NYU Preprint: Unsupervised Discovery of 6 Model Styles That Affect Math Reasoning Accuracy
kchonyc · x · 2026-10-04
A preprint by Ioana Marinescu, Eric Oermann and Kyunghyun Cho shows language models learn style and content jointly — and that style can be both discovered and controlled.
- The team built an algorithm that disentangles content from style representations in model outputs, discovering recurring styles without supervision.
- Analyzing 100K+ verified reasoning traces from nine teacher models, they found six recurring but imbalanced styles.
- Fine-tuning smaller student models with importance weighting to follow these styles beats standard fine-tuning on Pass@k across six math reasoning benchmarks.
- Crucially, style affects correctness: the probability of solving a problem depends on the conditioned style, and different problems benefit from different styles.
The takeaway: stylistic variation in model-generated data is an exploitable source of both control and improved reasoning performance.
More from Research
- LEV: Looped decision model lifts 2.6B accuracy 63.2% to 76.2% with adaptive compute — jiqizhixin · 2026-11-19
- Aurora2-400M-A20M: a sparse MoE pretrained from scratch in 36 hours on one RTX Pro 6000 — Tall_Abrocoma_3533 · 2026-11-19
- Researcher calls METR/Anthropic '2x speedup' RSI analysis statistically dubious — jessi_cata · 2026-11-19
- Linear Probes Could Reveal Which Persona a Base Model Is Playing, Researchers Suggest — cephaloform · 2026-11-19
- ETH's Anymal robot completes 10km autonomous alpine hike with VLM-guided ROAM system — ChongZzZhang · 2026-11-19
- OpenVax codebase grows from 78K to 565K lines in 15 months — iskander · 2026-11-19