LIMO paper shows 800 high-quality reasoning examples beats 100,000 — "less is more" thesis for reasoning training reproduced at scale
The LIMO paper ("Less is More for Reasoning") has been reproduced at the frontier-model scale: 800 carefully curated high-quality reasoning examples produce better downstream reasoning performance than 100,000+ examples of mixed-quality data. The result reframes the data-quality vs data-quantity debate for reasoning specifically — opposite to the conventional pretraining intuition that more is better.
The empirical finding is the headline. LIMO showed in early 2025 that 800 reasoning examples filtered for clear logical structure, correct answers, and explicit step-by-step reasoning produced models that outperformed baselines trained on hundreds of thousands of mixed-quality reasoning examples. The reproduction at frontier-model scale (Gemini 3.1 Pro and Claude Opus 4.7 ablations published in arXiv preprints this cycle) confirms the finding holds at the largest model sizes — it's not a quirk of small-model training dynamics.
The structural implication for the field: data-quality curation is now a higher-leverage activity than data-quantity scaling for reasoning capabilities specifically. Pretraining still benefits from raw scale (every frontier model is trained on multi-trillion-token corpora). But the reasoning fine-tune layer is bounded by example quality, not example count. That changes the cost structure of reasoning-model development: a small curation team producing 800-2000 gold-standard reasoning examples is more valuable than a large data-collection operation producing 500,000 mixed examples. The labs that figure out the curation methodology first own a real competitive advantage. Whether this also holds for non-reasoning capability surfaces (multimodal alignment, agent tool-use, code-completion) is the next research question.
arXiv — LIMO Less is More for Reasoning → · Phys.org — AI breakthrough in math problem → · Tech Jacks Solutions — AI Math Results Four Reasoning Breakthroughs 30 Days →