{"slug": "mathematical-transfer-in-llms-follows-reasoning-approach-more-than-topic", "title": "Mathematical Transfer in LLMs Follows Reasoning Approach More Than Topic", "summary": "Fine-tuning large language models on math problems that share a reasoning approach with the target task outperforms training on same-topic problems, according to an arXiv paper (2610.00331v1) that tested two counterbalanced 2x2 designs covering 2,000 and 800 problems. Same-approach sources beat same-topic sources in all 40 seed-pooled model-target comparisons across five base models and three training seeds per design, with model-level advantages of 8.2 to 16.2 percentage points in the primary design (mean 10.8) and 12.0 to 16.0 in the second (mean 14.3), and all ten model-level 95% confidence intervals excluding zero. The authors report the same-approach advantage runs opposite to embedding and lexical similarity measures, which ranked same-topic sources as closer to targets.", "body_md": "arXiv:2610.00331v1 Announce Type: new \nAbstract: When selecting mathematical training data for LLMs, a natural organizing principle is topic: probability examples for probability targets. An alternative is reasoning approach: worked solutions that share a solution method with the target, even when the mathematical domain differs. We ask which relation produces greater transfer after fine-tuning. We evaluate two counterbalanced $2\\times2$ designs: probability and combinatorics crossed with invariant reasoning and double counting (2,000 problems), and number theory and geometry crossed with complement and pigeonhole reasoning (800 problems). In each design, every cell serves as the held-out target in turn: same-approach (SA) sources share the target's method but change the topic, while same-topic (ST) sources share the topic but change the method. Every source appears once in each role, so additive source-quality effects cancel from the equally weighted aggregate contrast. Across five base models and three training seeds per design, SA outperforms ST in all 40 seed-pooled model--target comparisons. Model-level advantages range from 8.2 to 16.2 percentage points in the primary design (mean: 10.8) and from 12.0 to 16.0 in the second design (mean: 14.3); all ten model-level 95% confidence intervals exclude zero. In both designs, ST sources are more similar to targets under embedding and lexical measures, so the SA advantage runs opposite to the measured ordering of statement-level resemblance. These findings identify reasoning approach as a more effective matching criterion than topic for mathematical transfer across the evaluated topic--approach combinations.", "url": "https://wpnews.pro/news/mathematical-transfer-in-llms-follows-reasoning-approach-more-than-topic", "canonical_source": "https://arxiv.org/abs/2610.00331", "published_at": "2026-10-02 04:00:00+00:00", "updated_at": "2026-10-02 04:15:30.664631+00:00", "lang": "en", "topics": ["large-language-models", "ai-research", "machine-learning", "artificial-intelligence"], "entities": ["arXiv", "2610.00331v1"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/mathematical-transfer-in-llms-follows-reasoning-approach-more-than-topic", "markdown": "https://wpnews.pro/news/mathematical-transfer-in-llms-follows-reasoning-approach-more-than-topic.md", "text": "https://wpnews.pro/news/mathematical-transfer-in-llms-follows-reasoning-approach-more-than-topic.txt", "jsonld": "https://wpnews.pro/news/mathematical-transfer-in-llms-follows-reasoning-approach-more-than-topic.jsonld"}}