What's Behind PPO's Collapse in Long-CoT? Value Optimization Holds the Secret
Reviewed by Pithpith:N4CGYDDUopen to challenge →
read the original abstract
Reinforcement learning (RL) is pivotal for enabling large language models (LLMs) to generate long chains of thought (CoT) for complex tasks like math and reasoning. However, Proximal Policy Optimization (PPO), effective in many RL scenarios, fails in long CoT tasks. This paper identifies that value initialization bias and reward signal decay are the root causes of PPO's failure. We propose Value-Calibrated PPO (VC-PPO) to address these issues. In VC-PPO, the value model is pretrained to tackle initialization bias, and the Generalized Advantage Estimation (GAE) computation is decoupled between the actor and critic to mitigate reward signal decay. Experiments on the American Invitational Mathematics Examination (AIME) show that VC-PPO significantly boosts PPO performance. Ablation studies show that techniques in VC-PPO are essential in enhancing PPO for long CoT tasks.
This paper has not been read by Pith yet.
Forward citations
Cited by 27 Pith papers
-
Lightning OPD: Efficient Post-Training for Large Reasoning Models with Offline On-Policy Distillation
Lightning OPD enforces teacher consistency by precomputing log-probabilities over SFT rollouts, matching standard OPD performance with bounded gradient discrepancy and achieving 4x speedup on math and code reasoning tasks.
-
Bringing Value Models Back: Generative Critics for Value Modeling in LLM Reinforcement Learning
GenAC introduces generative critics with chain-of-thought reasoning and in-context conditioning to improve value approximation and downstream RL performance in LLMs compared to value-based and value-free baselines.
-
Small Generalizable Prompt Predictive Models Can Steer Efficient RL Post-Training of Large Reasoning Models
GPS trains a small model on optimization history to predict prompt difficulty and select intermediate-difficulty diverse batches, yielding better training efficiency, final performance, and test-time allocation than b...
-
MURPHY: Feedback-Aware GRPO with Retrospective Credit Assignment for Multi-Turn Code Generation
MURPHY improves code generation pass rates by up to 6% through retrospective credit assignment on multi-turn feedback trees using max or mean reward propagation.
-
The Art of Scaling Reinforcement Learning Compute for LLMs
A 400k+ GPU-hour study shows RL scaling in LLMs follows predictable sigmoidal trajectories, with most design choices affecting efficiency rather than the performance asymptote, enabling accurate large-scale prediction...
-
Blending Supervised and Reinforcement Fine-Tuning with Prefix Sampling
Prefix-RFT blends SFT and RFT via prefix sampling from demonstrations to outperform standalone SFT, RFT, and mixed-policy baselines on math reasoning problems.
-
Reinforcement Learning for Reasoning in Large Language Models with One Training Example
One training example via RLVR boosts LLM math reasoning from 17.6% to 35.7% average across six benchmarks.
-
Small RL Controller, Large Language Model: RL-Guided Adaptive Sampling for Test-Time Scaling
RL-trained lightweight controller using answer statistics improves trade-offs among correctness, latency, and total samples in adaptive sampling for LLM test-time scaling.
-
Extreme Region Policy Distillation
ERPD decouples aggressive off-policy optimization on fixed trajectories from trust-region distillation to achieve comparable or better LLM performance with substantially smaller KL divergence.
-
Holder Policy Optimisation
HölderPO unifies token aggregation in GRPO via the Hölder mean with dynamic p annealing, reporting 54.9% average math-benchmark accuracy and 93.8% ALFWorld success.
-
Holder Policy Optimisation
HölderPO unifies token-level aggregation in GRPO via the Hölder mean with a tunable p parameter and annealing schedule, delivering 54.9% average accuracy on math benchmarks and 93.8% success on ALFWorld.
-
Forge: Quality-Aware Reinforcement Learning for NP-Hard Optimization in LLMs
OPT-BENCH trains LLMs on NP-hard optimization via quality-aware RLVR, achieving 93.1% success rate and 46.6% quality ratio on Qwen2.5-7B while outperforming GPT-4o and transferring gains to other domains.
-
Experience Sharing in Mutual Reinforcement Learning for Heterogeneous Language Models
Mutual Reinforcement Learning allows heterogeneous LLMs to exchange experience through mechanisms like Peer Rollout Pooling, Cross-Policy GRPO Advantage Sharing, and Success-Gated Transfer, with outcome-level sharing ...
-
Segment-Aligned Policy Optimization for Multi-Modal Reasoning
SAPO introduces segment-level policy optimization using a step-wise MDP abstraction to better align RL updates with reasoning structure in multi-modal LLM tasks.
-
Unleashing Implicit Rewards: Prefix-Value Learning for Distribution-Level Optimization
IPVRM learns prefix values to produce reliable step rewards from sequence outcomes using TD learning, enabling distribution-level RL that improves reasoning when paired with calibrated rewards.
-
Lightning OPD: Efficient Post-Training for Large Reasoning Models with Offline On-Policy Distillation
Lightning OPD is an offline on-policy distillation method that matches standard OPD performance at 4x efficiency by enforcing teacher consistency between SFT and distillation phases.
-
ReSeek: A Self-Correcting Framework for Search Agents with Instructive Rewards
ReSeek adds self-correction via a JUDGE action and a dense instructive reward (correctness plus utility) to RL training of search agents, yielding higher success and faithfulness on a new contamination-resistant benchmark.
-
VAPO: Efficient and Reliable Reinforcement Learning for Advanced Reasoning Tasks
VAPO achieves 60.4 on AIME 2024 with Qwen 32B, outperforming prior methods by over 10 points through targeted fixes for value bias, sequence length variation, and sparse rewards.
-
DAPO: An Open-Source LLM Reinforcement Learning System at Scale
DAPO introduces decoupled clipping and dynamic sampling for LLM RL, achieving 50 on AIME 2024 with Qwen2.5-32B while fully open-sourcing code, data, and the verl-based training system.
-
CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents
CompactionRL trains LLM agents to generate context summaries during RL rollouts, enabling long-horizon task completion under fixed context budgets with consistent gains on SWE-bench Verified and Terminal-Bench 2.0.
-
Modularized Reinforcement Learning on LLMs: From MDP Creation to Exploration and Learning
Survey mapping RL techniques onto LLM training and highlighting gaps in value-based, off-policy, and bootstrapping methods.
-
VIMPO: Value-Implicit Policy Optimization for LLMs
VIMPO derives a policy-implied value function from optimality conditions for critic-free RL in LLMs and shows gains over GRPO on math benchmarks.
-
RLVR without Ineffective Samples: Group Prioritized Off-Policy Optimization for LLM Reasoning
POPO uses recency-based prioritized group replay and decoupled off-policy optimization to avoid zero-variance ineffective samples in RLVR, accelerating LLM reasoning finetuning with fewer rollouts.
-
Trust Region On-Policy Distillation
TrOPD stabilizes on-policy distillation for LLMs with trust-region learning, outlier estimation, and off-policy guidance, outperforming prior OPD methods on reasoning and code benchmarks.
-
UI-TARS-2 Technical Report: Advancing GUI Agent with Multi-Turn Reinforcement Learning
UI-TARS-2 reaches 88.2 on Online-Mind2Web, 47.5 on OSWorld, 50.6 on WindowsAgentArena, and 73.3 on AndroidWorld while attaining 59.8 mean normalized score on a 15-game suite through multi-turn RL and scalable data generation.
-
Seed1.5-VL Technical Report
Seed1.5-VL is a compact multimodal model that sets new records on dozens of vision-language benchmarks and outperforms prior systems on agent-style tasks.
-
Reinforcement Learning from Human Feedback
The book introduces the origins, mathematical setup, and optimization stages of RLHF including reward modeling, reinforcement learning, rejection sampling, and direct alignment algorithms.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.