Yuxuan Tong

Jiazheng Xu, Xiao Liu, Yuchen Wu et al. · tsinghua

We present a comprehensive solution to learn and improve text-to-image models from human preference feedback. To begin with, we build ImageReward -- the first general-purpose text-to-image human preference reward model -- to effectively encode human preferences. Its training is based on our systematic annotation pipeline including rating and ranking, which collects 137k expert comparisons to date. In human evaluation, ImageReward outperforms existing scoring models and metrics, making it a promising automatic metric for evaluating text-to-image synthesis. On top of it, we propose Reward Feedback Learning (ReFL), a direct tuning algorithm to optimize diffusion models against a scorer. Both automatic and human evaluation support ReFL's advantages over compared methods. All code and datasets are provided at \url{https://github.com/THUDM/ImageReward}.

Qiying Yu, Zheng Zhang, Ruofei Zhu et al. · tsinghua

Inference scaling empowers LLMs with unprecedented reasoning ability, with reinforcement learning as the core technique to elicit complex reasoning. However, key technical details of state-of-the-art reasoning LLMs are concealed (such as in OpenAI o1 blog and DeepSeek R1 technical report), thus the community still struggles to reproduce their RL training results. We propose the $\textbf{D}$ecoupled Clip and $\textbf{D}$ynamic s$\textbf{A}$mpling $\textbf{P}$olicy $\textbf{O}$ptimization ($\textbf{DAPO}$) algorithm, and fully open-source a state-of-the-art large-scale RL system that achieves 50 points on AIME 2024 using Qwen2.5-32B base model. Unlike previous works that withhold training details, we introduce four key techniques of our algorithm that make large-scale LLM RL a success. In addition, we open-source our training code, which is built on the verl framework, along with a carefully curated and processed dataset. These components of our open-source system enhance reproducibility and support future research in large-scale LLM RL.

Edward Yeo, Yuxuan Tong, Morry Niu et al.

Scaling inference compute enhances reasoning in large language models (LLMs), with long chains-of-thought (CoTs) enabling strategies like backtracking and error correction. Reinforcement learning (RL) has emerged as a crucial method for developing these capabilities, yet the conditions under which long CoTs emerge remain unclear, and RL training requires careful design choices. In this study, we systematically investigate the mechanics of long CoT reasoning, identifying the key factors that enable models to generate long CoT trajectories. Through extensive supervised fine-tuning (SFT) and RL experiments, we present four main findings: (1) While SFT is not strictly necessary, it simplifies training and improves efficiency; (2) Reasoning capabilities tend to emerge with increased training compute, but their development is not guaranteed, making reward shaping crucial for stabilizing CoT length growth; (3) Scaling verifiable reward signals is critical for RL. We find that leveraging noisy, web-extracted solutions with filtering mechanisms shows strong potential, particularly for out-of-distribution (OOD) tasks such as STEM reasoning; and (4) Core abilities like error correction are inherently present in base models, but incentivizing these skills effectively for complex tasks via RL demands significant compute, and measuring their emergence requires a nuanced approach. These insights provide practical guidance for optimizing training strategies to enhance long CoT reasoning in LLMs. Our code is available at: https://github.com/eddycmu/demystify-long-cot.

Yingru Li, Jiawei Xu, Ziniu Li et al.

Reinforcement Learning (RL) for Large Language Models (LLMs) often suffers from training collapse in long-horizon tasks due to exploding gradient variance. To mitigate this, a baseline is commonly introduced for advantage computation; however, traditional value models remain difficult to optimize, and standard group-based baselines overlook sequence heterogeneity. Although classic optimal baseline theory can achieve global variance reduction, it neglects token heterogeneity and requires prohibitive gradient-based computation. In this work, we derive the Optimal Token Baseline (OTB) from first principles, proving that gradient updates should be weighted inversely to their cumulative gradient norm. To ensure efficiency, we propose the Logit-Gradient Proxy that approximates the gradient norm using only forward-pass probabilities. Our method achieves training stability and matches the performance of large group sizes ($N=32$) with only $N=4$, reducing token consumption by over 65% across single-turn and tool-integrated reasoning tasks.

Chaobo Jia, Ruipeng Wan, Ting Sun et al.

LLM-based game generation promises to turn natural-language specifications into executable games, but progress is limited by the lack of reliable automated verification. Unlike conventional code generation, game correctness is defined over long-horizon interaction: a game may appear correct while violating core mechanics such as state updates, interaction rules, and phase transitions. Existing Agent-as-a-Verifier approaches collapse verification into open-ended gameplay, making verdicts reachability-bound, time-consuming, coverage-limited, and sensitive to the agent's gameplay ability. We present GameGen-Verifier, an automated verification paradigm for LLM-generated games that decomposes a specification into verifiable keypoints and grounds them into independent verification units. Each unit patches the game runtime into a concrete target state, executes a bounded interaction, and judges the outcome against the keypoint assertion. We implement GGV-Harness, a scalable agentic harness providing concurrency management, runtime isolation, and fault recovery. On VeriGame, our dataset of 100 games across seven genres, GameGen-Verifier achieves up to 92.2% accuracy against human judgments versus 58.8% for the coverage-enforced Agent-as-a-Verifier baseline, while reducing wall-clock time by up to 16.6x.

Yingru Li, Jiacai Liu, Jiawei Xu et al.

Policy gradient methods for Large Language Models optimize a policy $π_θ$ via a surrogate objective computed from samples of a rollout policy $π_{\text{roll}}$. However, modern LLM-RL pipelines suffer from unavoidable implementation divergences -- backend discrepancies, Mixture-of-Experts routing discontinuities, and distributed training staleness -- causing off-policy mismatch ($π_{\text{roll}} \neq π_θ$) and approximation errors between the surrogate and the true objective. We demonstrate that classical trust region bounds on this error scale as $O(T^2)$ with sequence length $T$, rendering them vacuous for long-horizon tasks. To address this, we derive a family of bounds -- both KL-based and TV-based -- including a Pinsker-Marginal bound ($O(T^{3/2})$), a Mixed bound ($O(T)$), and an Adaptive bound that strictly generalizes the Pinsker-Marginal bound via per-position importance-ratio decomposition. Taking the minimum over all bounds yields the tightest known guarantee across all divergence regimes. Crucially, all bounds depend on the maximum token-level divergence $D_{\mathrm{KL}}^{\mathrm{tok,max}}$ (or $D_{\mathrm{TV}}^{\mathrm{tok,max}}$), a sequence-level quantity that cannot be controlled by token-independent methods like PPO clipping. We propose Trust Region Masking (TRM), which masks entire sequences violating the trust region, enabling the first non-vacuous monotonic improvement guarantees for long-horizon LLM-RL.

Yingru Li, Jiawei Xu, Jiacai Liu et al.

Reinforcement Learning (RL) for Large Language Models (LLMs) faces a fundamental tension: the numerical divergence between high-throughput inference engines and numerically precise training engines. Although these systems share the same parameters, they produce slightly different probability distributions, creating a training-inference mismatch. We prove that the bound on the log-probability divergence arising from this mismatch scales as $(1-p)$, where $p$ is the token probability. This scaling induces a highly asymmetric effect: the bound vanishes for high-probability tokens but remains significant for low-probability tokens in the distribution tail. When sampled, these tail tokens introduce systematically biased errors that accumulate over sequences, thereby destabilizing gradient estimation. Instead of applying post-hoc corrections, we propose Dynamic Vocabulary Pruning (DVP), which constrains the RL objective to a dynamically determined ''safe'' vocabulary that excludes the extreme tail. This strategy trades large, destabilizing numerical errors for a small, bounded optimization bias. We validate DVP empirically by demonstrating stable training, and theoretically by deriving strict bounds on the induced bias.

Guangming Sheng, Yuxuan Tong, Borui Wan et al.

Reinforcement learning (RL) post-training for Large Language Models (LLMs) is now scaling to large clusters and running for extended durations to enhance model reasoning performance. However, the scalability of existing RL frameworks is limited, as extreme long-tail skewness in RL trajectory generation causes severe GPU underutilization. Current asynchronous RL systems attempt to mitigate this, but they rely on global weight synchronization between the actor and all rollouts, which creates a rigid model update schedule. This global synchronization is ill-suited for the highly skewed and evolving distribution of trajectory generation latency in RL training, crippling training efficiency. Our key insight is that efficient scaling requires breaking this lockstep through trajectory-level asynchrony, which generates and consumes each trajectory independently. We propose Laminar, a scalable and robust RL post-training system built on a fully decoupled architecture. First, we replace global updates with a tier of relay workers acting as a distributed parameter service. This enables asynchronous and fine-grained weight synchronization, allowing rollouts to pull the latest weight anytime without stalling the actor's training loop. Second, a dynamic repack mechanism consolidates long-tail trajectories onto a few dedicated rollouts, maximizing generation throughput. The fully decoupled design also isolates failures, ensuring robustness for long-running jobs. Our evaluation on a 1024-GPU cluster shows that Laminar achieves up to 5.48$\times$ training throughput speedup over state-of-the-art systems, while reducing model convergence time.

Yuxuan Tong, Xiwen Zhang, Rui Wang et al.

Solving mathematical problems requires advanced reasoning abilities and presents notable challenges for large language models. Previous works usually synthesize data from proprietary models to augment existing datasets, followed by instruction tuning to achieve top-tier results. However, our analysis of these datasets reveals severe biases towards easy queries, with frequent failures to generate any correct response for the most challenging queries. Hypothesizing that difficult queries are crucial to learn complex reasoning, we propose Difficulty-Aware Rejection Tuning (DART), a method that allocates difficult queries more trials during the synthesis phase, enabling more extensive training on difficult samples. Utilizing DART, we have created new datasets for mathematical problem-solving that focus more on difficult queries and are substantially smaller than previous ones. Remarkably, our synthesis process solely relies on a 7B-sized open-weight model, without reliance on the commonly used proprietary GPT-4. We fine-tune various base models on our datasets ranging from 7B to 70B in size, resulting in a series of strong models called DART-MATH. In comprehensive in-domain and out-of-domain evaluation on 6 mathematical benchmarks, DART-MATH outperforms vanilla rejection tuning significantly, being superior or comparable to previous arts, despite using much smaller datasets and no proprietary models. Furthermore, our results position our synthetic datasets as the most effective and cost-efficient publicly available resources for advancing mathematical problem-solving.