FedMOA: Federated GRPO for Personalized Reasoning LLMs under Heterogeneous Rewards

Group Relative Policy Optimization (GRPO) has recently emerged as an effective approach for improving the reasoning capabilities of large language models through online multi-objective reinforcement learning. While personalization on private data is increasingly vital, traditional Reinforcement Learning (RL) alignment is often memory-prohibitive for on-device federated learning due to the overhead of maintaining a separate critic network. GRPO's critic-free architecture enables feasible on-device training, yet transitioning to a federated setting introduces systemic challenges: heterogeneous reward definitions, imbalanced multi-objective optimization, and high training costs. We propose FedMOA, a federated GRPO framework for multi-objective alignment under heterogeneous rewards. FedMOA stabilizes local training through an online adaptive weighting mechanism via hypergradient descent, which prioritizes primary reasoning as auxiliary objectives saturate. On the server side, it utilizes a task- and accuracy-aware aggregation strategy to prioritize high-quality updates. Experiments on mathematical reasoning and code generation benchmarks demonstrate that FedMOA consistently outperforms federated averaging, achieving accuracy gains of up to 2.2% while improving global performance, personalization, and multi-objective balance.

Towards Building Non-Fine-Tunable Foundation Models

Open-sourcing foundation models (FMs) enables broad reuse but also exposes model trainers to economic and safety risks from unrestricted downstream fine-tuning. We address this problem by building non-fine-tunable foundation models: models that remain broadly usable in their released form while yielding limited adaptation gains under task-agnostic unauthorized fine-tuning. We propose Private Mask Pre-Training (PMP), a pre-training framework that concentrates representation learning into a sparse subnetwork identified early in training. The binary mask defining this subnetwork is kept private, and only the final dense weights are released. This forces unauthorized fine-tuning without access to the mask to update parameters misaligned with pretraining subspace, inducing an intrinsic mismatch between the fine-tuning objective and the pre-training geometry. We provide theoretical analysis showing that this mismatch destabilizes gradient-based adaptation and bounds fine-tuning gains. Empirical results on large language models demonstrating that PMP preserves base model performance while consistently degrading unauthorized fine-tuning across a wide range of downstream tasks, with the strength of non-fine-tunability controlled by the mask ratio.

Causal-Driven Feature Evaluation for Cross-Domain Image Classification

Out-of-distribution (OOD) generalization remains a fundamental challenge in real-world classification, where test distributions often differ substantially from training data. Most existing approaches pursue domain-invariant representations, implicitly assuming that invariance implies reliability. However, features that are invariant across domains are not necessarily causally effective for prediction. In this work, we revisit OOD classification from a causal perspective and propose to evaluate learned representations based on their necessity and sufficiency under distribution shift. We introduce an explicit segment-level framework that directly measures causal effectiveness across domains, providing a more faithful criterion than invariance alone. Experiments on multi-domain benchmarks demonstrate consistent improvements in OOD performance, particularly under challenging domain shifts, highlighting the value of causal evaluation for robust generalization.

Decoupling Return-to-Go for Efficient Decision Transformer

The Decision Transformer (DT) has established a powerful sequence modeling approach to offline reinforcement learning. It conditions its action predictions on Return-to-Go (RTG), using it both to distinguish trajectory quality during training and to guide action generation at inference. In this work, we identify a critical redundancy in this design: feeding the entire sequence of RTGs into the Transformer is theoretically unnecessary, as only the most recent RTG affects action prediction. We show that this redundancy can impair DT's performance through experiments. To resolve this, we propose the Decoupled DT (DDT). DDT simplifies the architecture by processing only observation and action sequences through the Transformer, using the latest RTG to guide the action prediction. This streamlined approach not only improves performance but also reduces computational cost. Our experiments show that DDT significantly outperforms DT and establishes competitive performance against state-of-the-art DT variants across multiple offline RL tasks.

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.

Zeros can be Informative: Masked Binary U-Net for Image Segmentation on Tensor Cores

Real-time image segmentation is a key enabler for AR/VR, robotics, drones, and autonomous systems, where tight accuracy, latency, and energy budgets must be met on resource-constrained edge devices. While U-Net offers a favorable balance of accuracy and efficiency compared to large transformer-based models, achieving real-time performance on high-resolution input remains challenging due to compute, memory, and power limits. Extreme quantization, particularly binary networks, is appealing for its hardware-friendly operations. However, two obstacles limit practicality: (1) severe accuracy degradation, and (2) a lack of end-to-end implementations that deliver efficiency on general-purpose GPUs. We make two empirical observations that guide our design. (1) An explicit zero state is essential: training with zero masking to binary U-Net weights yields noticeable sparsity. (2) Quantization sensitivity is uniform across layers. Motivated by these findings, we introduce Masked Binary U-Net (MBU-Net), obtained through a cost-aware masking strategy that prioritizes masking where it yields the highest accuracy-per-cost, reconciling accuracy with near-binary efficiency. To realize these gains in practice, we develop a GPU execution framework that maps MBU-Net to Tensor Cores via a subtractive bit-encoding scheme, efficiently implementing masked binary weights with binary activations. This design leverages native binary Tensor Core BMMA instructions, enabling high throughput and energy savings on widely available GPUs. Across 3 segmentation benchmarks, MBU-Net attains near full-precision accuracy (3% average drop) while delivering 2.04x speedup and 3.54x energy reductions over a 16-bit floating point U-Net.

Policy-Conditioned Policies for Multi-Agent Task Solving

In multi-agent tasks, the central challenge lies in the dynamic adaptation of strategies. However, directly conditioning on opponents' strategies is intractable in the prevalent deep reinforcement learning paradigm due to a fundamental ``representational bottleneck'': neural policies are opaque, high-dimensional parameter vectors that are incomprehensible to other agents. In this work, we propose a paradigm shift that bridges this gap by representing policies as human-interpretable source code and utilizing Large Language Models (LLMs) as approximate interpreters. This programmatic representation allows us to operationalize the game-theoretic concept of \textit{Program Equilibrium}. We reformulate the learning problem by utilizing LLMs to perform optimization directly in the space of programmatic policies. The LLM functions as a point-wise best-response operator that iteratively synthesizes and refines the ego agent's policy code to respond to the opponent's strategy. We formalize this process as \textit{Programmatic Iterated Best Response (PIBR)}, an algorithm where the policy code is optimized by textual gradients, using structured feedback derived from game utility and runtime unit tests. We demonstrate that this approach effectively solves several standard coordination matrix games and a cooperative Level-Based Foraging environment.

Making Large Language Models Efficient Dense Retrievers

Recent work has shown that directly fine-tuning large language models (LLMs) for dense retrieval yields strong performance, but their substantial parameter counts make them computationally inefficient. While prior studies have revealed significant layer redundancy in LLMs for generative tasks, it remains unclear whether similar redundancy exists when these models are adapted for retrieval tasks, which require encoding entire sequences into fixed representations rather than generating tokens iteratively. To this end, we conduct a comprehensive analysis of layer redundancy in LLM-based dense retrievers. We find that, in contrast to generative settings, MLP layers are substantially more prunable, while attention layers remain critical for semantic aggregation. Building on this insight, we propose EffiR, a framework for developing efficient retrievers that performs large-scale MLP compression through a coarse-to-fine strategy (coarse-grained depth reduction followed by fine-grained width reduction), combined with retrieval-specific fine-tuning. Across diverse BEIR datasets and LLM backbones, EffiR achieves substantial reductions in model size and inference cost while preserving the performance of full-size models.

CoordAR: One-Reference 6D Pose Estimation of Novel Objects via Autoregressive Coordinate Map Generation

Object 6D pose estimation, a crucial task for robotics and augmented reality applications, becomes particularly challenging when dealing with novel objects whose 3D models are not readily available. To reduce dependency on 3D models, recent studies have explored one-reference-based pose estimation, which requires only a single reference view instead of a complete 3D model. However, existing methods that rely on real-valued coordinate regression suffer from limited global consistency due to the local nature of convolutional architectures and face challenges in symmetric or occluded scenarios owing to a lack of uncertainty modeling. We present CoordAR, a novel autoregressive framework for one-reference 6D pose estimation of unseen objects. CoordAR formulates 3D-3D correspondences between the reference and query views as a map of discrete tokens, which is obtained in an autoregressive and probabilistic manner. To enable accurate correspondence regression, CoordAR introduces 1) a novel coordinate map tokenization that enables probabilistic prediction over discretized 3D space; 2) a modality-decoupled encoding strategy that separately encodes RGB appearance and coordinate cues; and 3) an autoregressive transformer decoder conditioned on both position-aligned query features and the partially generated token sequence. With these novel mechanisms, CoordAR significantly outperforms existing methods on multiple benchmarks and demonstrates strong robustness to symmetry, occlusion, and other challenges in real-world tests.

Is It Truly Necessary to Process and Fit Minutes-Long Reference Videos for Personalized Talking Face Generation?

Talking Face Generation (TFG) aims to produce realistic and dynamic talking portraits, with broad applications in fields such as digital education, film and television production, e-commerce live streaming, and other related areas. Currently, TFG methods based on Neural Radiated Field (NeRF) or 3D Gaussian sputtering (3DGS) are received widespread attention. They learn and store personalized features from reference videos of each target individual to generate realistic speaking videos. To ensure models can capture sufficient 3D information and successfully learns the lip-audio mapping, previous studies usually require meticulous processing and fitting several minutes of reference video, which always takes hours. The computational burden of processing and fitting long reference videos severely limits the practical application value of these methods.However, is it really necessary to fit such minutes of reference video? Our exploratory case studies show that using some informative reference video segments of just a few seconds can achieve performance comparable to or even better than the full reference video. This indicates that video informative quality is much more important than its length. Inspired by this observation, we propose the ISExplore (short for Informative Segment Explore), a simple-yet-effective segment selection strategy that automatically identifies the informative 5-second reference video segment based on three key data quality dimensions: audio feature diversity, lip movement amplitude, and number of camera views. Extensive experiments demonstrate that our approach increases data processing and training speed by more than 5x for NeRF and 3DGS methods, while maintaining high-fidelity output. Project resources are available at xx.