The new Principia benchmark reveals that state-of-the-art video generation models struggle significantly with Newtonian physics, scoring poorly on relational consistency. Despite high performance on existing metrics like VBench, these models fail to maintain predictable physical relationships between objects in a scene.
Puffin-World introduces a unified multimodal architecture that integrates physical understanding, geometry, and appearance without relying on external offline modules. By training on the massive Puffin-16M dataset, the model enables physically consistent 3D world generation and dynamic simulation.
A new benchmark called IdeaAMBIG reveals that while LLMs are proficient at clarifying research methods when defects are identified, they struggle significantly to locate those defects in the first place. This research highlights a critical bottleneck in the automation of scientific implementation.
SenseNova-U1.5 introduces an 8B-MoT architecture that eliminates VAEs and encoders, moving toward a fully end-to-end framework for visual generation and reasoning. By leveraging spatially coherent patch reconstruction and multi-expert on-policy distillation, the model achieves high-fidelity 4K generation without traditional pipeline bottlenecks.
Researchers have demonstrated that reasoning capabilities are language-agnostic, allowing models to reason effectively in non-English languages without requiring specific reasoning supervision in those languages. By optimizing data composition, the new Tiny Aya L2-Thinker model achieves a 93% L2 reasoning rate across 60 languages at a 3.35B parameter scale.
RoboSPA introduces a rigorous diagnostic framework for Vision-Language-Action models, highlighting critical failures in spatial reasoning and long-horizon planning. By testing 280 task variants, the benchmark reveals that current state-of-the-art agents struggle significantly as complexity scales.
A new benchmark, WearableQA, challenges LLMs to interpret complex, longitudinal wearable data across 4,084 multiple-choice questions. By testing reasoning on real-world physiological signals, the dataset exposes significant gaps in current model performance, with most systems failing to exceed 60% accuracy.
Show-Harness introduces a semantic interface that allows off-the-shelf vision-language models to control robots directly, bypassing the need for specialized embodiment-specific training. This approach leverages discrete action units to ground high-level VLM reasoning into physical robot movements.
Researchers have introduced a novel framework that separates world-state logic from visual generation, enabling persistent, rule-based interactions in video world models. By utilizing a lightweight engine to manage entity states and state-augmented 3D bounding boxes, the system achieves unprecedented control over long-horizon video generation.
Researchers have introduced Terminal-Universe, a framework that reconstructs executable development environments from existing terminal-based agent trajectories. This method creates 37.3k verifiable task environments, significantly boosting the performance of code-generation models like Qwen3.5.
DRACO introduces a dynamic rubric-based approach to reinforcement learning, enabling agents to improve performance in long-horizon tasks without needing programmatic verifiers. By redistributing trajectory-level scores into fine-grained, per-step advantages, it outperforms standard GRPO baselines on complex benchmarks.
EditVid introduces a training-free framework that unifies instruction-guided and subject-guided video editing. By combining sparse causal memory, token injection, and soft latent blending, it achieves superior fidelity and coherence compared to existing baselines.
RISE introduces a method to construct synthetic teachers for language model training by extrapolating from a model's own training trajectory. This approach eliminates the need for external teachers or privileged conditioning, enabling recursive, self-improving policy distillation.
WorldSculpt introduces a novel paradigm for generating compositional 3D representations of densely cluttered scenes by adapting single-object generative priors to multi-view observations. This approach enables the reconstruction of complex environments as collections of individual meshes, solving significant occlusion challenges without requiring scene-level training.
A new model, BDH-CQ, introduces recurrent latent reasoning to solve complex tasks without verbalizing intermediate steps. By achieving 29.5% pass@2 on ARC-AGI-1 at a cost of $0.0007 per task, it establishes a new efficiency benchmark for reasoning models.
UniMate introduces a topology-aware diffusion transformer capable of generating articulated motion for arbitrary 3D skeletons without per-skeleton fine-tuning. By utilizing a novel spectral approach to position embeddings and a diverse dataset, it enables zero-shot animation across varied skeletal structures.
Researchers have introduced a new framework for training terminal agents by dynamically evolving environments off-policy. This approach addresses the stagnation of learning signals in static benchmarks, yielding significant performance gains on Terminal-Bench 2.1.
Researchers have introduced 'Compile by Training,' a method that converts natural-language specifications into lightweight, standalone neural functions. This approach eliminates the need for repeated calls to large, remote LLMs by distilling task-specific logic into compact, deployable adapters.