Jianfei Gao

Chengqi Lyu, Songyang Gao, Yuzhe Gu et al.

Reasoning abilities, especially those for solving complex math problems, are crucial components of general intelligence. Recent advances by proprietary companies, such as o-series models of OpenAI, have made remarkable progress on reasoning tasks. However, the complete technical details remain unrevealed, and the techniques that are believed certainly to be adopted are only reinforcement learning (RL) and the long chain of thoughts. This paper proposes a new RL framework, termed OREAL, to pursue the performance limit that can be achieved through \textbf{O}utcome \textbf{RE}w\textbf{A}rd-based reinforcement \textbf{L}earning for mathematical reasoning tasks, where only binary outcome rewards are easily accessible. We theoretically prove that behavior cloning on positive trajectories from best-of-N (BoN) sampling is sufficient to learn the KL-regularized optimal policy in binary feedback environments. This formulation further implies that the rewards of negative samples should be reshaped to ensure the gradient consistency between positive and negative samples. To alleviate the long-existing difficulties brought by sparse rewards in RL, which are even exacerbated by the partial correctness of the long chain of thought for reasoning tasks, we further apply a token-level reward model to sample important tokens in reasoning trajectories for learning. With OREAL, for the first time, a 7B model can obtain 94.0 pass@1 accuracy on MATH-500 through RL, being on par with 32B models. OREAL-32B also surpasses previous 32B models trained by distillation with 95.0 pass@1 accuracy on MATH-500. Our investigation also indicates the importance of initial policy models and training queries for RL. Code, models, and data will be released to benefit future research\footnote{https://github.com/InternLM/OREAL}.

Kai Liu, Ruohui Wang, Jianfei Gao et al.

Over the past few years, as large language models have ushered in an era of intelligence emergence, there has been an intensified focus on scaling networks. Currently, many network architectures are designed manually, often resulting in sub-optimal configurations. Although Neural Architecture Search (NAS) methods have been proposed to automate this process, they suffer from low search efficiency. This study introduces Differentiable Model Scaling (DMS), increasing the efficiency for searching optimal width and depth in networks. DMS can model both width and depth in a direct and fully differentiable way, making it easy to optimize. We have evaluated our DMS across diverse tasks, ranging from vision tasks to NLP tasks and various network architectures, including CNNs and Transformers. Results consistently indicate that our DMS can find improved structures and outperforms state-of-the-art NAS methods. Specifically, for image classification on ImageNet, our DMS improves the top-1 accuracy of EfficientNet-B0 and Deit-Tiny by 1.4% and 0.6%, respectively, and outperforms the state-of-the-art zero-shot NAS method, ZiCo, by 1.3% while requiring only 0.4 GPU days for searching. For object detection on COCO, DMS improves the mAP of Yolo-v8-n by 2.0%. For language modeling, our pruned Llama-7B outperforms the prior method with lower perplexity and higher zero-shot classification accuracy. We will release our code in the future.

Junhao Shen, Haiteng Zhao, Yuzhe Gu et al.

Enhancing large vision-language models (LVLMs) with visual slow-thinking reasoning is crucial for solving complex multimodal tasks. However, since LVLMs are mainly trained with vision-language alignment, it is difficult to adopt on-policy reinforcement learning (RL) to develop the slow thinking ability because the rollout space is restricted by its initial abilities. Off-policy RL offers a way to go beyond the current policy, but directly distilling trajectories from external models may cause visual hallucinations due to mismatched visual perception abilities across models. To address these issues, this paper proposes SOPHIA, a simple and scalable Semi-Off-Policy RL for vision-language slow-tHInking reAsoning. SOPHIA builds a semi-off-policy behavior model by combining on-policy visual understanding from a trainable LVLM with off-policy slow-thinking reasoning from a language model, assigns outcome-based rewards to reasoning, and propagates visual rewards backward. Then LVLM learns slow-thinking reasoning ability from the obtained reasoning trajectories using propagated rewards via off-policy RL algorithms. Extensive experiments with InternVL2.5 and InternVL3.0 with 8B and 38B sizes show the effectiveness of SOPHIA. Notably, SOPHIA improves InternVL3.0-38B by 8.50% in average, reaching state-of-the-art performance among open-source LVLMs on multiple multimodal reasoning benchmarks, and even outperforms some closed-source models (e.g., GPT-4.1) on the challenging MathVision and OlympiadBench, achieving 49.08% and 49.95% pass@1 accuracy, respectively. Analysis shows SOPHIA outperforms supervised fine-tuning and direct on-policy RL methods, offering a better policy initialization for further on-policy training.

Changyong Shu, Yifan Liu, Jianfei Gao et al.

Knowledge distillation (KD) has been proven to be a simple and effective tool for training compact models. Almost all KD variants for dense prediction tasks align the student and teacher networks' feature maps in the spatial domain, typically by minimizing point-wise and/or pair-wise discrepancy. Observing that in semantic segmentation, some layers' feature activations of each channel tend to encode saliency of scene categories (analogue to class activation mapping), we propose to align features channel-wise between the student and teacher networks. To this end, we first transform the feature map of each channel into a probabilty map using softmax normalization, and then minimize the Kullback-Leibler (KL) divergence of the corresponding channels of the two networks. By doing so, our method focuses on mimicking the soft distributions of channels between networks. In particular, the KL divergence enables learning to pay more attention to the most salient regions of the channel-wise maps, presumably corresponding to the most useful signals for semantic segmentation. Experiments demonstrate that our channel-wise distillation outperforms almost all existing spatial distillation methods for semantic segmentation considerably, and requires less computational cost during training. We consistently achieve superior performance on three benchmarks with various network structures. Code is available at: https://git.io/Distiller

Zhonghan Zhao, Wenwei Zhang, Haian Huang et al.

Reasoning before action and imagining potential outcomes (i.e., world models) are essential for embodied agents operating in complex open-world environments. Yet, prior work either incorporates only one of these abilities in an end-to-end agent or integrates multiple specialized models into an agent system, limiting the learning efficiency and generalization of the policy. Thus, this paper makes the first attempt to synergize Reasoning and Imagination in an end-to-end Generalist policy, termed RIG. To train RIG in an end-to-end manner, we construct a data pipeline that progressively integrates and enriches the content of imagination and reasoning in the trajectories collected from existing agents. The joint learning of reasoning and next image generation explicitly models the inherent correlation between reasoning, action, and dynamics of environments, and thus exhibits more than $17\times$ sample efficiency improvements and generalization in comparison with previous works. During inference, RIG first reasons about the next action, produces potential action, and then predicts the action outcomes, which offers the agent a chance to review and self-correct based on the imagination before taking real actions. Experimental results show that the synergy of reasoning and imagination not only improves the robustness, generalization, and interoperability of generalist policy but also enables test-time scaling to enhance overall performance.

Songyang Gao, Yuzhe Gu, Zijian Wu et al.

Large Reasoning Models (LRMs) have expanded the mathematical reasoning frontier through Chain-of-Thought (CoT) techniques and Reinforcement Learning with Verifiable Rewards (RLVR), capable of solving AIME-level problems. However, the performance of LRMs is heavily dependent on the extended reasoning context length. For solving ultra-hard problems like those in the International Mathematical Olympiad (IMO), the required reasoning complexity surpasses the space that an LRM can explore in a single round. Previous works attempt to extend the reasoning context of LRMs but remain prompt-based and built upon proprietary models, lacking systematic structures and training pipelines. Therefore, this paper introduces Intern-S1-MO, a long-horizon math agent that conducts multi-round hierarchical reasoning, composed of an LRM-based multi-agent system including reasoning, summary, and verification. By maintaining a compact memory in the form of lemmas, Intern-S1-MO can more freely explore the lemma-rich reasoning spaces in multiple reasoning stages, thereby breaking through the context constraints for IMO-level math problems. Furthermore, we propose OREAL-H, an RL framework for training the LRM using the online explored trajectories to simultaneously bootstrap the reasoning ability of LRM and elevate the overall performance of Intern-S1-MO. Experiments show that Intern-S1-MO can obtain 26 out of 35 points on the non-geometry problems of IMO2025, matching the performance of silver medalists. It also surpasses the current advanced LRMs on inference benchmarks such as HMMT2025, AIME2025, and CNMO2025. In addition, our agent officially participates in CMO2025 and achieves a score of 102/126 under the judgment of human experts, reaching the gold medal level.

Jianfei Gao, Bruno Ribeiro

This work formalizes the associational task of predicting node attribute evolution in temporal graphs from the perspective of learning equivariant representations. We show that node representations in temporal graphs can be cast into two distinct frameworks: (a) The most popular approach, which we denote as time-and-graph, where equivariant graph (e.g., GNN) and sequence (e.g., RNN) representations are intertwined to represent the temporal evolution of node attributes in the graph; and (b) an approach that we denote as time-then-graph, where the sequences describing the node and edge dynamics are represented first, then fed as node and edge attributes into a static equivariant graph representation that comes after. Interestingly, we show that time-then-graph representations have an expressivity advantage over time-and-graph representations when both use component GNNs that are not most-expressive (e.g., 1-Weisfeiler-Lehman GNNs). Moreover, while our goal is not necessarily to obtain state-of-the-art results, our experiments show that time-then-graph methods are capable of achieving better performance and efficiency than state-of-the-art time-and-graph methods in some real-world tasks, thereby showcasing that the time-then-graph framework is a worthy addition to the graph ML toolbox.

Jianfei Gao, Mohamed A. Zahran, Amit Sheoran et al.

We consider the task of learning a parametric Continuous Time Markov Chain (CTMC) sequence model without examples of sequences, where the training data consists entirely of aggregate steady-state statistics. Making the problem harder, we assume that the states we wish to predict are unobserved in the training data. Specifically, given a parametric model over the transition rates of a CTMC and some known transition rates, we wish to extrapolate its steady state distribution to states that are unobserved. A technical roadblock to learn a CTMC from its steady state has been that the chain rule to compute gradients will not work over the arbitrarily long sequences necessary to reach steady state ---from where the aggregate statistics are sampled. To overcome this optimization challenge, we propose $\infty$-SGD, a principled stochastic gradient descent method that uses randomly-stopped estimators to avoid infinite sums required by the steady state computation, while learning even when only a subset of the CTMC states can be observed. We apply $\infty$-SGD to a real-world testbed and synthetic experiments showcasing its accuracy, ability to extrapolate the steady state distribution to unobserved states under unobserved conditions (heavy loads, when training under light loads), and succeeding in difficult scenarios where even a tailor-made extension of existing methods fails.