{"id":"arxiv:2410.21272","title":"Arithmetic Without Algorithms: Language Models Solve Math With a Bag of Heuristics","date":"2024-10-28","section":"Reasoning and the \"physics\" of language models","quality":0.602412,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/","source_url":"https://arxiv.org/abs/2410.21272","text_available":true,"authors":["Yaniv Nikankin","Anja Reusch","Aaron Mueller","Yonatan Belinkov"],"outgoing":[{"source":"arxiv:2410.21272","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:2410.21272","target":"arxiv:2406.03445","evidence":"id","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:2410.21272","target":"arxiv:2407.21783","evidence":"id","id":"arxiv:2407.21783","title":"The Llama 3 Herd of Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de3cebb4c28bc5d41a843279f219dca14476284991ea5c7c0aeab7068a40e306/"}],"incoming":[{"source":"arxiv:2502.00873","target":"arxiv:2410.21272","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:2502.19981","target":"arxiv:2410.21272","evidence":"id","id":"arxiv:2502.19981","title":"The Lookahead Limitation: Why Multi-Operand Addition is Hard for LLMs","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fc024e63f0774183f41cf84f17bb5a4238ff43a8ec18a6f8afb9a1abcd113488/"},{"source":"arxiv:2508.02513","target":"arxiv:2410.21272","evidence":"id","id":"arxiv:2508.02513","title":"Modular Arithmetic: Language Models Solve Math Digit by Digit","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/83aede757707b76dc76fc15431a436e5dec05563a9f2f28ebcc650775c878bcc/"},{"source":"arxiv:2510.00184","target":"arxiv:2410.21272","evidence":"id","id":"arxiv:2510.00184","title":"Why Can't Transformers Learn Multiplication? Reverse-Engineering Reveals Long-Range Dependency Pitfalls","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/7e395b2ae5bfc2a4d8bfc1e8669b4f324739b6a9b11161cba150bfdffc31746c/"},{"source":"arxiv:2512.16902","target":"arxiv:2410.21272","evidence":"title","id":"arxiv:2512.16902","title":"In-Context Algebra","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5aa7087a88619eafd156aad9217e16f6d1ce284c65534322239be0e02f04e553/"},{"source":"arxiv:2602.10416","target":"arxiv:2410.21272","evidence":"title","id":"arxiv:2602.10416","title":"AI-rithmetic","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a644c16d12f96071c95ea707eb5c5b3cc44be3dda2f9fe49a5f38735695fdad9/"},{"source":"arxiv:2606.03982","target":"arxiv:2410.21272","evidence":"title","id":"arxiv:2606.03982","title":"Language Models Compare Quantities Using Number-specific and Unit-specific Heuristics","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/"}],"text_parts":[{"start":0,"end":14979,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-1.html"},{"start":14979,"end":29819,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-2.html"},{"start":29819,"end":45629,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-3.html"},{"start":45629,"end":61621,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-4.html"},{"start":61621,"end":62979,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-5.html"}],"contents":[{"title":"Arithmetic Without Algorithms: Language Models Solve Math with a Bag of Heuristics","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-1.html#text"},{"title":"Abstract","offset":86,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-1.html#text"},{"title":"1 Introduction","offset":1506,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-1.html#text"},{"title":"2 Arithmetic circuit discovery","offset":5827,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-1.html#text"},{"title":"2.1 Circuit discovery and evaluation","offset":6180,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-1.html#text"},{"title":"Models and Data","offset":6222,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-1.html#text"},{"title":"Method","offset":7979,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-1.html#text"},{"title":"Results","offset":9724,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-1.html#text"},{"title":"2.2 Identifying answer-promoting components","offset":12351,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-1.html#text"},{"title":"3 MLP neurons implement arithmetic heuristics","offset":14196,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-1.html#text"},{"title":"3.1 Decomposing circuit MLPs to individual neurons","offset":14246,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-1.html#text"},{"title":"3.2 MLP neurons act as memorized heuristics","offset":17204,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-2.html#text"},{"title":"4 Arithmetic prompts are answered with a bag of heuristics","offset":20240,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-2.html#text"},{"title":"4.1 Classifying neurons to heuristic types","offset":20807,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-2.html#text"},{"title":"4.2 Heuristic types are combined to answer arithmetic prompts","offset":21927,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-2.html#text"},{"title":"Knocking out neurons by heuristic type.","offset":22295,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-2.html#text"},{"title":"Knocking out neurons by prompt.","offset":24362,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-2.html#text"},{"title":"4.3 Failure modes of the bag of heuristics","offset":25540,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-2.html#text"},{"title":"5 Tracking heuristics development across training steps","offset":27391,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-2.html#text"},{"title":"6 Related work","offset":31125,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-3.html#text"},{"title":"Memorization and generalization in LLMs.","offset":32148,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-3.html#text"},{"title":"Arithmetic reasoning interpretability.","offset":33529,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-3.html#text"},{"title":"7 Conclusions","offset":34628,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-3.html#text"},{"title":"8 Limitations and discussion","offset":36197,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-3.html#text"},{"title":"Acknowledgments","offset":37150,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-3.html#text"},{"title":"Appendix A Llama3-8B circuit evaluation results","offset":38036,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-3.html#text"},{"title":"Appendix B Llama3-8B circuit additional components","offset":39175,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-3.html#text"},{"title":"B.1 MLP0","offset":39442,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-3.html#text"},{"title":"B.2 Attention head patterns","offset":41710,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-3.html#text"},{"title":"Appendix C Multi-layer perceptron implementation details","offset":43903,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-3.html#text"},{"title":"Appendix D Neuron intersection between operators","offset":46208,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-4.html#text"},{"title":"Appendix E Heuristic types descriptions","offset":46873,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-4.html#text"},{"title":"Appendix F Heuristic Neuron Classification Algorithm","offset":51231,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-4.html#text"},{"title":"Appendix G Additional implementation details","offset":53868,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-4.html#text"},{"title":"Arithmetic prompts.","offset":54242,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-4.html#text"},{"title":"Circuit discovery via activation patching.","offset":55239,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-4.html#text"},{"title":"Circuit faithfulness evaluation.","offset":55584,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-4.html#text"},{"title":"Linear probing for correct answers.","offset":56619,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-4.html#text"},{"title":"Heuristic classification.","offset":57257,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-4.html#text"},{"title":"Appendix H Model accuracies on arithmetic prompts","offset":57781,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-4.html#text"},{"title":"Appendix I Results on additional models","offset":58462,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-4.html#text"},{"title":"Appendix J Additional examples for arithmetic heuristics","offset":60922,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-4.html#text"},{"title":"Instructions for reporting errors","offset":62015,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/part-5.html#text"}],"text_code_points":62979,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/397b826ba289bd4f14f457562942cb092c38550a4c49451e95ffcc7b376c8cb0/source.md"}