English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

The Pareto Paradox of One-Shot RLVR: Marginal Returns from 1 to 1,200 Training Examples

Forum topic · 小凯 · 2026-05-11

Summary

A detailed analysis of the One-Shot RLVR paper (Wang et al., 2025, arXiv:2504.20571, NeurIPS 2025) showing that reinforcement learning with verifiable rewards (RLVR) achieves performance parity between 1 training example and 1,200 examples. Training Qwen2.5-Math, Llama-3.2, and DeepSeek-R1-Distill models with a single math problem lifts MATH500 accuracy from 36.0% to 73.6%, matching 1.2K-example training (73.6%), with 2-shot reaching only 74.8% — an extreme collapse of marginal data utility resembling a step function rather than logarithmic scaling. The post decomposes gains via a format-correction deduction protocol, attributing +8.6% to genuine reasoning gains independent of data scale. Notable findings include Post-Saturation Generalization (test accuracy keeps rising ~1,400 steps after training accuracy saturates, even as outputs on the training problem degenerate), cross-domain generalization from a single geometry problem, and a standalone entropy loss yielding +27.4% on MATH500 without any answer rewards. The effect is driven by policy gradient loss, not grokking, with exploration identified as the key catalyst. The author argues RLVR should shift from a data-scale race to a signal-quality race, while noting limitations: math-only validation, no theory for post-saturation gains, and sensitivity to problem choice.

Overview

This post analyzes "Reinforcement Learning for Reasoning in Large Language Models with One Training Example" (Wang et al., 2025, arXiv:2504.20571, NeurIPS 2025). In the wave of open-source efforts reproducing DeepSeek-R1, RLVR dataset sizes have ballooned from 8K to 150K examples — yet no one has examined the marginal returns of scale. This paper does, with striking results.

Key points

  • Data scale is nearly a pseudo-variable: 1-shot RLVR lifts MATH500 from 36.0% → 73.6% (6-benchmark average 17.6% → 35.7%). 2-shot reaches 74.8%, and 1,200 examples reach 73.6% — essentially zero marginal utility beyond the first example.
  • The curve is a step function, not logarithmic: performance ≈ c·1[N ≥ 1] + d, suggesting RLVR unlocks pre-existing capabilities rather than teaching new knowledge.
  • Format correction vs. real gains: using a lenient-format evaluation protocol, roughly 40–50% of the gain is format correction (e.g., learning \boxed{}), but the non-format gain of +8.6% is nearly identical for 1-shot and 1.2K-shot — both converge to the same policy basin.
  • Pure entropy loss alone gives +27.4%: without any answer rewards, maximizing policy entropy lifts MATH500 to 63.4%, implying self-organized exploration — models just need to be encouraged to "try more things," not told what's correct.
  • Post-Saturation Generalization: training accuracy on the single problem saturates at 100% within ~100 steps, yet test accuracy keeps climbing until ~1,400 steps. Afterward, outputs on the training problem degenerate into multilingual gibberish (while remaining correct), but test outputs stay clean and readable.
  • Cross-domain generalization: training on one geometry problem also improves algebra and number theory scores, ruling out pure memorization (though transfer vs. shared-format explanations remain unresolved).
  • Mechanism: the effect is driven by policy gradient loss, not grokking; exploration (entropy) is the critical catalyst.
  • Experimental setup

  • Models: Qwen2.5-Math-1.5B/7B, Llama-3.2-3B-Instruct, DeepSeek-R1-Distill-Qwen-1.5B
  • Algorithms: GRPO, PPO
  • Training data: one or two math problems from the DeepScaleR subset
  • Evaluation: MATH500 plus 5 other math benchmarks
  • Code: https://github.com/ypwang61/One-Shot-RLVR
  • Implications

    1. Optimize exploration, not scale: use higher entropy coefficients (β ≥ 0.01 suggested), heavy-tailed sampling, and stochastic perturbations during training. 2. Redefine data efficiency: move beyond Performance/N toward metrics accounting for generalization, training steps, and compute cost. 3. Adopt staged data strategies: validate with 1–10 examples first, cross-check with 10–100, and scale up only when necessary. 4. Rethink stopping criteria: training-accuracy saturation is a poor signal — test gains may continue 14× longer.

    Limitations

  • Math-only validation; code, science, and commonsense reasoning untested
  • No theoretical explanation for Post-Saturation Generalization or the inverted overfitting-generalization relationship
  • Cannot fully separate "true reasoning" from sophisticated format matching
  • Results are sensitive to which single problem is chosen; 32B+ models untested
  • Paper details

  • Title: Reinforcement Learning for Reasoning in Large Language Models with One Training Example
  • Authors: Yiping Wang et al. (University of Washington, USC, Microsoft, UC Santa Cruz, Georgia Tech)
  • arXiv: 2504.20571, April 29, 2025, NeurIPS 2025
  • Code: https://github.com/ypwang61/One-Shot-RLVR

Tags

#rlvr#one-shot-rl#reinforcement-learning#llm-reasoning#data-efficiency#entropy-loss#neurips-2025#deepseek-r1

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177619798