Researchers have introduced a capability-driven data infrastructure that moves away from static dataset optimization toward a curriculum-based, dependency-aware training pipeline. This approach, which scales to 440 million images, demonstrates how aligning data supervision with generative capability acquisition improves model performance.
SPADE introduces a self-play framework where a single LLM generates its own training environments as executable code. This approach allows agents to scale beyond static datasets, achieving significant performance gains across math, science, and tool-use benchmarks.
Dots Studio has released the dots3-note preview, a 280B parameter multimodal Mixture-of-Experts model supporting 512K context. This open-weight release targets high-performance inference across text, image, video, and audio modalities.
ClawGym II introduces a unified framework for optimizing agents through complex, opaque harnesses using sandbox-based execution and trajectory reconstruction. This approach enables stable reinforcement learning on long-horizon tasks, yielding significant performance gains on benchmarks like ClawGym-Bench.
A new benchmark, StartupBench, reveals that even the most capable AI agents struggle to complete more than 30% of real-world, market-validated tasks. By moving away from researcher-designed tests to actual startup product workflows, the research highlights critical gaps in instruction following and domain expertise.
The newly released CPI-Bench addresses the limitations of existing image editing benchmarks by introducing multi-image, practical, and reasoning-based evaluation criteria. It aims to bridge the gap between academic model performance and real-world deployment efficacy.
Marionette introduces a modular architecture for interactive world modeling that separates geometric state prediction from visual rendering. By delegating physics to a zero-parameter renderer, the system achieves superior long-horizon stability and controllability compared to monolithic latent-space models.
Latent Dynamics Reasoning (LDR) introduces a novel approach to video world modeling by integrating kinematic laws into latent spaces rather than relying solely on pixel-level diffusion. This method demonstrates superior generalization and efficiency, outperforming traditional video diffusion models in physical reasoning tasks.
Intern-S2-Preview introduces a 397B parameter scientific foundation model designed for long-horizon reasoning and multimodal scientific tasks. It utilizes a novel Memory Decoder architecture to enable specialized domain adaptation without modifying the primary model weights.
A new research framework, OmniScientist, introduces a perception layer that allows AI agents to reason directly over raw, heterogeneous scientific data rather than relying on precomputed summaries. By integrating multi-modal inputs like video, audio, and 3D structures, the system successfully automates end-to-end research workflows across diverse scientific disciplines.
Alaya-EVOKE introduces an externalized, camera-indexed world state bank to decouple persistent memory from the denoiser context, enabling long-horizon, low-latency video generation. By redesigning the teacher model for linear-scaling supervision, the system maintains consistent world geometry without the memory explosion typical of traditional key-value caching.
AutoDesign introduces a meta-harness optimization framework that enables code agents to recursively improve their own design harnesses through rollout feedback. This approach outperforms existing commercial systems in academic poster generation by leveraging long-horizon agentic loops.
DreamX-Phi 1.0 introduces a specialized architecture for robotic manipulation that prioritizes geometric faithfulness over mere visual realism. By leveraging SE(3) transformations and multi-stage distillation, the model achieves state-of-the-art performance in the WorldArena 2.0 Challenge.
Qwen3.8-27B introduces a 27B parameter dense model optimized for complex agentic workflows, featuring native vision-language capabilities and configurable reasoning depth. It outperforms its predecessors across coding and multimodal benchmarks, positioning itself as a high-efficiency alternative for production environments.
The new PlayWorld benchmark introduces multi-modal Agent Players to evaluate video world models, addressing the critical challenge of assessing long-horizon spatial and physical consistency. By moving beyond fixed action sequences, this framework exposes significant reliability gaps in current state-of-the-art models.
A new research paper demonstrates that stronger AI models can construct inference-time harnesses to significantly boost the performance of weaker models without requiring parameter updates. This method, termed strong-to-weak scaffolding, effectively offloads reasoning into deterministic code and structured routing, nearly doubling target model accuracy on Theory-of-Mind benchmarks.
The newly proposed Cognitive Embodied Agents Architecture (CEAA) offers a modular framework designed to unify high-level reasoning with real-time execution in virtual environments. By integrating established paradigms like Sense-Think-Act and Belief-Desire-Intention, the architecture addresses the persistent divide between complex cognitive models and game-engine-constrained agent control.
A new benchmark, Cultivar, introduces source-contrastive evaluation to address data contamination and locale-specific performance gaps in multilingual translation models. By benchmarking 32 open-weight models, researchers demonstrate that current translation systems often struggle with non-US cultural contexts and exhibit signs of overfitting.