,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-1.html#text"},{"title":"4 Cue-Combination Hypothesis: Quantity Decisions Depend on Cue Agreement","offset":11537,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-1.html#text"},{"title":"5 Surrogate Analysis: Number and Unit Differences Best Predict LM Behavior","offset":14398,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-1.html#text"},{"title":"5.1 Experimental Setup","offset":14553,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-1.html#text"},{"title":"Implementation details.","offset":18660,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-2.html#text"},{"title":"5.2 Results","offset":19478,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-2.html#text"},{"title":"6 Causal Analysis: Number and Unit Differences Steer LM Decisions","offset":21857,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-2.html#text"},{"title":"6.1 Experimental Setup","offset":22161,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-2.html#text"},{"title":"6.2 Results","offset":26329,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-2.html#text"},{"title":"7 Related Work","offset":28303,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-2.html#text"},{"title":"8 Conclusion","offset":30720,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-2.html#text"},{"title":"Limitations","offset":31601,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-2.html#text"},{"title":"Ethical Considerations","offset":33236,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-3.html#text"},{"title":"Appendix A Task Formulation and Overall Experimental Setup","offset":33795,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-3.html#text"},{"title":"A.1 Prompt Templates","offset":34189,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-3.html#text"},{"title":"A.2 Choice of Prompt for Main Analyses","offset":35497,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-3.html#text"},{"title":"A.3 Generation and Evaluation","offset":36350,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-3.html#text"},{"title":"A.4 Effects of Number Formatting, Unit Notation, and Tokenization","offset":37064,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-3.html#text"},{"title":"A.5 Motivation for Analyzing the Quantity Margin in Log Space","offset":38426,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-3.html#text"},{"title":"Appendix B On the Terms “Heuristic”, “Heuristic Aggregation”, and “Bag of Heuristics”","offset":40356,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-3.html#text"},{"title":"Heuristic.","offset":41800,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-3.html#text"},{"title":"Heuristic Aggregation.","offset":42501,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-3.html#text"},{"title":"Appendix C Behavioral Observation: LMs Become Less Accurate Near the Boundary","offset":43031,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-3.html#text"},{"title":"C.1 Results Across LMs","offset":43518,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-3.html#text"},{"title":"C.2 Effects of Prompt Templates","offset":44438,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-3.html#text"},{"title":"C.3 Effects of Unit Notation","offset":45884,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-3.html#text"},{"title":"C.4 Generated Outputs from Reasoning LMs","offset":46834,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-3.html#text"},{"title":"Appendix D Cue-Combination Hypothesis: Quantity Decisions Depend on Cue Agreement","offset":47444,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-3.html#text"},{"title":"Appendix E Surrogate Analysis: Number and Unit Differences Best Predict LM Behavior","offset":48529,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-4.html#text"},{"title":"E.1 Surrogate Feature Design","offset":49143,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-4.html#text"},{"title":"E.2 Additional Surrogate Results","offset":52164,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-4.html#text"},{"title":"Appendix F Causal Analysis: Number and Unit Differences Steer LM Decisions","offset":54143,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-4.html#text"},{"title":"F.1 Distributed Alignment Search","offset":55103,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-4.html#text"},{"title":"F.2 Robustness Across LMs, Unit Settings, Prompts, and training random seeds","offset":58563,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-4.html#text"},{"title":"F.3 Robustness to Intervention Dimensionality","offset":60929,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-4.html#text"},{"title":"Instructions for reporting errors","offset":62426,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/part-4.html#text"}],"text_code_points":63390,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/source.md"}