Research Papers 论文研究 4h ago Updated 2h ago 更新于 2小时前 47

Verifier-Induced Support Reshaping in On-Policy Optimization 验证器诱导的支持重塑在策略优化中

On-policy RL with verifiable rewards (RLVR) can optimize the current objective while simultaneously making successful behaviors for other objectives too rare to sample, a phenomenon termed "verifier-induced support reshaping" Math-RLVR improves instruction-following pass@1 by 6.5pp on IFEval with Qwen3-8B-Base but reduces best@32 by 9.8pp, demonstrating a critical divergence between single-sample and multi-sample performance IF-RLVR shifts mathematical reasoning from step-by-step openings toward On-policy强化学习(RLVR)在优化当前目标时,会导致后续目标的"有效可奖励支持"萎缩,成功行为变得过于罕见难以采样 Math-RLVR提升指令遵循平均成功率,但使成功响应的prompt数量下降;IFEval上pass@1提升6.5pp,best@32下降9.8pp IF-RLVR将数学响应从逐步推理转向直接回答,降低后续Math-RLVR的reward variation 变化集中在响应前几个token,RLVR本质是对基础策略已有opening的重排序,而非创造新能力 参考策略约束、路由先验和on-policy蒸馏仅能部分保留跨任务支持,端点改进不保证未来可训练性或联合能力

62
Hot 热度
75
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • On-policy RL with verifiable rewards (RLVR) can optimize the current objective while simultaneously making successful behaviors for other objectives too rare to sample, a phenomenon termed "verifier-induced support reshaping"
  • Math-RLVR improves instruction-following pass@1 by 6.5pp on IFEval with Qwen3-8B-Base but reduces best@32 by 9.8pp, demonstrating a critical divergence between single-sample and multi-sample performance
  • IF-RLVR shifts mathematical reasoning from step-by-step openings toward direct answers, reducing reward variation and lowering best@k across sampling budgets
  • Token-distribution and intervention analyses reveal these effects concentrate in the first few response tokens, with RLVR primarily reranking existing base-policy openings rather than generating new ones
  • Reference-policy constraints, routing priors, and on-policy distillation only partially preserve cross-task support, proving that endpoint improvements do not guarantee future trainability or joint capability

Why It Matters

This research reveals a fundamental tension in on-policy RLVR training: optimizing for one capability can actively degrade the model's ability to learn other capabilities by shrinking the effective rewardable support. For practitioners running multi-objective or sequential RLVR pipelines, this means endpoint gains on one benchmark may mask deteriorating trainability on others, requiring careful monitoring of best@k and support diversity rather than relying solely on pass@1 improvements.

Technical Details

  • Verifier-induced support reshaping: Defined as the phenomenon where on-policy RLVR improves the current objective while making successful trajectories for later objectives too rare to sample within a fixed rollout budget; effective rewardable support is quantified as successful trajectories reachable within that budget
  • Bidirectional training experiments: Conducted across two model families on mathematical reasoning and constrained instruction following, including sequential training where one verifier is applied after the other (Math-RLVR then IF-RLVR and vice versa)
  • Key empirical findings: On IFEval with Qwen3-8B-Base, Math-RLVR yields pass@1 +6.5pp but best@32 -9.8pp; IF-RLVR shifts math responses from step-by-step to direct answers and reduces reward variation for subsequent Math-RLVR
  • Causal mechanism: Token-distribution analyses and controlled opening interventions show changes concentrate in the first few response tokens; RLVR mainly reranks openings already present in the base policy, and the selected opening causally affects math searchability
  • Mitigation attempts: Reference-policy constraints, routing priors, and on-policy distillation were tested but only partially preserved cross-task support; composite benchmarks MathIF and ReasonIF showed marginal gains translate only partly into responses that are both correct and constraint-following

Industry Insight

  • Multi-objective RLVR pipelines should monitor best@k and support diversity metrics alongside pass@1, as single-sample improvements can mask shrinking explorability for downstream objectives
  • Sequential training strategies that optimize one capability before another risk irreversible support collapse; joint or interleaved optimization with support-preserving regularizers may be necessary for robust multi-task RLVR
  • The finding that RLVR primarily reranks existing openings rather than creating new ones suggests that base model diversity and initial rollout budgets are critical bottlenecks—investing in broader pre-training coverage and larger sampling budgets may yield better returns than aggressive on-policy optimization alone

TL;DR

  • On-policy强化学习(RLVR)在优化当前目标时,会导致后续目标的"有效可奖励支持"萎缩,成功行为变得过于罕见难以采样
  • Math-RLVR提升指令遵循平均成功率,但使成功响应的prompt数量下降;IFEval上pass@1提升6.5pp,best@32下降9.8pp
  • IF-RLVR将数学响应从逐步推理转向直接回答,降低后续Math-RLVR的reward variation
  • 变化集中在响应前几个token,RLVR本质是对基础策略已有opening的重排序,而非创造新能力
  • 参考策略约束、路由先验和on-policy蒸馏仅能部分保留跨任务支持,端点改进不保证未来可训练性或联合能力

为什么值得看

该研究揭示了RLVR训练中一个关键但被忽视的"支持坍塌"现象,对多任务联合训练和持续学习具有重要警示意义。研究结果提醒从业者:单点指标优化可能以牺牲模型长期可训练性和跨任务泛化为代价。

技术解析

  • 有效可奖励支持定义:在固定rollout预算内可达的成功轨迹集合,用于量化模型的实际可训练空间
  • 双向训练实验:在数学推理和约束指令遵循两个任务族上进行交替RLVR训练,观察跨任务支持变化
  • 关键发现:Math-RLVR使IF任务pass@1提升6.5pp但best@32下降9.8pp;IF-RLVR使数学响应从step-by-step转向直接回答
  • Token级分析:通过token分布分析和controlled opening干预,证明变化集中在前几个响应token
  • 缓解方法测试:参考策略约束、routing priors、on-policy distillation仅能部分保留跨任务支持;MathIF和ReasonIF联合任务显示边际增益无法完全转化为正确且遵循约束的响应

行业启示

  • 多任务RLVR训练需警惕"支持坍塌"风险,单目标优化可能以牺牲跨任务可训练性为代价,建议引入支持保留机制
  • 评估指标应兼顾pass@k和best@k,单一pass@1提升可能掩盖模型探索能力的退化
  • 联合能力训练需关注opening分布变化,前几个token的因果影响决定了后续搜索空间,应在训练设计中显式建模跨任务支持保留

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。

Research 科学研究 Training 训练 Evaluation 评测