New research from Microsoft and academic collaborators has overturned that assumption. Using Reinforcement Learning with Verifiable Rewards (RLVR)<\/strong> and just a single training example<\/strong>, researchers achieved results on par with models trained on over a thousand examples<\/em>, sometimes even better.<\/p>\n This improvement isn\u2019t just incremental progress. It\u2019s a rethinking of how we fine-tune large language models (LLMs) for reasoning tasks. In this post, we\u2019ll unpack what 1-shot RLVR is, how it works, and what it means for developers building math agents, automated tutors, and reasoning copilots.<\/p>\n \n<\/blockquote>\n RLVR is a flavor of reinforcement learning where the model is trained using verifiable reward signals,<\/strong> typically 0\/1 based on whether the output is correct. In contrast to reward models used in Rlhf<\/a>, RLVR uses hard ground truth.<\/p>\n What the authors discovered is that if you apply RLVR to a base model (e.g., Qwen2.5-Math-1.5B) and train it on just one carefully selected math example<\/strong>, performance on benchmark tasks can nearly double<\/strong>.<\/p>\n Here\u2019s what happens when you train Qwen2.5-Math-1.5B on just one<\/em> example:<\/p>\n Even using two examples<\/strong> yielded 74.8%<\/strong> on MATH500 and 36.6%<\/strong> average, slightly outperforming the full 1.2k dataset<\/strong> the example was selected from.<\/p>\n This result wasn\u2019t limited to a fluke. Many different examples produced ~30% or more gains when used individually.<\/p>\n The paper introduces several hypotheses and findings:<\/p>\n If you\u2019re building LLM-powered reasoning tools, math solvers, science tutors, or data agents, this technique offers enormous leverage:<\/p>\n Imagine building an AI tutor that learns from a single problem and generalizes across the curriculum. That future just got closer.<\/p>\n The authors evaluated on the ARC-Challenge and ARC-Easy, non-mathematical reasoning benchmarks.\u00a0<\/p>\n Here\u2019s what they found for Qwen2.5-Math-1.5B:<\/p>\n That\u2019s a gain over even full-dataset RLVR. Training on a math problem helped the model become a better commonsense reasoner.<\/p>\n Using historical training variance to select high-impact examples (\u03c01 and \u03c013) worked well. But surprisingly, many examples work<\/strong>, even those with low variance.<\/p>\n There\u2019s no perfect recipe yet, but the early insight is promising:<\/p>\n \u201cAlmost all examples improve performance when used in 1-shot RLVR.\u201d<\/p>\n For some models, particularly distilled ones like DeepSeek-R1-Distill-Qwen-1.5B, performance gains from 1-shot RLVR were more modest (~6.9%). But moving to 4-shot or 16-shot setups showed steady improvement.<\/p>\n This implies that model family and training history matter,<\/strong> but the general trend holds: you need far less data than we thought<\/em>.<\/p>\n One of the paper\u2019s most surprising discoveries is that entropy loss alone<\/strong>, even without rewards, can yield large gains.<\/p>\n Example: Training Qwen2.5-Math-1.5B with only entropy loss improves MATH500 from 36.0% to 63.4%<\/strong> in 20 steps.<\/p>\n This reveals a powerful principle:<\/p>\n Letting models explore more freely helps them generalize even from one example.<\/p>\n Post-saturation generalization may remind some of grokking, where models suddenly generalize after long periods of overfitting.<\/p>\n But ablation studies show 1-shot RLVR isn\u2019t the same:<\/p>\n This paper serves as a timely reminder. More data isn\u2019t always the answer. Better data, better selection, and reinforcement learning, even from one example, can unlock powerful capabilities in your base models.<\/p>\n For developers<\/strong>, this means<\/p>\n While the results of 1-shot RLVR are impressive in research, applying them at scale requires the right tools and infrastructure. That\u2019s where Adaptive Engine<\/strong> comes in.<\/p>\n Whether you\u2019re fine-tuning models on a single math problem or optimizing agents across business domains, Adaptive gives you the full flywheel:<\/p>\n Outperform frontier models with reinforcement fine-tuning<\/strong> that works, even with limited data. Adaptive makes it easy to run GRPO or PPO on open models with just a few examples and verifiable rewards.<\/p>\n Before you deploy, you need confidence. Adaptive supports personalized, production-aligned evaluations<\/strong>, so you can benchmark improvements on your real-world workloads, not just abstract benchmarks.<\/p>\n With fast, efficient inference<\/strong>, Adaptive lets you host tuned models wherever you need them, on cloud, edge, or hybrid infrastructure. High performance, low latency.<\/p>\n From day-one experimentation to at-scale deployment, Adaptive helps you:<\/p>\n \n The post Reinforcement Learning from One Example?<\/a> appeared first on Towards Data Science<\/a>.<\/p>\n<\/div>\n![\"\"](\"https:\/\/i0.wp.com\/contributor.insightmediagroup.io\/wp-content\/uploads\/2025\/04\/unnamed-71-1024x322.png?resize=1024%2C322&ssl=1\")
\n
1-Shot RLVR: What Is It?<\/h2>\n
The Numbers That Stun<\/h2>\n
\n
Why Does This Approach Work?<\/h2>\n
\n
Implications for Developers<\/h2>\n
\n
Beyond Math: Early Signs of Transfer<\/h2>\n
\n
What Makes a Good Example?<\/h2>\n
When One Isn\u2019t Enough<\/h2>\n
The Role of Entropy: Why Exploration Matters<\/h2>\n
1-Shot \u2260 Grokking<\/h2>\n
\n
The Future: Smarter Data, Smaller Footprints<\/h2>\n
\n
How Adaptive Helps You Go from Prototype to Production<\/h2>\n
Adapt<\/h3>\n
Evaluate<\/h3>\n
\n\u00a0<\/em><\/strong>Serve<\/h3>\n
\n