Xiangyuan Wang

MiniMax, Aili Chen, Aonian Li et al.

We introduce the MiniMax-M2 series, a family of Mixture-of-Experts language models built around the principle that mini activations can unleash maximum real-world intelligence. The flagship M2 contains 229.9B total parameters with only 9.8B activated per token. Designed end-to-end for agentic deployment, the M2 series rests on three components: (i) agent-driven data pipelines producing large-scale, verifiable trajectories across agentic coding and agentic cowork, each grounded in an executable workspace and an artifact-aligned reward; (ii) Forge, a scalable agent-native RL system that adapts to long-horizon agent trajectories, paired with windowed-FIFO scheduling, prefix-tree merging, inference optimization, and a clean training-inference-agent decoupling that supports both white-box and black-box agents; (iii) the latest M2.7 checkpoint takes an early step toward self-evolution -- autonomously debugging training runs and modifying its own scaffold. Across M2 through M2.7, this combination translates a mini-activation footprint into frontier-tier performance on agentic coding, deep search, office-task, and reasoning benchmarks.

Haozhe Wu, Jia Jia, Junliang Xing et al.

Audio-Driven Face Animation is an eagerly anticipated technique for applications such as VR/AR, games, and movie making. With the rapid development of 3D engines, there is an increasing demand for driving 3D faces with audio. However, currently available 3D face animation datasets are either scale-limited or quality-unsatisfied, which hampers further developments of audio-driven 3D face animation. To address this challenge, we propose MMFace4D, a large-scale multi-modal 4D (3D sequence) face dataset consisting of 431 identities, 35,904 sequences, and 3.9 million frames. MMFace4D exhibits two compelling characteristics: 1) a remarkably diverse set of subjects and corpus, encompassing actors spanning ages 15 to 68, and recorded sentences with durations ranging from 0.7 to 11.4 seconds. 2) It features synchronized audio and 3D mesh sequences with high-resolution face details. To capture the subtle nuances of 3D facial expressions, we leverage three synchronized RGBD cameras during the recording process. Upon MMFace4D, we construct a non-autoregressive framework for audio-driven 3D face animation. Our framework considers the regional and composite natures of facial animations, and surpasses contemporary state-of-the-art approaches both qualitatively and quantitatively. The code, model, and dataset will be publicly available.

Xiangyuan Wang, Honghao Cai, Yunhao Bai et al.

High-quality training triplets (source-target image pairs with precise editing instructions) are a critical bottleneck for scaling instruction-guided image editing models. Vision-language models (VLMs) are widely used for automated instruction synthesis, but we identify three systematic failure modes in image-pair settings: orientation inconsistency (e.g., left/right confusion), viewpoint ambiguity, and insufficient fine-grained attribute description. Human evaluation shows that over 47% of instructions from strong baseline VLMs contain critical errors unusable for downstream training. We propose EditCaption, a scalable two-stage post-training pipeline for VLM-based instruction synthesis. Stage 1 builds a 100K supervised fine-tuning (SFT) dataset by combining GLM automatic annotation, EditScore-based filtering, and human refinement for spatial, directional, and attribute-level accuracy. Stage 2 collects 10K human preference pairs targeting the three failure modes and applies direct preference optimization (DPO) for alignment beyond SFT alone. On Eval-400, ByteMorph-Bench, and HQ-Edit, fine-tuned Qwen3-VL models outperform open-source baselines; the 235B model reaches 4.712 on Eval-400 (vs. Gemini-3-Pro 4.706, GPT-4.1 4.220, Kimi-K2.5 4.111) and 4.588 on ByteMorph-Bench (vs. Gemini-3-Pro 4.522, GPT-4.1 3.412). Human evaluation shows critical errors falling from 47.75% to 23% and correctness rising from 41.75% to 66%. The work offers a practical path to scalable, human-aligned instruction synthesis for image editing data.

Honghao Cai, Xiangyuan Wang, Yunhao Bai et al.

Large diffusion transformers (DiTs) follow global editing instructions well but consistently leak local edits into unrelated regions, because joint-attention architectures offer no explicit channel telling the network where to apply the edit. We introduce REDEdit, a co-trained, instruction- and region-aware adapter framework that retrofits a frozen DiT into a precise local editor without modifying its backbone weights. A lightweight Block Adapter at every transformer block injects a structured condition stream that factorizes what to edit (instruction semantics) from where to edit (spatial mask); a learned SpatialGate routes the adapter signal selectively into the edit region while keeping the rest of the image near-identical to the source; and a Region-Aware Loss focuses the training objective on the changing pixels. Because these components make the backbone's internal representation mask-aware end-to-end, a thin MaskPredictor head trained jointly with the editor can ground the edit region directly from the instruction and source image eliminating any user-mask requirement at deployment. We evaluate on two complementary benchmarks: MagicBrush (paired ground-truth targets) to measure pixel-level preservation and edit accuracy, and Emu-Edit Test (no ground-truth images, 9 diverse edit categories) to stress-test instruction following and generalization across edit types. On both, REDEdit achieves state-of-the-art results, simultaneously outperforming mask-free and oracle-mask baselines. A seven-variant ablation cleanly isolates the contribution of each component.

Zhaosong Lu, Xiangyuan Wang

We study a class of nonconvex-nonconcave minimax problems in which the inner maximization problem satisfies a local Kurdyka-Łojasiewicz (KL) condition that may vary with the outer minimization variable. In contrast to the global KL or Polyak-Łojasiewicz (PL) conditions commonly assumed in the literature -- which are significantly stronger and often too restrictive in practice -- this local KL condition accommodates a broader range of practical scenarios. However, it also introduces new analytical challenges. In particular, as an optimization algorithm progresses toward a stationary point of the problem, the region over which the KL condition holds may shrink, resulting in a more intricate and potentially ill-conditioned landscape. To address this challenge, we show that the associated maximal function is locally Hölder smooth. Leveraging this key property, we develop an inexact proximal gradient method for solving the minimax problem, where the inexact gradient of the maximal function is computed by applying a proximal gradient method to a KL-structured subproblem. Under mild assumptions, we establish complexity guarantees for computing an approximate stationary point of the minimax problem.

Zhaosong Lu, Xiangyuan Wang

We study a class of constrained nonconvex--nonconcave minimax problems in which the inner maximization involves potentially complex constraints. Under the assumption that the inner problem of a novel lifted minimax problem satisfies a local Kurdyka-Łojasiewicz (KL) condition, we show that the maximal function of the original problem enjoys a local Hölder smoothness property. We also propose a sequential convex programming (SCP) method for solving constrained optimization problems and establish its convergence rate under a local KL condition. Leveraging these results, we develop an inexact proximal gradient method for the original minimax problem, where the inexact gradient of the maximal function is computed via the SCP method applied to a locally KL-structured subproblem. Finally, we establish complexity guarantees for the proposed method in computing an approximate stationary point of the original minimax problem.

Chao Yu, Yuanqing Wang, Zhen Guo et al.

Reinforcement learning (RL) has demonstrated immense potential in advancing artificial general intelligence, agentic intelligence, and embodied intelligence. However, the inherent heterogeneity and dynamicity of RL workflows often lead to low hardware utilization and slow training on existing systems. In this paper, we present RLinf, a high-performance RL training system based on our key observation that the major roadblock to efficient RL training lies in system flexibility. To maximize flexibility and efficiency, RLinf is built atop a novel RL system design paradigm called macro-to-micro flow transformation (M2Flow), which automatically breaks down high-level, easy-to-compose RL workflows at both the temporal and spatial dimensions, and recomposes them into optimized execution flows. Supported by RLinf worker's adaptive communication capability, we devise context switching and elastic pipelining to realize M2Flow transformation, and a profiling-guided scheduling policy to generate optimal execution plans. Extensive evaluations on both reasoning RL and embodied RL tasks demonstrate that RLinf consistently outperforms state-of-the-art systems, achieving 1.1x-2.13x speedup in end-to-end training throughput.

Haitian Zhang, Xiangyuan Wang, Chang Xu et al.

Fusing Events and RGB images for object detection leverages the robustness of Event cameras in adverse environments and the rich semantic information provided by RGB cameras. However, two critical mismatches: low-latency Events \textit{vs.}~high-latency RGB frames; temporally sparse labels in training \textit{vs.}~continuous flow in inference, significantly hinder the high-frequency fusion-based object detection. To address these challenges, we propose the \textbf{F}requency-\textbf{A}daptive Low-Latency \textbf{O}bject \textbf{D}etector (FAOD). FAOD aligns low-frequency RGB frames with high-frequency Events through an Align Module, which reinforces cross-modal style and spatial proximity to address the Event-RGB Mismatch. We further propose a training strategy, Time Shift, which enforces the module to align the prediction from temporally shifted Event-RGB pairs and their original representation, that is, consistent with Event-aligned annotations. This strategy enables the network to use high-frequency Event data as the primary reference while treating low-frequency RGB images as supplementary information, retaining the low-latency nature of the Event stream toward high-frequency detection. Furthermore, we observe that these corrected Event-RGB pairs demonstrate better generalization from low training frequency to higher inference frequencies compared to using Event data alone. Extensive experiments on the PKU-DAVIS-SOD and DSEC-Detection datasets demonstrate that our FAOD achieves SOTA performance. Specifically, in the PKU-DAVIS-SOD Dataset, FAOD achieves 9.8 points improvement in terms of the mAP in fully paired Event-RGB data with only a quarter of the parameters compared to SODFormer, and even maintains robust performance (only a 3 points drop in mAP) under 80$\times$ Event-RGB frequency mismatch.

Hongwei Xu, Leijia Dai, Jianxing Fu et al.

We propose a real time deep learning framework for video-based facial expression capture. Our process uses a high-end facial capture pipeline based on FACEGOOD to capture facial expression. We train a convolutional neural network to produce high-quality continuous blendshape weight output from video training. Since this facial capture is fully automated, our system can drastically reduce the amount of labor involved in the development of modern narrative-driven video games or films involving realistic digital doubles of actors and potentially hours of animated dialogue per character. We demonstrate compelling animation inference in challenging areas such as eyes and lips.