STEP3-VL-10B Technical Report

We present STEP3-VL-10B, a lightweight open-source foundation model designed to redefine the trade-off between compact efficiency and frontier-level multimodal intelligence. STEP3-VL-10B is realized through two strategic shifts: first, a unified, fully unfrozen pre-training strategy on 1.2T multimodal tokens that integrates a language-aligned Perception Encoder with a Qwen3-8B decoder to establish intrinsic vision-language synergy; and second, a scaled post-training pipeline featuring over 1k iterations of reinforcement learning. Crucially, we implement Parallel Coordinated Reasoning (PaCoRe) to scale test-time compute, allocating resources to scalable perceptual reasoning that explores and synthesizes diverse visual hypotheses. Consequently, despite its compact 10B footprint, STEP3-VL-10B rivals or surpasses models 10$\times$-20$\times$ larger (e.g., GLM-4.6V-106B, Qwen3-VL-235B) and top-tier proprietary flagships like Gemini 2.5 Pro and Seed-1.5-VL. Delivering best-in-class performance, it records 92.2% on MMBench and 80.11% on MMMU, while excelling in complex reasoning with 94.43% on AIME2025 and 75.95% on MathVision. We release the full model suite to provide the community with a powerful, efficient, and reproducible baseline.

Distribution-Aligned Sequence Distillation for Superior Long-CoT Reasoning

In this report, we introduce DASD-4B-Thinking, a lightweight yet highly capable, fully open-source reasoning model. It achieves SOTA performance among open-source models of comparable scale across challenging benchmarks in mathematics, scientific reasoning, and code generation -- even outperforming several larger models. We begin by critically reexamining a widely adopted distillation paradigm in the community: SFT on teacher-generated responses, also known as sequence-level distillation. Although a series of recent works following this scheme have demonstrated remarkable efficiency and strong empirical performance, they are primarily grounded in the SFT perspective. Consequently, these approaches focus predominantly on designing heuristic rules for SFT data filtering, while largely overlooking the core principle of distillation itself -- enabling the student model to learn the teacher's full output distribution so as to inherit its generalization capability. Specifically, we identify three critical limitations in current practice: i) Inadequate representation of the teacher's sequence-level distribution; ii) Misalignment between the teacher's output distribution and the student's learning capacity; and iii) Exposure bias arising from teacher-forced training versus autoregressive inference. In summary, these shortcomings reflect a systemic absence of explicit teacher-student interaction throughout the distillation process, leaving the essence of distillation underexploited. To address these issues, we propose several methodological innovations that collectively form an enhanced sequence-level distillation training pipeline. Remarkably, DASD-4B-Thinking obtains competitive results using only 448K training samples -- an order of magnitude fewer than those employed by most existing open-source efforts. To support community research, we publicly release our models and the training dataset.

Puzzle it Out: Local-to-Global World Model for Offline Multi-Agent Reinforcement Learning

Offline multi-agent reinforcement learning (MARL) aims to solve cooperative decision-making problems in multi-agent systems using pre-collected datasets. Existing offline MARL methods primarily constrain training within the dataset distribution, resulting in overly conservative policies that struggle to generalize beyond the support of the data. While model-based approaches offer a promising solution by expanding the original dataset with synthetic data generated from a learned world model, the high dimensionality, non-stationarity, and complexity of multi-agent systems make it challenging to accurately estimate the transitions and reward functions in offline MARL. Given the difficulty of directly modeling joint dynamics, we propose a local-to-global (LOGO) world model, a novel framework that leverages local predictions-which are easier to estimate-to infer global state dynamics, thus improving prediction accuracy while implicitly capturing agent-wise dependencies. Using the trained world model, we generate synthetic data to augment the original dataset, expanding the effective state-action space. To ensure reliable policy learning, we further introduce an uncertainty-aware sampling mechanism that adaptively weights synthetic data by prediction uncertainty, reducing approximation error propagation to policies. In contrast to conventional ensemble-based methods, our approach requires only an additional encoder for uncertainty estimation, significantly reducing computational overhead while maintaining accuracy. Extensive experiments across 8 scenarios against 8 baselines demonstrate that our method surpasses state-of-the-art baselines on standard offline MARL benchmarks, establishing a new model-based baseline for generalizable offline multi-agent learning.

Mon3tr: Monocular 3D Telepresence with Pre-built Gaussian Avatars as Amortization

Immersive telepresence aims to transform human interaction in AR/VR applications by enabling lifelike full-body holographic representations for enhanced remote collaboration. However, existing systems rely on hardware-intensive multi-camera setups and demand high bandwidth for volumetric streaming, limiting their real-time performance on mobile devices. To overcome these challenges, we propose Mon3tr, a novel Monocular 3D telepresence framework that integrates 3D Gaussian splatting (3DGS) based parametric human modeling into telepresence for the first time. Mon3tr adopts an amortized computation strategy, dividing the process into a one-time offline multi-view reconstruction phase to build a user-specific avatar and a monocular online inference phase during live telepresence sessions. A single monocular RGB camera is used to capture body motions and facial expressions in real time to drive the 3DGS-based parametric human model, significantly reducing system complexity and cost. The extracted motion and appearance features are transmitted at < 0.2 Mbps over WebRTC's data channel, allowing robust adaptation to network fluctuations. On the receiver side, e.g., Meta Quest 3, we develop a lightweight 3DGS attribute deformation network to dynamically generate corrective 3DGS attribute adjustments on the pre-built avatar, synthesizing photorealistic motion and appearance at ~ 60 FPS. Extensive experiments demonstrate the state-of-the-art performance of our method, achieving a PSNR of > 28 dB for novel poses, an end-to-end latency of ~ 80 ms, and > 1000x bandwidth reduction compared to point-cloud streaming, while supporting real-time operation from monocular inputs across diverse scenarios. Our demos can be found at https://mon3tr3d.github.io.

Interleaved Tool-Call Reasoning for Protein Function Understanding

Recent advances in large language models (LLMs) have highlighted the effectiveness of chain-of-thought reasoning in symbolic domains such as mathematics and programming. However, our study shows that directly transferring such text-based reasoning paradigms to protein function understanding is ineffective: reinforcement learning mainly amplifies superficial keyword patterns while failing to introduce new biological knowledge, resulting in limited generalization. We argue that protein function prediction is a knowledge-intensive scientific task that fundamentally relies on external biological priors and computational tools rather than purely internal reasoning. To address this gap, we propose PFUA, a tool-augmented protein reasoning agent that unifies problem decomposition, tool invocation, and grounded answer generation. Instead of relying on long unconstrained reasoning traces, PFUA integrates domain-specific tools to produce verifiable intermediate evidence. Experiments on four benchmarks demonstrate that PFUA consistently outperforms text-only reasoning models with an average performance improvement of 103%.

ByteLoom: Weaving Geometry-Consistent Human-Object Interactions through Progressive Curriculum Learning

Human-object interaction (HOI) video generation has garnered increasing attention due to its promising applications in digital humans, e-commerce, advertising, and robotics imitation learning. However, existing methods face two critical limitations: (1) a lack of effective mechanisms to inject multi-view information of the object into the model, leading to poor cross-view consistency, and (2) heavy reliance on fine-grained hand mesh annotations for modeling interaction occlusions. To address these challenges, we introduce ByteLoom, a Diffusion Transformer (DiT)-based framework that generates realistic HOI videos with geometrically consistent object illustration, using simplified human conditioning and 3D object inputs. We first propose an RCM-cache mechanism that leverages Relative Coordinate Maps (RCM) as a universal representation to maintain object's geometry consistency and precisely control 6-DoF object transformations in the meantime. To compensate HOI dataset scarcity and leverage existing datasets, we further design a training curriculum that enhances model capabilities in a progressive style and relaxes the demand of hand mesh. Extensive experiments demonstrate that our method faithfully preserves human identity and the object's multi-view geometry, while maintaining smooth motion and object manipulation.

Where Did This Sentence Come From? Tracing Provenance in LLM Reasoning Distillation

Reasoning distillation has attracted increasing attention. It typically leverages a large teacher model to generate reasoning paths, which are then used to fine-tune a student model so that it mimics the teacher's behavior in training contexts. However, previous approaches have lacked a detailed analysis of the origins of the distilled model's capabilities. It remains unclear whether the student can maintain consistent behaviors with the teacher in novel test-time contexts, or whether it regresses to its original output patterns, raising concerns about the generalization of distillation models. To analyse this question, we introduce a cross-model Reasoning Distillation Provenance Tracing framework. For each action (e.g., a sentence) produced by the distilled model, we obtain the predictive probabilities assigned by the teacher, the original student, and the distilled model under the same context. By comparing these probabilities, we classify each action into different categories. By systematically disentangling the provenance of each action, we experimentally demonstrate that, in test-time contexts, the distilled model can indeed generate teacher-originated actions, which correlate with and plausibly explain observed performance on distilled model. Building on this analysis, we further propose a teacher-guided data selection method. Unlike prior approach that rely on heuristics, our method directly compares teacher-student divergences on the training data, providing a principled selection criterion. We validate the effectiveness of our approach across multiple representative teacher models and diverse student models. The results highlight the utility of our provenance-tracing framework and underscore its promise for reasoning distillation. We hope to share Reasoning Distillation Provenance Tracing and our insights into reasoning distillation with the community.

TrafficSimAgent: A Hierarchical Agent Framework for Autonomous Traffic Simulation with MCP Control

Traffic simulation is important for transportation optimization and policy making. While existing simulators such as SUMO and MATSim offer fully-featured platforms and utilities, users without too much knowledge about these platforms often face significant challenges when conducting experiments from scratch and applying them to their daily work. To solve this challenge, we propose TrafficSimAgent, an LLM-based agent framework that serves as an expert in experiment design and decision optimization for general-purpose traffic simulation tasks. The framework facilitates execution through cross-level collaboration among expert agents: high-level expert agents comprehend natural language instructions with high flexibility, plan the overall experiment workflow, and invoke corresponding MCP-compatible tools on demand; meanwhile, low-level expert agents select optimal action plans for fundamental elements based on real-time traffic conditions. Extensive experiments across multiple scenarios show that TrafficSimAgent effectively executes simulations under various conditions and consistently produces reasonable outcomes even when user instructions are ambiguous. Besides, the carefully designed expert-level autonomous decision-driven optimization in TrafficSimAgent yields superior performance when compared with other systems and SOTA LLM based methods.

Evolutionary Neural Architecture Search with Dual Contrastive Learning

Evolutionary Neural Architecture Search (ENAS) has gained attention for automatically designing neural network architectures. Recent studies use a neural predictor to guide the process, but the high computational costs of gathering training data -- since each label requires fully training an architecture -- make achieving a high-precision predictor with { limited compute budget (i.e., a capped number of fully trained architecture-label pairs)} crucial for ENAS success. This paper introduces ENAS with Dual Contrastive Learning (DCL-ENAS), a novel method that employs two stages of contrastive learning to train the neural predictor. In the first stage, contrastive self-supervised learning is used to learn meaningful representations from neural architectures without requiring labels. In the second stage, fine-tuning with contrastive learning is performed to accurately predict the relative performance of different architectures rather than their absolute performance, which is sufficient to guide the evolutionary search. Across NASBench-101 and NASBench-201, DCL-ENAS achieves the highest validation accuracy, surpassing the strongest published baselines by 0.05\% (ImageNet16-120) to 0.39\% (NASBench-101). On a real-world ECG arrhythmia classification task, DCL-ENAS improves performance by approximately 2.5 percentage points over a manually designed, non-NAS model obtained via random search, while requiring only 7.7 GPU-days.

UrbanV2X: A Multisensory Vehicle-Infrastructure Dataset for Cooperative Navigation in Urban Areas

Due to the limitations of a single autonomous vehicle, Cellular Vehicle-to-Everything (C-V2X) technology opens a new window for achieving fully autonomous driving through sensor information sharing. However, real-world datasets supporting vehicle-infrastructure cooperative navigation in complex urban environments remain rare. To address this gap, we present UrbanV2X, a comprehensive multisensory dataset collected from vehicles and roadside infrastructure in the Hong Kong C-V2X testbed, designed to support research on smart mobility applications in dense urban areas. Our onboard platform provides synchronized data from multiple industrial cameras, LiDARs, 4D radar, ultra-wideband (UWB), IMU, and high-precision GNSS-RTK/INS navigation systems. Meanwhile, our roadside infrastructure provides LiDAR, GNSS, and UWB measurements. The entire vehicle-infrastructure platform is synchronized using the Precision Time Protocol (PTP), with sensor calibration data provided. We also benchmark various navigation algorithms to evaluate the collected cooperative data. The dataset is publicly available at https://polyu-taslab.github.io/UrbanV2X/.