SEDM: Scalable Self-Evolving Distributed Memory for Agents

Long-term multi-agent systems inevitably generate vast amounts of trajectories and historical interactions, which makes efficient memory management essential for both performance and scalability. Existing methods typically depend on vector retrieval and hierarchical storage, yet they are prone to noise accumulation, uncontrolled memory expansion, and limited generalization across domains. To address these challenges, we present SEDM, Self-Evolving Distributed Memory, a verifiable and adaptive framework that transforms memory from a passive repository into an active, self-optimizing component. SEDM integrates verifiable write admission based on reproducible replay, a self-scheduling memory controller that dynamically ranks and consolidates entries according to empirical utility, and cross-domain knowledge diffusion that abstracts reusable insights to support transfer across heterogeneous tasks. Evaluations on benchmark datasets demonstrate that SEDM improves reasoning accuracy while reducing token overhead compared with strong memory baselines, and further enables knowledge distilled from fact verification to enhance multi-hop reasoning. The results highlight SEDM as a scalable and sustainable memory mechanism for open-ended multi-agent collaboration. The code will be released in the later stage of this project.

Towards Affordance-Aware Robotic Dexterous Grasping with Human-like Priors

A dexterous hand capable of generalizable grasping objects is fundamental for the development of general-purpose embodied AI. However, previous methods focus narrowly on low-level grasp stability metrics, neglecting affordance-aware positioning and human-like poses which are crucial for downstream manipulation. To address these limitations, we propose AffordDex, a novel framework with two-stage training that learns a universal grasping policy with an inherent understanding of both motion priors and object affordances. In the first stage, a trajectory imitator is pre-trained on a large corpus of human hand motions to instill a strong prior for natural movement. In the second stage, a residual module is trained to adapt these general human-like motions to specific object instances. This refinement is critically guided by two components: our Negative Affordance-aware Segmentation (NAA) module, which identifies functionally inappropriate contact regions, and a privileged teacher-student distillation process that ensures the final vision-based policy is highly successful. Extensive experiments demonstrate that AffordDex not only achieves universal dexterous grasping but also remains remarkably human-like in posture and functionally appropriate in contact location. As a result, AffordDex significantly outperforms state-of-the-art baselines across seen objects, unseen instances, and even entirely novel categories.

Make Your MoVe: Make Your 3D Contents by Adapting Multi-View Diffusion Models to External Editing

As 3D generation techniques continue to flourish, the demand for generating personalized content is rapidly rising. Users increasingly seek to apply various editing methods to polish generated 3D content, aiming to enhance its color, style, and lighting without compromising the underlying geometry. However, most existing editing tools focus on the 2D domain, and directly feeding their results into 3D generation methods (like multi-view diffusion models) will introduce information loss, degrading the quality of the final 3D assets. In this paper, we propose a tuning-free, plug-and-play scheme that aligns edited assets with their original geometry in a single inference run. Central to our approach is a geometry preservation module that guides the edited multi-view generation with original input normal latents. Besides, an injection switcher is proposed to deliberately control the supervision extent of the original normals, ensuring the alignment between the edited color and normal views. Extensive experiments show that our method consistently improves both the multi-view consistency and mesh quality of edited 3D assets, across multiple combinations of multi-view diffusion models and editing methods.

Decoder-Hybrid-Decoder Architecture for Efficient Reasoning with Long Generation

Recent advances in language modeling have demonstrated the effectiveness of State Space Models (SSMs) for efficient sequence modeling. While hybrid architectures such as Samba and the decoder-decoder architecture, YOCO, have shown promising performance gains over Transformers, prior works have not investigated the efficiency potential of representation sharing between SSM layers. In this paper, we introduce the Gated Memory Unit (GMU), a simple yet effective mechanism for efficient memory sharing across layers. We apply it to create SambaY, a decoder-hybrid-decoder architecture that incorporates GMUs in the cross-decoder to share memory readout states from a Samba-based self-decoder. SambaY significantly enhances decoding efficiency, preserves linear pre-filling time complexity, and boosts long-context performance, all while eliminating the need for explicit positional encoding. Through extensive scaling experiments, we demonstrate that our model exhibits a significantly lower irreducible loss compared to a strong YOCO baseline, indicating superior performance scalability under large-scale compute regimes. Our largest model enhanced with Differential Attention, Phi4-mini-Flash-Reasoning, achieves significantly better performance than Phi4-mini-Reasoning on reasoning tasks such as Math500, AIME24/25, and GPQA Diamond without any reinforcement learning, while delivering up to 10x higher decoding throughput on 2K-length prompts with 32K generation length under the vLLM inference framework. We release our training codebase on open-source data at https://github.com/microsoft/ArchScale.

Consistent World Models via Foresight Diffusion

Diffusion and flow-based models have enabled significant progress in generation tasks across various modalities and have recently found applications in world modeling. However, unlike typical generation tasks that encourage sample diversity, world models entail different sources of uncertainty and require consistent samples aligned with the ground-truth trajectory, which is a limitation we empirically observe in diffusion models. We argue that a key bottleneck in learning consistent diffusion-based world models lies in the suboptimal predictive ability, which we attribute to the entanglement of condition understanding and target denoising within shared architectures and co-training schemes. To address this, we propose Foresight Diffusion (ForeDiff), a diffusion-based world modeling framework that enhances consistency by decoupling condition understanding from target denoising. ForeDiff incorporates a separate deterministic predictive stream to process conditioning inputs independently of the denoising stream, and further leverages a pretrained predictor to extract informative representations that guide generation. Extensive experiments on robot video prediction and scientific spatiotemporal forecasting show that ForeDiff improves both predictive accuracy and sample consistency over strong baselines, offering a promising direction for diffusion-based world models.

TeleOpBench: A Simulator-Centric Benchmark for Dual-Arm Dexterous Teleoperation

Teleoperation is a cornerstone of embodied-robot learning, and bimanual dexterous teleoperation in particular provides rich demonstrations that are difficult to obtain with fully autonomous systems. While recent studies have proposed diverse hardware pipelines-ranging from inertial motion-capture gloves to exoskeletons and vision-based interfaces-there is still no unified benchmark that enables fair, reproducible comparison of these systems. In this paper, we introduce TeleOpBench, a simulator-centric benchmark tailored to bimanual dexterous teleoperation. TeleOpBench contains 30 high-fidelity task environments that span pick-and-place, tool use, and collaborative manipulation, covering a broad spectrum of kinematic and force-interaction difficulty. Within this benchmark we implement four representative teleoperation modalities-(i) MoCap, (ii) VR device, (iii) arm-hand exoskeletons, and (iv) monocular vision tracking-and evaluate them with a common protocol and metric suite. To validate that performance in simulation is predictive of real-world behavior, we conduct mirrored experiments on a physical dual-arm platform equipped with two 6-DoF dexterous hands. Across 10 held-out tasks we observe a strong correlation between simulator and hardware performance, confirming the external validity of TeleOpBench. TeleOpBench establishes a common yardstick for teleoperation research and provides an extensible platform for future algorithmic and hardware innovation.

Bridging Quantized Artificial Neural Networks and Neuromorphic Hardware

Neuromorphic hardware has been proposed and also been produced for decades. One of the main goals of this hardware is to leverage distributed computing and event-driven circuit design and achieve power-efficient AI system. The name ``neuromorphic'' is derived from its spiking and local computational nature, which mimics the fundamental activity of an animal's nervous system. Neurons as well as distributed computing cores of neuromorphic hardware use single bit data, called a spike, for inter-communication. To construct a spiking model for neuromorphic hardware, the conventional approach is to build spiking neural networks (SNNs). SNN replaces the nonlinearity part of artificial neural networks (ANNs) in the realm of deep learning with spiking neurons, where the spiking neuron mimic the basic behavior of bio-neurons. However, there is still a performance gap between SNN and ANN counterpart. In this paper, we explore a new path from ANN to neuromorphic hardware. The SDANN framework is proposed to directly implement quantized ANN on hardware, eliminating the need for tuning the trainable parameters or any performance degradation. With the power of quantized ANN, our SDANN provides a lower bound of the functionality of neuromorphic hardware. Meanwhile, we have also proposed scaling methods in case of the limited bit-width support in hardware. Spike sparsity methods are also provided for further energy optimization. Experiments on various tasks demonstrate the usefulness of our SDANN framework. Beyond toy examples and software implementation, we successfully deploy the spiking models of SDANN on real neuromorphic hardware, demonstrating the feasibility of the SDANN framework.

Phi-4-Mini-Reasoning: Exploring the Limits of Small Reasoning Language Models in Math

Chain-of-Thought (CoT) significantly enhances formal reasoning capabilities in Large Language Models (LLMs) by training them to explicitly generate intermediate reasoning steps. While LLMs readily benefit from such techniques, improving reasoning in Small Language Models (SLMs) remains challenging due to their limited model capacity. Recent work by Deepseek-R1 demonstrates that distillation from LLM-generated synthetic data can substantially improve the reasoning ability of SLM. However, the detailed modeling recipe is not disclosed. In this work, we present a systematic training recipe for SLMs that consists of four steps: (1) large-scale mid-training on diverse distilled long-CoT data, (2) supervised fine-tuning on high-quality long-CoT data, (3) Rollout DPO leveraging a carefully curated preference dataset, and (4) Reinforcement Learning (RL) with Verifiable Reward. We apply our method on Phi-4-Mini, a compact 3.8B-parameter model. The resulting Phi-4-Mini-Reasoning model exceeds, on math reasoning tasks, much larger reasoning models, e.g., outperforming DeepSeek-R1-Distill-Qwen-7B by 3.2 points and DeepSeek-R1-Distill-Llama-8B by 7.7 points on Math-500. Our results validate that a carefully designed training recipe, with large-scale high-quality CoT data, is effective to unlock strong reasoning capabilities even in resource-constrained small models.

Decoupled Global-Local Alignment for Improving Compositional Understanding

Contrastive Language-Image Pre-training (CLIP) has achieved success on multiple downstream tasks by aligning image and text modalities. However, the nature of global contrastive learning limits CLIP's ability to comprehend compositional concepts, such as relations and attributes. Although recent studies employ global hard negative samples to improve compositional understanding, these methods significantly compromise the model's inherent general capabilities by forcibly distancing textual negative samples from images in the embedding space. To overcome this limitation, we introduce a Decoupled Global-Local Alignment (DeGLA) framework that improves compositional understanding while substantially mitigating losses in general capabilities. To optimize the retention of the model's inherent capabilities, we incorporate a self-distillation mechanism within the global alignment process, aligning the learnable image-text encoder with a frozen teacher model derived from an exponential moving average. Under the constraint of self-distillation, it effectively mitigates the catastrophic forgetting of pretrained knowledge during fine-tuning. To improve compositional understanding, we first leverage the in-context learning capability of Large Language Models (LLMs) to construct about 2M high-quality negative captions across five types. Subsequently, we propose the Image-Grounded Contrast (IGC) loss and Text-Grounded Contrast (TGC) loss to enhance vision-language compositionally. Extensive experimental results demonstrate the effectiveness of the DeGLA framework. Compared to previous state-of-the-art methods, DeGLA achieves an average enhancement of 3.5% across the VALSE, SugarCrepe, and ARO benchmarks. Concurrently, it obtains an average performance improvement of 13.0% on zero-shot classification tasks across eleven datasets. Our code will be released at https://github.com/xiaoxing2001/DeGLA

An Optimal Discriminator Weighted Imitation Perspective for Reinforcement Learning

We introduce Iterative Dual Reinforcement Learning (IDRL), a new method that takes an optimal discriminator-weighted imitation view of solving RL. Our method is motivated by a simple experiment in which we find training a discriminator using the offline dataset plus an additional expert dataset and then performing discriminator-weighted behavior cloning gives strong results on various types of datasets. That optimal discriminator weight is quite similar to the learned visitation distribution ratio in Dual-RL, however, we find that current Dual-RL methods do not correctly estimate that ratio. In IDRL, we propose a correction method to iteratively approach the optimal visitation distribution ratio in the offline dataset given no addtional expert dataset. During each iteration, IDRL removes zero-weight suboptimal transitions using the learned ratio from the previous iteration and runs Dual-RL on the remaining subdataset. This can be seen as replacing the behavior visitation distribution with the optimized visitation distribution from the previous iteration, which theoretically gives a curriculum of improved visitation distribution ratios that are closer to the optimal discriminator weight. We verify the effectiveness of IDRL on various kinds of offline datasets, including D4RL datasets and more realistic corrupted demonstrations. IDRL beats strong Primal-RL and Dual-RL baselines in terms of both performance and stability, on all datasets.