Moonshot AI has released Kimi-K3, a 2.8 trillion parameter Mixture-of-Experts model featuring a 1-million-token context window and native multimodal capabilities. This release introduces the Kimi Delta Attention architecture and marks a significant shift toward open-weight frontier models capable of long-horizon autonomous engineering.
SWE-Pruner Pro introduces an integrated approach to context management for coding agents by leveraging internal model representations to prune irrelevant tool outputs. This method achieves significant token savings and performance gains on standard benchmarks without requiring external classifiers.
FVAttn introduces a training-free sparse attention system that mitigates load imbalance in multi-GPU video generation. By utilizing runtime load balancing and slack-aware augmentation, it achieves significant speedups over standard FlashAttention.
Researchers have introduced the 'Data Pyramid,' a structured framework for categorizing the diverse data sources required for training embodied AI agents. This taxonomy helps developers navigate the trade-offs between scalability and physical alignment, providing a roadmap for building more capable robotic systems.
ClinFusion is a new vision-centric multimodal LLM designed to unify 2D and 3D medical image understanding through a novel cascaded encoder architecture. By introducing specialized benchmarks like MedIF-Bench and RoI-grounded evaluation, it sets a new state-of-the-art in clinical report generation and instruction following.
The Kimi Team has released Kimi K3, a 2.8T parameter Mixture-of-Experts model featuring 104 billion active parameters and a 1-million-token context window. This release introduces architectural innovations like Kimi Delta Attention and Stable LatentMoE, marking a 2.5x improvement in scaling efficiency over its predecessor.
Researchers have identified a critical failure mode in on-policy diffusion distillation called Negative Branch Asymmetry, where classifier-free guidance leads to antagonistic error dynamics. The proposed Positive-Direction Matching objective offers a branch-aware solution to improve knowledge transfer in complex tasks like video control.
Researchers have introduced a controlled, multi-turn environment to isolate the mechanisms behind long-horizon planning in foundation models. The study reveals that explicit world modeling through chain-of-thought state transitions and specific post-training distillation techniques are critical for robust agentic performance.
O-VAD introduces a training-free, agentic framework for industrial video anomaly detection that mimics human inspection by tracking object state evolution. By focusing on spatial-temporal dynamics rather than domain-specific retraining, it provides a more interpretable and flexible solution for complex manufacturing environments.
Skill Self-Play (Skill-SP) introduces a co-evolutionary framework that balances task diversity with verification reliability in LLM training. By utilizing a tripartite architecture of a proposer, solver, and skill controller, the method enables models to expand their capabilities through autonomous, verifiable self-play.
A new framework, Agentic Context Management (ACM), shifts the focus from simple storage-retrieval to a holistic lifecycle approach to improve agent performance and cost-efficiency. By implementing five core primitives, developers can achieve linear token costs while maintaining high fidelity in long-running agent interactions.
SceneActBench introduces a new framework for evaluating vision-language model agents that perform actions within complex 3D scenes. By testing across five distinct tasks using a unified agent-environment loop, the benchmark reveals significant performance gaps in current proprietary models.
FlashRT introduces a novel chain-of-program agent harness designed to automate the complex deployment of real-time multimodal pipelines. By iteratively transforming reference implementations into optimized multi-GPU configurations, it achieves significant latency and throughput gains across diverse hardware platforms.
Researchers have introduced Experiential Learning (EL), a novel post-training framework that replaces traditional scalar reward signals with high-bandwidth textual feedback. By transitioning from an 'LLM-as-a-Judge' to an 'LLM-as-a-Coach' architecture, the method improves generalization and mitigates reward hacking in complex, non-verifiable tasks.
A new benchmark, ActiveVision, reveals that frontier multimodal large language models struggle with active, closed-loop visual observation. Despite high reasoning capabilities, these models fail to perform the iterative gaze redirection required for complex visual tasks, highlighting a significant architectural gap.
Researchers have introduced Differentiable Logic Gate Networks (Diff-Logic) to address the latency and memory bottlenecks of running neural networks on edge devices. By compiling models into pure Boolean circuits, this approach achieves significant performance gains in EEG classification tasks while maintaining efficiency on power-constrained hardware.
GraphVid introduces a novel graph-conditioned framework for video generation, enabling precise multi-object control through structured interaction graphs. By moving beyond traditional pixel-level trajectory tracking, the model achieves superior performance with fewer parameters and training data.
Visual Contrastive Self-Distillation (VCSD) introduces a novel approach to on-policy self-distillation that eliminates the need for external teachers or privileged information. By leveraging image-content erasure as a contrastive signal, VCSD improves performance across Qwen3-VL model scales without adding inference-time overhead.