JoyAI-LLM Flash: Advancing Mid-Scale LLMs with Token Efficiency

We introduce JoyAI-LLM Flash, an efficient Mixture-of-Experts (MoE) language model designed to redefine the trade-off between strong performance and token efficiency in the sub-50B parameter regime. JoyAI-LLM Flash is pretrained on a massive corpus of 20 trillion tokens and further optimized through a rigorous post-training pipeline, including supervised fine-tuning (SFT), Direct Preference Optimization (DPO), and large-scale reinforcement learning (RL) across diverse environments. To improve token efficiency, JoyAI-LLM Flash strategically balances \emph{thinking} and \emph{non-thinking} cognitive modes and introduces FiberPO, a novel RL algorithm inspired by fibration theory that decomposes trust-region maintenance into global and local components, providing unified multi-scale stability control for LLM policy optimization. To enhance architectural sparsity, the model comprises 48B total parameters while activating only 2.7B parameters per forward pass, achieving a substantially higher sparsity ratio than contemporary industry leading models of comparable scale. To further improve inference throughput, we adopt a joint training-inference co-design that incorporates dense Multi-Token Prediction (MTP) and Quantization-Aware Training (QAT). We release the checkpoints for both JoyAI-LLM-48B-A3B Base and its post-trained variants on Hugging Face to support the open-source community.

COMPASS: Complete Multimodal Fusion via Proxy Tokens and Shared Spaces for Ubiquitous Sensing

Missing modalities remain a major challenge for multimodal sensing, because most existing methods adapt the fusion process to the observed subset by dropping absent branches, using subset-specific fusion, or reconstructing missing features. As a result, the fusion head often receives an input structure different from the one seen during training, leading to incomplete fusion and degraded cross-modal interaction. We propose COMPASS, a missing-modality fusion framework built on the principle of fusion completeness: the fusion head always receives a fixed N-slot multimodal input, with one token per modality slot. For each missing modality, COMPASS synthesizes a target-specific proxy token from the observed modalities using pairwise source-to-target generators in a shared latent space, and aggregates them into a single replacement token. To make these proxies both representation-compatible and task-informative, we combine proxy alignment, shared-space regularization, and per-proxy discriminative supervision. Experiments on XRF55, MM-Fi, and OctoNet under diverse single- and multiple-missing settings show that COMPASS outperforms prior methods on the large majority of scenarios. Our results suggest that preserving a modality-complete fusion interface is a simple and effective design principle for robust multimodal sensing.

Scheduling LLM Inference with Uncertainty-Aware Output Length Predictions

To schedule LLM inference, the \textit{shortest job first} (SJF) principle is favorable by prioritizing requests with short output lengths to avoid head-of-line (HOL) blocking. Existing methods usually predict a single output length for each request to facilitate scheduling. We argue that such a \textit{point estimate} does not match the \textit{stochastic} decoding process of LLM inference, where output length is \textit{uncertain} by nature and determined by when the end-of-sequence (EOS) token is sampled. Hence, the output length of each request should be fitted with a distribution rather than a single value. With an in-depth analysis of empirical data and the stochastic decoding process, we observe that output length follows a heavy-tailed distribution and can be fitted with the log-t distribution. On this basis, we propose a simple metric called Tail Inflated Expectation (TIE) to replace the output length in SJF scheduling, which adjusts the expectation of a log-t distribution with its tail probabilities to account for the risk that a request generates long outputs. To evaluate our TIE scheduler, we compare it with three strong baselines, and the results show that TIE reduces the per-token latency by $2.31\times$ for online inference and improves throughput by $1.42\times$ for offline data generation.

Conditional Polarization Guidance for Camouflaged Object Detection

Camouflaged object detection (COD) aims to identify targets that are highly blended with their backgrounds. Recent works have shown that the optical characteristics of polarization cues play a significant role in improving camouflaged object detection. However, most existing polarization-based approaches depend on complex visual encoders and fusion mechanisms, leading to increased model complexity and computational overhead, while failing to fully explore how polarization can explicitly guide hierarchical RGB representation learning. To address these limitations, we propose CPGNet, an asymmetric RGB-polarization framework that introduces a conditional polarization guidance mechanism to explicitly regulate RGB feature learning for camouflaged object detection. Specifically, we design a lightweight polarization interaction module that jointly models these complementary cues and generates reliable polarization guidance in a unified manner. Unlike conventional feature fusion strategies, the proposed conditional guidance mechanism dynamically modulates RGB features using polarization priors, enabling the network to focus on subtle discrepancies between camouflaged objects and their backgrounds. Furthermore, we introduce a polarization edge-guided frequency refinement strategy that enhances high-frequency components under polarization constraints, effectively breaking camouflage patterns. Finally, we develop an iterative feedback decoder to perform coarse-to-fine feature calibration and progressively refine camouflage prediction. Extensive experiments on polarization datasets across multiple tasks, along with evaluations on non-polarization datasets, demonstrate that CPGNet consistently outperforms state-of-the-art methods.

FED-Bench: A Cross-Granular Benchmark for Disentangled Evaluation of Facial Expression Editing

Facial expression image editing requires fine-grained control to strictly preserve human identity and background while precisely manipulating expression. However, existing editing benchmarks primarily focus on general scenarios, lacking high-quality facial images and corresponding editing instructions. Furthermore, current evaluation metrics exhibit systemic biases in this task, often favoring lazy editing or overfit editing. To bridge these gaps, we propose FED-Bench, a comprehensive benchmark featuring rigorous testing and an accurate evaluation suite. First, we carefully construct a benchmark of 747 triplets through a cascaded and scalable pipeline, each comprising an original image, an editing instruction, and a ground-truth image for precise evaluation. Second, we introduce FED-Score, a cross-granularity evaluation protocol that disentangles assessment into three dimensions: Alignment for verifying instruction following, Fidelity for testing image quality and identity preservation, and Relative Expression Gain for quantifying the magnitude of expression changes, effectively mitigating the aforementioned evaluation biases. Third, we benchmark 18 image editing models, revealing that current approaches struggle to simultaneously achieve high fidelity and accurate expression manipulation, with fine-grained instruction following identified as the primary bottleneck. Finally, leveraging the scalable characteristic of introduced benchmark engine, we provide a 20k+ in-the-wild facial training set and demonstrate its effectiveness by fine-tuning a baseline model that achieves significant performance gains. Our benchmark and related code will be made publicly open soon.

ManipArena: Comprehensive Real-world Evaluation of Reasoning-Oriented Generalist Robot Manipulation

Vision-Language-Action (VLA) models and world models have recently emerged as promising paradigms for general-purpose robotic intelligence, yet their progress is hindered by the lack of reliable evaluation protocols that reflect real-world deployment. Existing benchmarks are largely simulator-centric, which provide controllability but fail to capture the reality gap caused by perception noise, complex contact dynamics, hardware constraints, and system latency. Moreover, fragmented real-world evaluations across different robot platforms prevent fair and reproducible comparison. To address these challenges, we introduce ManipArena, a standardized evaluation framework designed to bridge simulation and real-world execution. ManipArena comprises 20 diverse tasks across 10,812 expert trajectories emphasizing reasoning-oriented manipulation tasks requiring semantic and spatial reasoning, supports multi-level generalization through controlled out-of-distribution settings, and incorporates long-horizon mobile manipulation beyond tabletop scenarios. The framework further provides rich sensory diagnostics, including low-level motor signals, and synchronized real-to-sim environments constructed via high-quality 3D scanning. Together, these features enable fair, realistic, and reproducible evaluation for both VLA and world model approaches, providing a scalable foundation for diagnosing and advancing embodied intelligence systems.

KAT-Coder-V2 Technical Report

We present KAT-Coder-V2, an agentic coding model developed by the KwaiKAT team at Kuaishou. KAT-Coder-V2 adopts a "Specialize-then-Unify" paradigm that decomposes agentic coding into five expert domains - SWE, WebCoding, Terminal, WebSearch, and General - each undergoing independent supervised fine-tuning and reinforcement learning, before being consolidated into a single model via on-policy distillation. We develop KwaiEnv, a modular infrastructure sustaining tens of thousands of concurrent sandbox instances, and scale RL training along task complexity, intent alignment, and scaffold generalization. We further propose MCLA for stabilizing MoE RL training and Tree Training for eliminating redundant computation over tree-structured trajectories with up to 6.2x speedup. KAT-Coder-V2 achieves 79.6% on SWE-bench Verified (vs. Claude Opus 4.6 at 80.8%), 88.7 on PinchBench (surpassing GLM-5 and MiniMax M2.7), ranks first across all three frontend aesthetics scenarios, and maintains strong generalist scores on Terminal-Bench Hard (46.8) and tau^2-Bench (93.9). Our model is publicly available at https://streamlake.com/product/kat-coder.

S2D2: Fast Decoding for Diffusion LLMs via Training-Free Self-Speculation

Block-diffusion language models offer a promising path toward faster-than-autoregressive generation by combining block-wise autoregressive decoding with within-block parallel denoising. However, in the few-step regime needed for practical acceleration, standard confidence-thresholded decoding is often brittle: aggressive thresholds hurt quality, while conservative thresholds require unnecessary denoising steps. Existing approaches that address this issue either require additional training or incur extra test-time compute. We present S2D2, a training-free self-speculative decoding framework for block-diffusion language models. Our key observation is that a block-diffusion model becomes autoregressive when the block size is reduced to one, allowing the same pretrained model to act as both drafter and verifier. S2D2 inserts a speculative verification step into standard block-diffusion decoding and uses lightweight routing policies to decide when verification is worth its cost. This yields a hybrid decoding trajectory in which diffusion proposes tokens in parallel, while the autoregressive mode acts as a local sequence-level critic. Across three mainstream block-diffusion families, S2D2 consistently improves the accuracy-speed tradeoff over strong confidence-thresholding baselines. On SDAR, we observe up to $4.7\times$ speedup over autoregressive decoding, and up to $1.57\times$ over a tuned dynamic decoding baseline while improving accuracy by up to $4.5$ points. On LLaDA2.1-Mini, S2D2 remains complementary to built-in self-correction, including a conservative setting where it is $4.4\times$ faster than the static baseline with slightly higher accuracy.

DIET: Learning to Distill Dataset Continually for Recommender Systems

Modern deep recommender models are trained under a continual learning paradigm, relying on massive and continuously growing streaming behavioral logs. In large-scale platforms, retraining models on full historical data for architecture comparison or iteration is prohibitively expensive, severely slowing down model development. This challenge calls for data-efficient approaches that can faithfully approximate full-data training behavior without repeatedly processing the entire evolving data stream. We formulate this problem as \emph{streaming dataset distillation for recommender systems} and propose \textbf{DIET}, a unified framework that maintains a compact distilled dataset which evolves alongside streaming data while preserving training-critical signals. Unlike existing dataset distillation methods that construct a static distilled set, DIET models distilled data as an evolving training memory and updates it in a stage-wise manner to remain aligned with long-term training dynamics. DIET enables effective continual distillation through principled initialization from influential samples and selective updates guided by influence-aware memory addressing within a bi-level optimization framework. Experiments on large-scale recommendation benchmarks demonstrate that DIET compresses training data to as little as \textbf{1-2\%} of the original size while preserving performance trends consistent with full-data training, reducing model iteration cost by up to \textbf{60$\times$}. Moreover, the distilled datasets produced by DIET generalize well across different model architectures, highlighting streaming dataset distillation as a scalable and reusable data foundation for recommender system development.

From Synthetic to Native: Benchmarking Multilingual Intent Classification in Logistics Customer Service

Multilingual intent classification is central to customer-service systems on global logistics platforms, where models must process noisy user queries across languages and hierarchical label spaces. Yet most existing multilingual benchmarks rely on machine-translated text, which is typically cleaner and more standardized than native customer requests and can therefore overestimate real-world robustness. We present a public benchmark for hierarchical multilingual intent classification constructed from real logistics customer-service logs. The dataset contains approximately 30K de-identified, stand-alone user queries curated from 600K historical records through filtering, LLM-assisted quality control, and human verification, and is organized into a two-level taxonomy with 13 parent and 17 leaf intents. English, Spanish, and Arabic are included as seen languages, while Indonesian, Chinese, and additional test-only languages support zero-shot evaluation. To directly measure the gap between synthetic and real evaluation, we provide paired native and machine-translated test sets and benchmark multilingual encoders, embedding models, and small language models under flat and hierarchical protocols. Results show that translated test sets substantially overestimate performance on noisy native queries, especially for long-tail intents and cross-lingual transfer, underscoring the need for more realistic multilingual intent benchmarks.