Jiajun Fan

Ye Li, Huanan Liu, Kangye Ji et al.

Vision-Language-Action (VLA) models are a powerful paradigm for generalist robotic control. However, their high computational cost and limited control frequency hinder real-time robotic manipulation, especially when large vision-language backbones and iterative action heads run at every control step. Existing VLA acceleration methods often optimize individual components or rely on fixed acceleration rules, treating different control steps with largely fixed computation and overlooking the non-uniform reasoning demands of sequential embodied control. Inspired by human motor control, where cognitive and feedback resources concentrate on goal-sensitive stages, we argue that VLA models should learn when to invest full computation and when to reuse prior computation. We propose ElegantVLA, a plug-in phase-adaptive inference framework that accelerates VLA models through intra-model dynamic compute scheduling. ElegantVLA introduces a lightweight scheduler that observes temporal representation similarity, robot-motion cues, and episode progress to jointly allocate computation across the vision encoder, LLM, and action head. For perception-language reasoning, the scheduler selects a five-level Vision-LLM compute mode, from full recomputation to multi-step temporal reuse, based on visual-language representation stability. For action generation, it selects a three-level denoising mode, reusing intermediate denoising states during stable motion while preserving full refinement for goal-sensitive stages. By coordinating these decisions, ElegantVLA offers a general acceleration framework for modern VLA pipelines with explicit action-generation modules, without modifying or retraining the base model. Experiments on GR00T and CogACT achieve up to 2.55x and 3.77x speedup, and on six real-world GR00T tasks ElegantVLA cuts computation by 2.18x while raising control frequency from 13.8 Hz to 26.3 Hz.

Hao Wang, Zhichao Chen, Jiajun Fan et al. · pku

Estimating conditional average treatment effect from observational data is highly challenging due to the existence of treatment selection bias. Prevalent methods mitigate this issue by aligning distributions of different treatment groups in the latent space. However, there are two critical problems that these methods fail to address: (1) mini-batch sampling effects (MSE), which causes misalignment in non-ideal mini-batches with outcome imbalance and outliers; (2) unobserved confounder effects (UCE), which results in inaccurate discrepancy calculation due to the neglect of unobserved confounders. To tackle these problems, we propose a principled approach named Entire Space CounterFactual Regression (ESCFR), which is a new take on optimal transport in the context of causality. Specifically, based on the framework of stochastic optimal transport, we propose a relaxed mass-preserving regularizer to address the MSE issue and design a proximal factual outcome regularizer to handle the UCE issue. Extensive experiments demonstrate that our proposed ESCFR can successfully tackle the treatment selection bias and achieve significantly better performance than state-of-the-art methods.

Ye Li, Chen Tang, Yuan Meng et al.

We introduce PRANCE, a Vision Transformer compression framework that jointly optimizes the activated channels and reduces tokens, based on the characteristics of inputs. Specifically, PRANCE~ leverages adaptive token optimization strategies for a certain computational budget, aiming to accelerate ViTs' inference from a unified data and architectural perspective. However, the joint framework poses challenges to both architectural and decision-making aspects. Firstly, while ViTs inherently support variable-token inference, they do not facilitate dynamic computations for variable channels. To overcome this limitation, we propose a meta-network using weight-sharing techniques to support arbitrary channels of the Multi-head Self-Attention and Multi-layer Perceptron layers, serving as a foundational model for architectural decision-making. Second, simultaneously optimizing the structure of the meta-network and input data constitutes a combinatorial optimization problem with an extremely large decision space, reaching up to around $10^{14}$, making supervised learning infeasible. To this end, we design a lightweight selector employing Proximal Policy Optimization for efficient decision-making. Furthermore, we introduce a novel "Result-to-Go" training mechanism that models ViTs' inference process as a Markov decision process, significantly reducing action space and mitigating delayed-reward issues during training. Extensive experiments demonstrate the effectiveness of PRANCE~ in reducing FLOPs by approximately 50\%, retaining only about 10\% of tokens while achieving lossless Top-1 accuracy. Additionally, our framework is shown to be compatible with various token optimization techniques such as pruning, merging, and sequential pruning-merging strategies. The code is available at \href{https://github.com/ChildTang/PRANCE}{https://github.com/ChildTang/PRANCE}.

Jiashuo Sun, Pengcheng Jiang, Saizhuo Wang et al.

Retrieval-Augmented Generation (RAG) systems remain brittle under realistic retrieval noise, even when the required evidence appears in the top-K results. A key reason is that retrievers and rerankers optimize solely for relevance, often selecting either trivial, answer-revealing passages or evidence that lacks the critical information required to answer the question, without considering whether the evidence is suitable for the generator. We propose BAR-RAG, which reframes the reranker as a boundary-aware evidence selector that targets the generator's Goldilocks Zone -- evidence that is neither trivially easy nor fundamentally unanswerable for the generator, but is challenging yet sufficient for inference and thus provides the strongest learning signal. BAR-RAG trains the selector with reinforcement learning using generator feedback, and adopts a two-stage pipeline that fine-tunes the generator under the induced evidence distribution to mitigate the distribution mismatch between training and inference. Experiments on knowledge-intensive question answering benchmarks show that BAR-RAG consistently improves end-to-end performance under noisy retrieval, achieving an average gain of 10.3 percent over strong RAG and reranking baselines while substantially improving robustness. Code is publicly avaliable at https://github.com/GasolSun36/BAR-RAG.

Hao Wang, Zhichao Chen, Yuan Shen et al.

Heterogeneous treatment effect (HTE) estimation from observational data poses significant challenges due to treatment selection bias. Existing methods address this bias by minimizing distribution discrepancies between treatment groups in latent space, focusing on global alignment. However, the fruitful aspect of local proximity, where similar units exhibit similar outcomes, is often overlooked. In this study, we propose Proximity-aware Counterfactual Regression (PCR) to exploit proximity for representation balancing within the HTE estimation context. Specifically, we introduce a local proximity preservation regularizer based on optimal transport to depict the local proximity in discrepancy calculation. Furthermore, to overcome the curse of dimensionality that renders the estimation of discrepancy ineffective, exacerbated by limited data availability for HTE estimation, we develop an informative subspace projector, which trades off minimal distance precision for improved sample complexity. Extensive experiments demonstrate that PCR accurately matches units across different treatment groups, effectively mitigates treatment selection bias, and significantly outperforms competitors. Code is available at https://anonymous.4open.science/status/ncr-B697.

Hao Wang, Jianxun Lian, Mingqi Wu et al.

Sequential user modeling, a critical task in personalized recommender systems, focuses on predicting the next item a user would prefer, requiring a deep understanding of user behavior sequences. Despite the remarkable success of Transformer-based models across various domains, their full potential in comprehending user behavior remains untapped. In this paper, we re-examine Transformer-like architectures aiming to advance state-of-the-art performance. We start by revisiting the core building blocks of Transformer-based methods, analyzing the effectiveness of the item-to-item mechanism within the context of sequential user modeling. After conducting a thorough experimental analysis, we identify three essential criteria for devising efficient sequential user models, which we hope will serve as practical guidelines to inspire and shape future designs. Following this, we introduce ConvFormer, a simple but powerful modification to the Transformer architecture that meets these criteria, yielding state-of-the-art results. Additionally, we present an acceleration technique to minimize the complexity associated with processing extremely long sequences. Experiments on four public datasets showcase ConvFormer's superiority and confirm the validity of our proposed criteria.

Hao Wang, Zhichao Chen, Jiajun Fan et al.

As a basic research problem for building effective recommender systems, post-click conversion rate (CVR) estimation has long been plagued by sample selection bias and data sparsity issues. To address the data sparsity issue, prevalent methods based on entire space multi-task model leverage the sequential pattern of user actions, i.e. exposure $\rightarrow$ click $\rightarrow$ conversion to construct auxiliary learning tasks. However, they still fall short of guaranteeing the unbiasedness of CVR estimates. This paper theoretically demonstrates two defects of these entire space multi-task models: (1) inherent estimation bias (IEB) for CVR estimation, where the CVR estimate is inherently higher than the ground truth; (2) potential independence priority (PIP) for CTCVR estimation, where the causality from click to conversion might be overlooked. This paper further proposes a principled method named entire space counterfactual multi-task model (ESCM$^2$), which employs a counterfactual risk minimizer to handle both IEB and PIP issues at once. To demonstrate the effectiveness of the proposed method, this paper explores its parameter tuning in practice, derives its analytic properties, and showcases its effectiveness in industrial CVR estimation, where ESCM$^2$ can effectively alleviate the intrinsic IEB and PIP issues and outperform baseline models.

Jiashuo Sun, Jimeng Shi, Yixuan Xie et al.

Retrieval-Augmented Generation (RAG) has become a standard approach for knowledge-intensive question answering, but existing systems remain brittle on multi-hop questions, where solving the task requires chaining multiple retrieval and reasoning steps. Key challenges are that current methods represent reasoning through free-form natural language, where intermediate states are implicit, retrieval queries can drift from intended entities, and errors are detected by the same model that produces them making self-reflection an unreliable, ungrounded signal. We observe that multi-hop question answering is a typical form of step-by-step computation, and that this structured process aligns closely with how code-specialized language models are trained to operate. Motivated by this, we introduce \pyrag, a framework that reformulates multi-hop RAG as program synthesis and execution. Instead of free-form reasoning trajectories, \pyrag represents the reasoning process as an executable Python program over retrieval and QA tools, exposing intermediate states as variables, producing deterministic feedback through execution, and yielding an inspectable trace of the entire reasoning process. This formulation further enables compiler-grounded self-repair and execution-driven adaptive retrieval without any additional training. Experiments on five QA benchmarks (PopQA, HotpotQA, 2WikiMultihopQA, MuSiQue, and Bamboogle) show that \pyrag consistently outperforms strong baselines under both training-free and RL-trained settings, with especially large gains on compositional multi-hop datasets. Our code, data and models are publicly available at https://github.com/GasolSun36/PyRAG.

Bangji Yang, Ruihan Guo, Jiajun Fan et al.

Generative models have achieved impressive fidelity in text-to-image synthesis, yet struggle with complex compositional prompts involving multiple constraints. We introduce \textbf{M3 (Multi-Modal, Multi-Agent, Multi-Round)}, a training-free framework that systematically resolves these failures through iterative inference-time refinement. M3 orchestrates off-the-shelf foundation models in a robust multi-agent loop: a Planner decomposes prompts into verifiable checklists, while specialized Checker, Refiner, and Editor agents surgically correct constraints one at a time, with a Verifier ensuring monotonic improvement. Applied to open-source models, M3 achieves remarkable results on the challenging OneIG-EN benchmark, with our Qwen-Image+M3 surpassing commercial flagship systems including Imagen4 (0.515) and Seedream 3.0 (0.530), reaching state-of-the-art performance (0.532 overall). This demonstrates that intelligent multi-agent reasoning can elevate open-source models beyond proprietary alternatives. M3 also substantially improves GenEval compositional metrics, effectively doubling spatial reasoning performance on hardened test sets. As a plug-and-play module compatible with any pre-trained T2I model, M3 establishes a new paradigm for compositional generation without costly retraining.

Xi Zhu, Ziqi Wang, Kai Mei et al.

Retrieval-augmented generation (RAG) improves large language models (LLMs) by incorporating external evidence, but it also introduces knowledge conflicts when retrieved contextual knowledge (CK) and parametric knowledge (PK) disagree or are both unreliable. Existing approaches mainly coordinate which source to use, without explicitly asking whether each answer path is correct. We argue that faithful RAG requires LLM self-awareness, namely the ability to recognize the limits of its own knowledge and reasoning. To ground this problem, we construct a model-specific, ground-truth-aligned knowledge-conflict benchmark by evaluating LLM backbones on PK-only and CK-conditioned answer paths over approximately 69K query-context instances per backbone, drawn from five conflict-QA datasets. We then introduce SABER, a Self-Aware Belief Estimator for RAG that requires no LLM fine-tuning. SABER combines a self-prior with PK-side and CK-side conditional reasoning representations from multi-trace inference, then estimates reliability beliefs with two lightweight predictors to drive a 4-cell decision over trust PK, trust CK, trust either, or abstain. Across four LLM backbones, SABER improves end-to-end accuracy and conflict-specific faithfulness over ten inference-time and fine-tuning baselines, with the largest gains on conflict-heavy datasets. Under abstention, SABER's risk-coverage curve Pareto-dominates every prompt-based abstainer, providing a tunable balance between coverage and answer risk. Our code is available at https://github.com/xizhu1022/SABER.

Jiajun Fan, Changnan Xiao

To obtain higher sample efficiency and superior final performance simultaneously has been one of the major challenges for deep reinforcement learning (DRL). Previous work could handle one of these challenges but typically failed to address them concurrently. In this paper, we try to tackle these two challenges simultaneously. To achieve this, we firstly decouple these challenges into two classic RL problems: data richness and exploration-exploitation trade-off. Then, we cast these two problems into the training data distribution optimization problem, namely to obtain desired training data within limited interactions, and address them concurrently via i) explicit modeling and control of the capacity and diversity of behavior policy and ii) more fine-grained and adaptive control of selective/sampling distribution of the behavior policy using a monotonic data distribution optimization. Finally, we integrate this process into Generalized Policy Iteration (GPI) and obtain a more general framework called Generalized Data Distribution Iteration (GDI). We use the GDI framework to introduce operator-based versions of well-known RL methods from DQN to Agent57. Theoretical guarantee of the superiority of GDI compared with GPI is concluded. We also demonstrate our state-of-the-art (SOTA) performance on Arcade Learning Environment (ALE), wherein our algorithm has achieved 9620.33% mean human normalized score (HNS), 1146.39% median HNS and surpassed 22 human world records using only 200M training frames. Our performance is comparable to Agent57's while we consume 500 times less data. We argue that there is still a long way to go before obtaining real superhuman agents in ALE.

Hao Li, Hao Jiang, Jiajun Fan et al.

Persistent homology, a fundamental technique within Topological Data Analysis (TDA), captures structural and shape characteristics of graphs, yet encounters computational difficulties when applied to dynamic directed graphs. This paper introduces the Dynamic Neural Dowker Network (DNDN), a novel framework specifically designed to approximate the results of dynamic Dowker filtration, aiming to capture the high-order topological features of dynamic directed graphs. Our approach creatively uses line graph transformations to produce both source and sink line graphs, highlighting the shared neighbor structures that Dowker complexes focus on. The DNDN incorporates a Source-Sink Line Graph Neural Network (SSLGNN) layer to effectively capture the neighborhood relationships among dynamic edges. Additionally, we introduce an innovative duality edge fusion mechanism, ensuring that the results for both the sink and source line graphs adhere to the duality principle intrinsic to Dowker complexes. Our approach is validated through comprehensive experiments on real-world datasets, demonstrating DNDN's capability not only to effectively approximate dynamic Dowker filtration results but also to perform exceptionally in dynamic graph classification tasks.

Ye Li, Yuan Meng, Zewen Sun et al.

Vision-Language-Action (VLA) models have attracted increasing attention for their strong control capabilities. However, their high computational cost and low execution frequency hinder their suitability for real-time tasks such as robotic manipulation and autonomous navigation. Existing VLA acceleration methods primarily focus on structural optimization, overlooking the fact that these models operate in sequential decision-making environments. As a result, temporal redundancy in sequential action generation and spatial redundancy in visual input remain unaddressed. To this end, we propose SP-VLA, a unified framework that accelerates VLA models by jointly scheduling models and pruning tokens. Specifically, we design an action-aware model scheduling mechanism that reduces temporal redundancy by dynamically switching between VLA model and a lightweight generator. Inspired by the human motion pattern of focusing on key decision points while relying on intuition for other actions, we categorize VLA actions into deliberative and intuitive, assigning the former to the VLA model and the latter to the lightweight generator, enabling frequency-adaptive execution through collaborative model scheduling. To address spatial redundancy, we further develop a spatio-semantic dual-aware token pruning method. Tokens are classified into spatial and semantic types and pruned based on their dual-aware importance to accelerate VLA inference. These two mechanisms work jointly to guide the VLA in focusing on critical actions and salient visual information, achieving effective acceleration while maintaining high accuracy. Extensive experiments show that our method achieves 1.5$\times$ lossless acceleration in LIBERO and 2.4$\times$ in SimplerEnv, with up to 6% average performance gain. Inference frequency and latency improve by 2.2$\times$ in SimplerEnv and 1.4$\times$ in LIBERO.

Jiajun Fan, Shuaike Shen, Chaoran Cheng et al.

Recent advancements in reinforcement learning (RL) have achieved great success in fine-tuning diffusion-based generative models. However, fine-tuning continuous flow-based generative models to align with arbitrary user-defined reward functions remains challenging, particularly due to issues such as policy collapse from overoptimization and the prohibitively high computational cost of likelihoods in continuous-time flows. In this paper, we propose an easy-to-use and theoretically sound RL fine-tuning method, which we term Online Reward-Weighted Conditional Flow Matching with Wasserstein-2 Regularization (ORW-CFM-W2). Our method integrates RL into the flow matching framework to fine-tune generative models with arbitrary reward functions, without relying on gradients of rewards or filtered datasets. By introducing an online reward-weighting mechanism, our approach guides the model to prioritize high-reward regions in the data manifold. To prevent policy collapse and maintain diversity, we incorporate Wasserstein-2 (W2) distance regularization into our method and derive a tractable upper bound for it in flow matching, effectively balancing exploration and exploitation of policy optimization. We provide theoretical analyses to demonstrate the convergence properties and induced data distributions of our method, establishing connections with traditional RL algorithms featuring Kullback-Leibler (KL) regularization and offering a more comprehensive understanding of the underlying mechanisms and learning behavior of our approach. Extensive experiments on tasks including target image generation, image compression, and text-image alignment demonstrate the effectiveness of our method, where our method achieves optimal policy convergence while allowing controllable trade-offs between reward maximization and diversity preservation.

Ziwen Wang, Jiajun Fan, Ruihan Guo et al.

Protein generative models have shown remarkable promise in protein design but still face limitations in success rate, due to the scarcity of high-quality protein datasets for supervised pretraining. We present ProteinZero, a novel framework that enables scalable, automated, and continuous self-improvement of the inverse folding model through online reinforcement learning. To achieve computationally tractable online feedback, we introduce efficient proxy reward models based on ESM-fold and a novel rapid ddG predictor that significantly accelerates evaluation speed. ProteinZero employs a general RL framework balancing multi-reward maximization, KL-divergence from a reference model, and a novel protein-embedding level diversity regularization that prevents mode collapse while promoting higher sequence diversity. Through extensive experiments, we demonstrate that ProteinZero substantially outperforms existing methods across every key metric in protein design, achieving significant improvements in structural accuracy, designability, thermodynamic stability, and sequence diversity. Most impressively, ProteinZero reduces design failure rates by approximately 36% - 48% compared to widely-used methods like ProteinMPNN, ESM-IF and InstructPLM, consistently achieving success rates exceeding 90% across diverse and complex protein folds. Notably, the entire RL run on CATH-4.3 can be done with a single 8 X GPU node in under 3 days, including reward computation. Our work establishes a new paradigm for protein design where models evolve continuously from their own generated outputs, opening new possibilities for exploring the vast protein design space.

Ziwen Wang, Jiajun Fan, Thao Nguyen et al.

Contrastive learning has proven to be highly efficient and adaptable in shaping representation spaces across diverse modalities by pulling similar samples together and pushing dissimilar ones apart. However, two key limitations persist: (1) Without explicit regulation of the embedding distribution, semantically related instances can inadvertently be pushed apart unless complementary signals guide pair selection, and (2) excessive reliance on large in-batch negatives and tailored augmentations hinders generalization. To address these limitations, we propose Variational Supervised Contrastive Learning (VarCon), which reformulates supervised contrastive learning as variational inference over latent class variables and maximizes a posterior-weighted evidence lower bound (ELBO) that replaces exhaustive pair-wise comparisons for efficient class-aware matching and grants fine-grained control over intra-class dispersion in the embedding space. Trained exclusively on image data, our experiments on CIFAR-10, CIFAR-100, ImageNet-100, and ImageNet-1K show that VarCon (1) achieves state-of-the-art performance for contrastive learning frameworks, reaching 79.36% Top-1 accuracy on ImageNet-1K and 78.29% on CIFAR-100 with a ResNet-50 encoder while converging in just 200 epochs; (2) yields substantially clearer decision boundaries and semantic organization in the embedding space, as evidenced by KNN classification, hierarchical clustering results, and transfer-learning assessments; and (3) demonstrates superior performance in few-shot learning than supervised baseline and superior robustness across various augmentation strategies.

Chaoran Cheng, Jiahan Li, Jiajun Fan et al.

Recent efforts have extended the flow-matching framework to discrete generative modeling. One strand of models directly works with the continuous probabilities instead of discrete tokens, which we colloquially refer to as Continuous-State Discrete Flow Matching (CS-DFM). Existing CS-DFM models differ significantly in their representations and geometric assumptions. This work presents a unified framework for CS-DFM models, under which the existing variants can be understood as operating on different $α$-representations of probabilities. Building upon the theory of information geometry, we introduce $α$-Flow, a family of CS-DFM models that adheres to the canonical $α$-geometry of the statistical manifold, and demonstrate its optimality in minimizing the generalized kinetic energy. Theoretically, we show that the flow matching loss for $α$-flow establishes a unified variational bound for the discrete negative log-likelihood. We comprehensively evaluate different instantiations of $α$-flow on various discrete generation domains to demonstrate their effectiveness in discrete generative modeling, including intermediate values whose geometries have never been explored before. $α$-flow significantly outperforms its discrete-state counterpart in image and protein sequence generation and better captures the entropy in language modeling.

Jiajun Fan, Roger Ren, Jingyuan Li et al.

The role of reasoning in Audio Large Language Models remains widely underexplored, as introducing a reasoning process often degrades rather than improves performance during inference, a phenomenon we term test-time inverse scaling, where longer reasoning chains yield progressively worse results. We demonstrate that this stems not from fundamental limitations of reasoning itself, but from inadequate training: models without proper guidance for the reasoning process produce hallucinatory, inconsistent reasoning that accumulates errors over longer chains. To address these challenges, we introduce CESAR (Consistent, Effective, and Scalable Audio Reasoners), shifting from outcome verification to rewarding the reasoning process. Our online reinforcement learning framework employs Group Relative Policy Optimization with a multi-faceted reward suite that incentivizes not only correctness and format but also consistency, structured analytical patterns, causal reasoning, domain-knowledge integration, and calibrated reasoning depth. CESAR resolves test-time inverse scaling, transforming reasoning from detriments into gains while revealing model-specific ``reasoning sweet spots", where performance peaks during test-time scaling. We achieve state-of-the-art results on MMAU Test-mini, substantially outperforming Gemini 2.5 Pro and GPT-4o Audio, and near-human-level performance on MMSU reasoning tasks. Through AI-as-judge evaluations and qualitative comparisons, we provide both quantitative and qualitative validation of our improved reasoning quality. Importantly, enhanced reasoning creates synergistic effects, simultaneously improving multimodal reasoning and perception capabilities. Overall, CESAR establishes a principled method for developing robust and scalable reasoning in Audio LLMs.

Jiajun Fan, Chaoran Cheng, Shuaike Shen et al.

Flow-based generative models have shown remarkable success in text-to-image generation, yet fine-tuning them with intermediate feedback remains challenging, especially for continuous-time flow matching models. Most existing approaches solely learn from outcome rewards, struggling with the credit assignment problem. Alternative methods that attempt to learn a critic via direct regression on cumulative rewards often face training instabilities and model collapse in online settings. We present AC-Flow, a robust actor-critic framework that addresses these challenges through three key innovations: (1) reward shaping that provides well-normalized learning signals to enable stable intermediate value learning and gradient control, (2) a novel dual-stability mechanism that combines advantage clipping to prevent destructive policy updates with a warm-up phase that allows the critic to mature before influencing the actor, and (3) a scalable generalized critic weighting scheme that extends traditional reward-weighted methods while preserving model diversity through Wasserstein regularization. Through extensive experiments on Stable Diffusion 3, we demonstrate that AC-Flow achieves state-of-the-art performance in text-to-image alignment tasks and generalization to unseen human preference models. Our results demonstrate that even with a computationally efficient critic model, we can robustly finetune flow models without compromising generative quality, diversity, or stability.

Jiajun Fan, Tong Wei, Chaoran Cheng et al.

Balancing exploration and exploitation during reinforcement learning fine-tuning of generative models presents a critical challenge, as existing approaches rely on fixed divergence regularization that creates an inherent dilemma: strong regularization preserves model capabilities but limits reward optimization, while weak regularization enables greater alignment but risks instability or reward hacking. We introduce Adaptive Divergence Regularized Policy Optimization (ADRPO), which automatically adjusts regularization strength based on advantage estimates-reducing regularization for high-value samples while applying stronger regularization to poor samples, enabling policies to navigate between exploration and aggressive exploitation according to data quality. Our implementation with Wasserstein-2 regularization for flow matching generative models achieves remarkable results on text-to-image generation, achieving better semantic alignment and diversity than offline methods like DPO and online methods with fixed regularization like ORW-CFM-W2. ADRPO enables a 2B parameter SD3 model to surpass much larger models with 4.8B and 12B parameters in attribute binding, semantic consistency, artistic style transfer, and compositional control while maintaining generation diversity. ADRPO generalizes to KL-regularized fine-tuning of both text-only LLMs and multi-modal reasoning models, enhancing existing online RL methods like GRPO. In LLM fine-tuning, ADRPO demonstrates an emergent ability to escape local optima through active exploration, while in multi-modal audio reasoning, it outperforms GRPO through superior step-by-step reasoning, enabling a 7B model to outperform substantially larger commercial models including Gemini 2.5 Pro and GPT-4o Audio, offering an effective plug-and-play solution to the exploration-exploitation challenge across diverse generative architectures and modalities.

Bangji Yang, Hongbo Ma, Jiajun Fan et al.

Large Language Models employing Chain-of-Thought reasoning achieve strong performance but suffer from excessive token consumption that inflates inference costs. Existing efficiency methods such as explicit length penalties, difficulty estimators, or multi-stage curricula either degrade reasoning quality or require complex training pipelines. We introduce Batched Contextual Reinforcement, a minimalist, single-stage training paradigm that unlocks efficient reasoning through a simple structural modification: training the model to solve N problems simultaneously within a shared context window, rewarded purely by per-instance accuracy. This formulation creates an implicit token budget that yields several key findings: (1) We identify a novel task-scaling law: as the number of concurrent problems N increases during inference, per-problem token usage decreases monotonically while accuracy degrades far more gracefully than baselines, establishing N as a controllable throughput dimension. (2) BCR challenges the traditional accuracy-efficiency trade-off by demonstrating a "free lunch" phenomenon at standard single-problem inference. Across both 1.5B and 4B model families, BCR reduces token usage by 15.8% to 62.6% while consistently maintaining or improving accuracy across five major mathematical benchmarks. (3) Qualitative analyses reveal emergent self-regulated efficiency, where models autonomously eliminate redundant metacognitive loops without explicit length supervision. (4) Crucially, we empirically demonstrate that implicit budget constraints successfully circumvent the adversarial gradients and catastrophic optimization collapse inherent to explicit length penalties, offering a highly stable, constraint-based alternative for length control. These results prove BCR practical, showing simple structural incentives unlock latent high-density reasoning in LLMs.

Jiajun Fan, Yuzheng Zhuang, Yuecheng Liu et al.

The exploration problem is one of the main challenges in deep reinforcement learning (RL). Recent promising works tried to handle the problem with population-based methods, which collect samples with diverse behaviors derived from a population of different exploratory policies. Adaptive policy selection has been adopted for behavior control. However, the behavior selection space is largely limited by the predefined policy population, which further limits behavior diversity. In this paper, we propose a general framework called Learnable Behavioral Control (LBC) to address the limitation, which a) enables a significantly enlarged behavior selection space via formulating a hybrid behavior mapping from all policies; b) constructs a unified learnable process for behavior selection. We introduce LBC into distributed off-policy actor-critic methods and achieve behavior control via optimizing the selection of the behavior mappings with bandit-based meta-controllers. Our agents have achieved 10077.52% mean human normalized score and surpassed 24 human world records within 1B training frames in the Arcade Learning Environment, which demonstrates our significant state-of-the-art (SOTA) performance without degrading the sample efficiency.

Jiajun Fan

The Arcade Learning Environment (ALE) is proposed as an evaluation platform for empirically assessing the generality of agents across dozens of Atari 2600 games. ALE offers various challenging problems and has drawn significant attention from the deep reinforcement learning (RL) community. From Deep Q-Networks (DQN) to Agent57, RL agents seem to achieve superhuman performance in ALE. However, is this the case? In this paper, to explore this problem, we first review the current evaluation metrics in the Atari benchmarks and then reveal that the current evaluation criteria of achieving superhuman performance are inappropriate, which underestimated the human performance relative to what is possible. To handle those problems and promote the development of RL research, we propose a novel Atari benchmark based on human world records (HWR), which puts forward higher requirements for RL agents on both final performance and learning efficiency. Furthermore, we summarize the state-of-the-art (SOTA) methods in Atari benchmarks and provide benchmark results over new evaluation metrics based on human world records. We concluded that at least four open challenges hinder RL agents from achieving superhuman performance from those new benchmark results. Finally, we also discuss some promising ways to handle those problems.

Jiajun Fan, Changnan Xiao, Yue Huang

Deep Q Network (DQN) firstly kicked the door of deep reinforcement learning (DRL) via combining deep learning (DL) with reinforcement learning (RL), which has noticed that the distribution of the acquired data would change during the training process. DQN found this property might cause instability for training, so it proposed effective methods to handle the downside of the property. Instead of focusing on the unfavourable aspects, we find it critical for RL to ease the gap between the estimated data distribution and the ground truth data distribution while supervised learning (SL) fails to do so. From this new perspective, we extend the basic paradigm of RL called the Generalized Policy Iteration (GPI) into a more generalized version, which is called the Generalized Data Distribution Iteration (GDI). We see massive RL algorithms and techniques can be unified into the GDI paradigm, which can be considered as one of the special cases of GDI. We provide theoretical proof of why GDI is better than GPI and how it works. Several practical algorithms based on GDI have been proposed to verify the effectiveness and extensiveness of it. Empirical experiments prove our state-of-the-art (SOTA) performance on Arcade Learning Environment (ALE), wherein our algorithm has achieved 9620.98% mean human normalized score (HNS), 1146.39% median HNS and 22 human world record breakthroughs (HWRB) using only 200M training frames. Our work aims to lead the RL research to step into the journey of conquering the human world records and seek real superhuman agents on both performance and efficiency.

Changnan Xiao, Haosen Shi, Jiajun Fan et al.

Policy-based reinforcement learning methods suffer from the policy collapse problem. We find valued-based reinforcement learning methods with ε-greedy mechanism are capable of enjoying three characteristics, Closed-form Diversity, Objective-invariant Exploration and Adaptive Trade-off, which help value-based methods avoid the policy collapse problem. However, there does not exist a parallel mechanism for policy-based methods that achieves all three characteristics. In this paper, we propose an entropy regularization free mechanism that is designed for policy-based methods, which achieves Closed-form Diversity, Objective-invariant Exploration and Adaptive Trade-off. Our experiments show that our mechanism is super sample-efficient for policy-based methods and boosts a policy-based baseline to a new State-Of-The-Art on Arcade Learning Environment.

Changnan Xiao, Haosen Shi, Jiajun Fan et al.

We study the problem of model-free reinforcement learning, which is often solved following the principle of Generalized Policy Iteration (GPI). While GPI is typically an interplay between policy evaluation and policy improvement, most conventional model-free methods assume the independence of the granularity and other details of the GPI steps, despite of the inherent connections between them. In this paper, we present a method that regularizes the inconsistency between policy evaluation and policy improvement, leading to a conflict averse GPI solution with reduced functional approximation error. To this end, we formulate a novel learning paradigm where taking the policy evaluation step is equivalent to some compensation of performing policy improvement, and thus effectively alleviates the gradient conflict between the two GPI steps. We also show that the form of our proposed solution is equivalent to performing entropy-regularized policy improvement and therefore prevents the policy from being trapped into suboptimal solutions. We conduct extensive experiments to evaluate our method on the Arcade Learning Environment (ALE). Empirical results show that our method outperforms several strong baselines in major evaluation domains.

Jiajun Fan, He Ba, Xian Guo et al.

Constructing agents with planning capabilities has long been one of the main challenges in the pursuit of artificial intelligence. Tree-based planning methods from AlphaGo to Muzero have enjoyed huge success in discrete domains, such as chess and Go. Unfortunately, in real-world applications like robot control and inverted pendulum, whose action space is normally continuous, those tree-based planning techniques will be struggling. To address those limitations, in this paper, we present a novel model-based reinforcement learning frameworks called Critic PI2, which combines the benefits from trajectory optimization, deep actor-critic learning, and model-based reinforcement learning. Our method is evaluated for inverted pendulum models with applicability to many continuous control systems. Extensive experiments demonstrate that Critic PI2 achieved a new state of the art in a range of challenging continuous domains. Furthermore, we show that planning with a critic significantly increases the sample efficiency and real-time performance. Our work opens a new direction toward learning the components of a model-based planning system and how to use them.