A formal theory proves model exploitation is essentially unavoidable on large policy sets in RL, generalizes reward hacking results, and derives a safe horizon for a relaxed version of exploitation.
super hub Canonical reference
Cosmos World Foundation Model Platform for Physical AI
Canonical reference. 80% of citing Pith papers cite this work as background.
247
Pith papers citing it
Background
80% of classified citations
abstract
Physical AI needs to be trained digitally first. It needs a digital twin of itself, the policy model, and a digital twin of the world, the world model. In this paper, we present the Cosmos World Foundation Model Platform to help developers build customized world models for their Physical AI setups. We position a world foundation model as a general-purpose world model that can be fine-tuned into customized world models for downstream applications. Our platform covers a video curation pipeline, pre-trained world foundation models, examples of post-training of pre-trained world foundation models, and video tokenizers. To help Physical AI builders solve the most critical problems of our society, we make Cosmos open-source and our models open-weight with permissive licenses available via https://github.com/nvidia-cosmos/cosmos-predict1.
hub tools
citation-role summary
background 39
method 6
baseline 3
dataset 1
citation-polarity summary
claims ledger
- abstract Physical AI needs to be trained digitally first. It needs a digital twin of itself, the policy model, and a digital twin of the world, the world model. In this paper, we present the Cosmos World Foundation Model Platform to help developers build customized world models for their Physical AI setups. We position a world foundation model as a general-purpose world model that can be fine-tuned into customized world models for downstream applications. Our platform covers a video curation pipeline, pre-trained world foundation models, examples of post-training of pre-trained world foundation models,
authors
co-cited works
representative citing papers
AnyFlow enables any-step video diffusion by distilling flow-map transitions over arbitrary time intervals with on-policy backward simulation.
SimWorld Studio deploys an evolving coding agent to create adaptive 3D environments that co-evolve with embodied learners, delivering 18-point success-rate gains over fixed environments in navigation benchmarks.
Physics-IQ benchmark reveals that generative video models exhibit limited physical understanding unrelated to their visual quality.
SENSE-VAD introduces the first synthetic benchmark dataset with per-frame labels for socially complex anomalies in autonomous driving scenes and shows existing video anomaly detectors fail on them.
A method to decompose 3D Gaussian splats into independent albedo and shading components for consistent texture editing in radiance fields.
JudgeFit produces per-VLM physical video evaluation taxonomies that improve held-out accuracy by a mean 32% relative to a single global schema across 16 models from eight families.
SWAAP is the first two-stage poisoning framework that identifies a harmful target world model via bilevel optimization and realizes it through stealth-constrained gradient matching on a limited fraction of fine-tuning transitions.
ThinkingVLA is a Mixture-of-Transformers VLA model that performs interleaved forward CoT for subgoal and image prediction followed by inverse CoT grounded on the predicted image to generate actions.
M* introduces the Walk Graph abstraction to serve arbitrary compositions of multimodal model components and reports latency and throughput gains over vLLM-Omni and other baselines on text-to-image, text-to-speech, and robotic planning workloads.
Self-distillation from a caption-conditioned video diffusion model to an image-and-prompt-conditioned executor, enhanced by RL from VLM feedback, enables task solving in world models.
World models introduce a stealthy poisoning vector into robot learning pipelines where malicious prompts or dynamics in teleoperated data activate only during synthetic trajectory generation, enabling backdoors in downstream policies.
ActProbe is an action-space detector that uses temporal consistency error and action chunk magnitude from policy outputs, mapped via LSTM-MLP, to predict failures earlier than baselines across policies and real-robot tasks.
A large-scale empirical study across tokenizers and diffusion backbones identifies Velocity Irreducible Variance (VIV) as one of the most stable predictors of latent diffusion generation quality.
VLMs formulate differentiable rewards from task-specific rules to enable test-time online LoRA optimization of VGMs, delivering 16.7-point gains on symbolic and general video reasoning benchmarks over VLM-as-solver and Best-of-N baselines.
SVI-Bench provides 35K hours of sports video with 9 tasks across four cognitive levels, revealing models drop from ~74% on action QA to 5% on agentic evidence integration.
YoCausal benchmark shows video diffusion models detect the arrow of time but lack genuine causal understanding relative to humans.
FlatSounds benchmark shows state-of-the-art V2A models rely more on text captions than visual input for physical and semantic accuracy, with captions improving correctness but degrading temporal alignment.
MiraBench defines action-conditioned reliability via three levels (physics adherence, action-following fidelity, optimism bias detection) and applies it to 12 model configurations using a 16,000-judgment human corpus, finding visual fidelity a poor proxy for action fidelity, no reliable scale benefi
JOPAT jointly models pixels, point tracks, and actions in a diffusion transformer and reports gains over pixel-only baselines on long-horizon robot tasks with occlusion and off-screen motion.
Aero-World adapts a pretrained latent diffusion transformer for action-conditioned aerial video generation by injecting inertial action tokens and using a frozen latent-space Physics Probe for inertial consistency supervision during LoRA finetuning, with a new AeroBench benchmark showing improved AA
BandTok tokenizes Mel-spectrograms as independent time-frequency band tokens from a single codebook and pairs it with 2D RoPE in an autoregressive model to improve music generation over residual multi-codebook tokenizers.
A training-free method reformulates camera control as geometric displacement fields applied via differentiable latent resampling, enabling control and bias probing in video diffusion models.
CreFlow combines LTL compositional rewards with credit-aware NFT and corrective reflow losses in online RL to improve embodied video diffusion models, raising downstream task success by 23.8 percentage points on eight bimanual manipulation tasks.
citing papers explorer
-
Imperfect World Models are Exploitable
A formal theory proves model exploitation is essentially unavoidable on large policy sets in RL, generalizes reward hacking results, and derives a safe horizon for a relaxed version of exploitation.
-
AnyFlow: Any-Step Video Diffusion Model with On-Policy Flow Map Distillation
AnyFlow enables any-step video diffusion by distilling flow-map transitions over arbitrary time intervals with on-policy backward simulation.
-
SimWorld Studio: Automatic Environment Generation with Evolving Coding Agent for Embodied Agent Learning
SimWorld Studio deploys an evolving coding agent to create adaptive 3D environments that co-evolve with embodied learners, delivering 18-point success-rate gains over fixed environments in navigation benchmarks.
-
Do generative video models understand physical principles?
Physics-IQ benchmark reveals that generative video models exhibit limited physical understanding unrelated to their visual quality.
-
SENSE-VAD: Sentient and Semantic Video Anomaly Detection for Autonomous Driving
SENSE-VAD introduces the first synthetic benchmark dataset with per-frame labels for socially complex anomalies in autonomous driving scenes and shows existing video anomaly detectors fail on them.
-
Intrinsic decomposition and editing of 3D Gaussian splats
A method to decompose 3D Gaussian splats into independent albedo and shading components for consistent texture editing in radiance fields.
-
Each Judge Its Own Yardstick: Discovering Per-VLM Taxonomies for Physical Video Evaluation
JudgeFit produces per-VLM physical video evaluation taxonomies that improve held-out accuracy by a mean 32% relative to a single global schema across 16 models from eight families.
-
Stealthy World Model Manipulation via Data Poisoning
SWAAP is the first two-stage poisoning framework that identifies a harmful target world model via bilevel optimization and realizes it through stealth-constrained gradient matching on a limited fraction of fine-tuning transitions.
-
ThinkingVLA: Interleaved Vision and Language Reasoning for Robotic Manipulation
ThinkingVLA is a Mixture-of-Transformers VLA model that performs interleaved forward CoT for subgoal and image prediction followed by inverse CoT grounded on the predicted image to generate actions.
-
M*: A Modular, Extensible, Serving System for Multimodal Models
M* introduces the Walk Graph abstraction to serve arbitrary compositions of multimodal model components and reports latency and throughput gains over vLLM-Omni and other baselines on text-to-image, text-to-speech, and robotic planning workloads.
-
World Model Self-Distillation: Training World Models to Solve General Tasks
Self-distillation from a caption-conditioned video diffusion model to an image-and-prompt-conditioned executor, enhanced by RL from VLM feedback, enables task solving in world models.
-
Targeting World Models to Compromise Robot Learning Pipelines
World models introduce a stealthy poisoning vector into robot learning pipelines where malicious prompts or dynamics in teleoperated data activate only during synthetic trajectory generation, enabling backdoors in downstream policies.
-
ActProbe: Action-Space Probe for Early Failure Detection of Generative Robot Policies
ActProbe is an action-space detector that uses temporal consistency error and action chunk magnitude from policy outputs, mapped via LSTM-MLP, to predict failures earlier than baselines across policies and real-robot tasks.
-
Diffusing in the Right Space: A Systematic Study of Latent Diffusability
A large-scale empirical study across tokenizers and diffusion backbones identifies Velocity Irreducible Variance (VIV) as one of the most stable predictors of latent diffusion generation quality.
-
VLMs are Good Teachers for Video Reasoning via Adaptive Test-Time Optimization
VLMs formulate differentiable rewards from task-specific rules to enable test-time online LoRA optimization of VGMs, delivering 16.7-point gains on symbolic and general video reasoning benchmarks over VLM-as-solver and Best-of-N baselines.
-
SVI-Bench: A Dynamic Microworld for Strategic Video Intelligence
SVI-Bench provides 35K hours of sports video with 9 tasks across four cognitive levels, revealing models drop from ~74% on action QA to 5% on agentic evidence integration.
-
YoCausal: How Far is Video Generation from World Model? A Causality Perspective
YoCausal benchmark shows video diffusion models detect the arrow of time but lack genuine causal understanding relative to humans.
-
Benchmarking Single-Factor Physical Video-to-Audio Generation
FlatSounds benchmark shows state-of-the-art V2A models rely more on text captions than visual input for physical and semantic accuracy, with captions improving correctness but degrading temporal alignment.
-
MiraBench: Evaluating Action-Conditioned Reliability in Robotic World Models
MiraBench defines action-conditioned reliability via three levels (physics adherence, action-following fidelity, optimism bias detection) and applies it to 12 model configurations using a 16,000-judgment human corpus, finding visual fidelity a poor proxy for action fidelity, no reliable scale benefi
-
Point Tracking Improves World Action Models
JOPAT jointly models pixels, point tracks, and actions in a diffusion transformer and reports gains over pixel-only baselines on long-horizon robot tasks with occlusion and off-screen motion.
-
Aero-World: Action-Conditioned Aerial Video Generation from Inertial Controls
Aero-World adapts a pretrained latent diffusion transformer for action-conditioned aerial video generation by injecting inertial action tokens and using a frozen latent-space Physics Probe for inertial consistency supervision during LoRA finetuning, with a new AeroBench benchmark showing improved AA
-
Modeling Music as a Time-Frequency Image: A 2D Tokenizer for Music Generation
BandTok tokenizes Mel-spectrograms as independent time-frequency band tokens from a single codebook and pairs it with 2D RoPE in an autoregressive model to improve music generation over residual multi-codebook tokenizers.
-
Probing into Camera Control of Video Models
A training-free method reformulates camera control as geometric displacement fields applied via differentiable latent resampling, enabling control and bias probing in video diffusion models.
-
CreFlow: Corrective Reflow for Sparse-Reward Embodied Video Diffusion RL
CreFlow combines LTL compositional rewards with credit-aware NFT and corrective reflow losses in online RL to improve embodied video diffusion models, raising downstream task success by 23.8 percentage points on eight bimanual manipulation tasks.
-
GenAI Powered Dynamic Causal Inference with Unstructured Data
A GenAI-based method extracts representations from unstructured data and uses a neural network to fit marginal structural models that recover causal effects of treatment feature sequences including their positions.
-
NoiseGate: Learning Per-Latent Timestep Schedules as Information Gating in World Action Models
NoiseGate learns per-latent timestep schedules as an information-gating policy in diffusion-based world action models, yielding consistent gains on RoboTwin manipulation tasks.
-
Do Joint Audio-Video Generation Models Understand Physics?
AV-Phys Bench shows that current joint audio-video models lack robust physical commonsense, with major drops on transitions and deliberate anti-physics prompts.
-
LiVeAction: a Lightweight, Versatile, and Asymmetric Neural Codec Design for Real-time Operation
LiVeAction is a lightweight asymmetric neural codec using an FFT-inspired encoder and variance-based training that outperforms generative tokenizers in rate-distortion while supporting real-time use on resource-constrained sensors across modalities.
-
EA-WM: Event-Aware Generative World Model with Structured Kinematic-to-Visual Action Fields
EA-WM generates more accurate robot world rollouts by projecting actions as structured visual fields in camera space and using event-aware bidirectional fusion to better capture interaction dynamics.
-
Action Agent: Agentic Video Generation Meets Flow-Constrained Diffusion
Action Agent pairs LLM-driven video generation with a flow-constrained diffusion transformer to produce velocity commands, raising video success to 86% and delivering 64.7% real-world navigation on a Unitree G1 humanoid.
-
Being-H0.7: A Latent World-Action Model from Egocentric Videos
Being-H0.7 adds future-aware latent reasoning to direct VLA policies via dual-branch alignment on latent queries, matching world-model benefits at VLA efficiency.
-
OmniShotCut: Holistic Relational Shot Boundary Detection with Shot-Query Transformer
OmniShotCut treats shot boundary detection as structured relational prediction via a shot-query Transformer, uses fully synthetic transitions for training data, and releases OmniShotCutBench for evaluation.
-
VistaBot: View-Robust Robot Manipulation via Spatiotemporal-Aware View Synthesis
VistaBot integrates 4D geometry estimation and spatiotemporal view synthesis into action policies to improve cross-view generalization by 2.6-2.8x on a new VGS metric in simulation and real tasks.
-
RoboWM-Bench: A Benchmark for Evaluating World Models in Robotic Manipulation
RoboWM-Bench evaluates video world models by converting their manipulation video predictions into executable actions validated in simulation, showing that visual plausibility does not guarantee physical executability.
-
EgoTL: Egocentric Think-Aloud Chains for Long-Horizon Tasks
EgoTL provides a new egocentric dataset with think-aloud chains and metric labels that benchmarks VLMs on long-horizon tasks and improves their planning, reasoning, and spatial grounding after finetuning.
-
ViVa: A Video-Generative Value Model for Robot Reinforcement Learning
ViVa turns a video generator into a value model for robot RL that jointly forecasts future states and task value, yielding better performance on real-world box assembly when integrated with RECAP.
-
MoRight: Motion Control Done Right
MoRight disentangles object and camera motion via canonical-view specification and temporal cross-view attention, while decomposing motion into active user-driven and passive consequence components to learn and apply causality in video generation.
-
SEM-ROVER: Semantic Voxel-Guided Diffusion for Large-Scale Driving Scene Generation
SEM-ROVER generates large multiview-consistent 3D urban driving scenes via semantic-conditioned diffusion on Σ-Voxfield voxel grids with progressive outpainting and deferred rendering.
-
A Frame is Worth One Token: Efficient Generative World Modeling with Delta Tokens
Delta tokens compress VFM feature differences into single tokens, enabling a lightweight generative world model that predicts diverse futures with far lower compute than existing approaches.
-
LongTail Driving Scenarios with Reasoning Traces: The KITScenes LongTail Dataset
KITScenes LongTail supplies multimodal driving data and multilingual expert reasoning traces to benchmark models on rare scenarios beyond basic safety metrics.
-
PlayWorld: Learning Robot World Models from Autonomous Play
PlayWorld learns high-fidelity robot world models from unsupervised self-play, producing physically consistent video predictions that outperform models trained on human data and enabling 65% better real-world policy performance via model-based RL.
-
SPIRAL: Self-Evolving Action-Conditioned Video Generation via Reflective Planning Agents
SPIRAL is a closed-loop think-act-reflect framework using PlanAgent, VideoGenerator, and CriticAgent plus GRPO self-evolution to improve long-horizon action-conditioned video generation, with new dataset and benchmark showing gains over open-loop baselines.
-
PhysMem: Scaling Test-Time Memory for Embodied Physical Reasoning
PhysMem enables VLM-based robot planners to learn and verify physical properties through test-time interaction and hypothesis testing, raising success on a brick insertion task from 23% to 76%.
-
CoMoVi: Co-Generation of 3D Human Motions and Realistic Videos
CoMoVi co-generates 3D human motions and 2D videos synchronously in a single diffusion denoising loop using 3D-to-2D projection and dual-branch diffusion with 3D-2D cross attentions.
-
Large Video Planner Enables Generalizable Robot Control
A video foundation model trained on human demonstrations generates zero-shot plans that convert to executable robot actions on novel scenes and tasks.
-
MVAD: A Benchmark Dataset for Multimodal AI-Generated Video-Audio Detection
MVAD is the first comprehensive benchmark dataset for AI-generated multimodal video-audio detection, with three realistic forgery patterns, high-quality outputs from state-of-the-art models, and diversity across visual styles and content categories.
-
Target-Bench: Can Video World Models Achieve Mapless Path Planning with Semantic Targets?
Target-Bench shows the best off-the-shelf video world model scores only 0.341 on semantic target-approaching and directional consistency, with fine-tuning on a small robot dataset yielding measurable gains.
-
EgoWalk: A Multimodal Dataset for Robot Navigation in the Wild
EgoWalk supplies 50 hours of real-world multimodal human navigation data in varied indoor/outdoor settings together with open pipelines that auto-generate language goal annotations and traversability masks.
-
SpatialScore: Towards Comprehensive Evaluation for Spatial Intelligence
Presents SpatialScore benchmark for MLLM spatial reasoning, evaluates 49 models showing large human gap, and supplies SpatialCorpus plus SpatialAgent to improve performance.
-
DreamGen: Unlocking Generalization in Robot Learning through Video World Models
DreamGen trains robot policies on synthetic trajectories from adapted video world models, enabling a humanoid robot to perform 22 new behaviors in seen and unseen environments from a single pick-and-place teleoperation dataset.