AI Skills AI技能 3h ago Updated 1h ago 更新于 1小时前 54

RLVR: From Human Feedback to Verifiable Truth: How Modern LLMs Actually Learn to Reason RLVR:从人类反馈到可验证真理:现代大模型如何真正学会推理

RLVR (Reinforcement Learning with Verifiable Rewards) replaces human preference-based reward models with deterministic verifiers that check correctness, fundamentally shifting training from "sounding right" to "being right" GRPO (Group Relative Policy Optimization) eliminates the need for a separate value model by using intra-group comparison, making RLVR computationally feasible at scale DeepSeek-R1 discovered spontaneous chain-of-thought reasoning and self-correction emergently through RLVR+GR RLVR(可验证奖励强化学习)用确定性验证函数替代人类偏好评分,让模型学会真正正确而非仅"听起来正确" GRPO算法通过组内相对优势估计消除独立价值模型,使RLVR训练在计算上可行且可扩展 DeepSeek-R1仅用约5%总计算量进行后训练,却涌现出自发链式思维和自我纠错能力 o3、Gemini 3、GPT-6 Astra等前沿模型均采用RLVR范式,证明训练信号质量比数据量和参数规模更重要 这一范式转变标志着AI训练从优化人类偏好转向优化客观正确性

72
Hot 热度
78
Quality 质量
82
Impact 影响力

Analysis 深度分析

TL;DR

  • RLVR (Reinforcement Learning with Verifiable Rewards) replaces human preference-based reward models with deterministic verifiers that check correctness, fundamentally shifting training from "sounding right" to "being right"
  • GRPO (Group Relative Policy Optimization) eliminates the need for a separate value model by using intra-group comparison, making RLVR computationally feasible at scale
  • DeepSeek-R1 discovered spontaneous chain-of-thought reasoning and self-correction emergently through RLVR+GRPO, using only ~5% of total compute on post-training
  • The approach triggered an industry-wide shift: OpenAI's o3, Gemini 3 Deep Think, and GPT-6 Astra all adopted RLVR-style training, proving that better training signals matter more than raw scale
  • RLVR is most effective in domains with verifiable outputs—mathematics and code—where binary pass/fail rewards eliminate human rater bias and fluent-but-wrong answers

Why It Matters

RLVR represents a paradigm shift in how frontier models are trained, moving the field away from the fundamental limitation of RLHF: optimizing for human preference rather than objective correctness. For AI practitioners, this means the path to genuinely reasoning models lies not in more data or larger architectures, but in designing better reward signals. The DeepSeek-R1 result—disproportionate intelligence gains from a small compute investment—demonstrates that training methodology can be a more powerful lever than brute-force scaling.

Technical Details

  • RLVR Core Mechanism: Replaces human-rater reward models with deterministic verifiers—symbolic checkers for math (comparing final numerical answers) and test suites for code (running automated tests). Reward is binary: 1 for correct, 0 for incorrect, with no ambiguity or formatting bias.
  • GRPO Algorithm: DeepSeek's Group Relative Policy Optimization estimates advantage by comparing a response against a group of sibling responses to the same prompt, eliminating the need for a separate value/critic network. This reduces memory and engineering complexity while preserving RL-quality training signal.
  • Training Efficiency: DeepSeek-R1's RL phase consumed approximately 147,000 GPU hours versus 2.8 million GPU hours for pretraining—a 5% post-training compute ratio that produced emergent reasoning capabilities the researchers did not explicitly program.
  • Algorithm Progression: PPO requires a parallel value model and is computationally expensive and unstable at scale; DPO removes the RL loop entirely but still optimizes for human preferences rather than correctness; GRPO+RLVR combines RL-quality signal with DPO-level simplicity under verifiable reward conditions.
  • Benchmark Results: OpenAI's o3 and Gemini 3 Deep Think adopted similar approaches, with Gemini 3 Deep Think achieving approximately 96% on ARC-AGI-1, demonstrating that improved training signals—not just scale—drive frontier performance.

Industry Insight

  • The RLVR paradigm suggests that the next major performance gains in frontier models will come from training methodology innovation rather than architectural scale or dataset expansion—teams should prioritize verifier design and reward signal quality.
  • RLVR's effectiveness is currently constrained to domains with verifiable outputs (math, code, logic); extending this approach to open-ended domains like creative writing, strategy, or scientific hypothesis generation remains an unsolved challenge and a potential competitive frontier.
  • The two-week RL training pause observed in GPT-6 Astra's development (mid-August 2026) signals that RLVR training is non-trivial to stabilize at scale—practitioners should anticipate significant iteration and debugging during the post-training phase, and budget accordingly.

TL;DR

  • RLVR(可验证奖励强化学习)用确定性验证函数替代人类偏好评分,让模型学会真正正确而非仅"听起来正确"
  • GRPO算法通过组内相对优势估计消除独立价值模型,使RLVR训练在计算上可行且可扩展
  • DeepSeek-R1仅用约5%总计算量进行后训练,却涌现出自发链式思维和自我纠错能力
  • o3、Gemini 3、GPT-6 Astra等前沿模型均采用RLVR范式,证明训练信号质量比数据量和参数规模更重要
  • 这一范式转变标志着AI训练从优化人类偏好转向优化客观正确性

为什么值得看

这篇文章揭示了2026年前沿AI模型能力跃迁的核心机制——RLVR范式正在重塑模型训练逻辑。对从业者而言,理解RLVR与GRPO的结合是把握当前AI竞赛本质的关键,也为后训练阶段的资源分配和算法选择提供了明确方向。

技术解析

  • RLVR核心机制:用确定性验证函数(数学答案匹配、代码测试套件)替代人类偏好评分,奖励信号为二元(正确=1,错误=0),彻底消除"流畅但错误"答案获得高奖励的可能性。
  • GRPO算法创新:DeepSeek提出的Group Relative Policy Optimization,对同一提示生成多个候选响应,通过组内相对优势估计替代独立价值模型,实现RL级训练信号与DPO级计算效率的结合。
  • DeepSeek-R1训练效率:后训练阶段仅用约147,000 GPU小时(占总计算量约5%),却涌现出自发链式思维、自我纠错、中途回溯等未显式训练的行为。
  • 前沿模型进展:OpenAI o3结合RLVR与测试时计算(test-time compute),Gemini 3 Deep Think在ARC-AGI-1榜单达96%,GPT-6 Astra(内部代号"Bel")于2026年9月3日启动。

行业启示

  • 训练范式转移:RLVR证明"训练信号质量"比"数据量/参数规模"更能驱动能力跃迁,行业应从追求更大模型转向优化后训练阶段的奖励机制设计。
  • 资源分配策略:DeepSeek-R1仅用5%计算量实现智能突破,提示实验室应重新评估后训练阶段的计算预算,GRPO+RLVR组合具有极高性价比。
  • 能力涌现新认知:自我纠错、链式思维等"类推理"行为可从正确性奖励中自发涌现,无需显式编程,这为设计下一代推理模型提供了新路径。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Training 训练 Research 科学研究 Alignment 对齐 Evaluation 评测