#1 ·
SO-Arena: Benchmarking the Capability Spectrum and Reliability of System One Models
Lichen Li, Ziping Dong, Sicheng Feng, Qi Li
BenchmarksSystem One ModelsReliability
Why it matters
System One models make fast, low-latency decisions without long token generation, which is exactly the regime you want for real-time agents and control loops. But aggregate accuracy on existing benchmarks hides when those fast decisions are unstable or refuse nothing. If you are wiring a fast decision head into a production loop, you need to know not just how often it is right, but whether it holds up when the option set shifts or when there is no valid option at all.
SO-Arena is a benchmark spanning 11 domains and five capability families. It grounds reference labels in reliable sources with manual verification, applies domain-specific difficulty criteria, and runs realistic interactive evaluation where models act iteratively in sandboxes. Beyond accuracy it measures decision stability under controlled option-set changes and missing-answer recognition when no valid option exists. Across 21 models, the headline result is that strong overall performance does not imply reliable decision-making: the best-performing System One model ranks only seventh in reliability.
arXiv | HuggingFace
#2 ·
BirdDreams: Joint Perspective and BEV Video Generation for a Plausible, Reactive Driving World Model
Junyoung Seo, Jaewoo Seo, Seungryong Kim, Laura Leal-Taixé
World ModelsAutonomous DrivingVideo Generation
Why it matters
Driving world models need two things that usually fight each other: plausible motion of the vehicles around you, and precise control over your own ego trajectory. Prior work had to be handed future vehicle trajectories as input to get control, which defeats the point of a world model. BirdDreams makes the scene layout part of what the model predicts, so ego control becomes a simple geometric transform on the model's own output instead of an external condition.
BirdDreams is an autoregressive driving world model that jointly generates perspective video and a schematic bird's-eye-view (BEV) video. The BEV, rendered from scene annotations at training, draws roads and vehicles as simple shapes at a fixed metric scale, reducing occlusion, while the perspective stream keeps details like traffic signals and object appearance. Two video diffusion transformers, one per stream, exchange joint attention so each informs the other. At inference it produces both videos from a single camera image, with no HD map or vehicle-state inputs, and translates/rotates the latest BEV frame along an input ego trajectory as a soft control condition. On PhysicalAI-AV and nuScenes it follows input ego trajectories more faithfully than prior driving world models, and human evaluation rates its surrounding-vehicle behavior as more realistic than a perspective-only baseline.
arXiv | HuggingFace
#3 ·
PhysLoop: Harnessing LLMs' Physics Awareness for Self-Improving Video Generation Models
Yuxuan Chen, Xiaojie Li, Hao Chen, Juncheng Ma
Video GenerationSelf-ImprovementPhysics
Why it matters
Video generation models look realistic but still get physical dynamics and causal consistency wrong, and the fine-grained physical supervision needed to fix that is expensive to obtain. PhysLoop uses an LLM's existing physics knowledge as the supervision source, guiding, selecting, and correcting generated videos so the model learns from physically consistent examples. The authors frame it as a milestone toward recursive self-improvement with limited human intervention, which is the pattern to watch if you build on top of video generation.
PhysLoop is a unified harness that guides and constrains LLMs across the whole loop of video generation, selection, and correction, distilling their physical knowledge into physically consistent videos. It supports controllable training-data construction as well as inference-time video repair. The team releases 100K videos spanning realistic indoor and outdoor scenes, along with generation tools and two evaluation datasets, PhysLoop-Gen and PhysLoop-Repair. They evaluate controllable generation and localized repair on those datasets and assess physical behavior on the external Physics-IQ Verified benchmark.
arXiv | HuggingFace
Outlook
The day's theme is the gap between plausible and reliable: a model can be strong on average and still make unstable decisions, a generated scene can look realistic and still break physics, and a driving simulation can follow a path only if its geometry is explicit. Expect follow-ups as System One architectures start shipping reliability metrics, and as world models close the loop between predicted layout and real-time ego control.