Tsinghua AI

A single scene can be represented by ~100K plane primitives (5MB), enabling video world models to maintain geometric consistency across hundreds of frames and large viewpoint changes.

Continuous, decomposed visual verification unlocks robust rule-governed generative world modeling, outperforming binary critics and enabling structured reasoning agents.

Agentic robots are no longer just predicting actions, but actively orchestrating their own learning and execution through goal-directed reasoning.

Ultralight action heads achieve state-of-the-art embodied task performance by learning to selectively retrieve visual information via shared latent bridge queries.

Imagined failures in a world model can pinpoint exactly which robot skill components need new demonstrations, boosting success from 13% to 75% with minimal data.

Robots no longer need human intervention to learn from their mistakes: foundation-model-guided exploration over modular skills can autonomously synthesize corrective rollouts that distill directly into low-level policy weights.

Fysiverse-3D-Vision is proposed, a unified vision-language-geometry framework for generative 3D scene reconstruction and executable asset construction from a single image that establishes a shared representation where spatial reasoning and geometric reconstruction mutually enhance each other, allowing object layouts to be inferred beyond the constraints of individual asset generators.

DACG-agent is presented, a drift-aware causal-graph agent that incrementally builds a causal knowledge graph from PubMed abstracts and applies a two-layer stopping policy with complementary roles: a KL-divergence monitor that detects posterior convergence (the accuracy layer), and a Bradley--Terry process reward model (PRM) whose online decline detection halts retrieval once evidence quality peaks (the efficiency layer).

A synthetic-to-real framework coupling physically validated data generation with a hierarchical, diffusion-based controller is proposed, which outperforms state-of-the-art manipulation policies in transport smoothness and dynamic stability.

This work presents OmniRisk, an omnidirectional planning framework that learns trajectory-level risk offline for efficient onboard evasion and enables efficient onboard avoidance, with real-world flights demonstrating consecutive evasive maneuvers at relative encounter speeds up to 15 m/s without fine-tuning.

Budgeted Express-Mesh is presented, a topology-routing co-design that adds a small number of traffic-aware express links under a fixed wire budget that improves high-load throughput over Mesh and random placement, and annealing further improves Greedy.

Scheduling imagination in VLA models can cut GPU costs by 80% while boosting performance and robustness in real-world tasks.

Multi-turn jailbreak attacks reveal that framing requests as actionable can significantly reduce model vulnerability, highlighting a critical leverage point in AI safety.

Time-varying transition dynamics in Poisson-Gamma models lead to a marked improvement in predictive accuracy for count time series.

Evolving causal models allows scientific agents to refine their beliefs and predictions more effectively, outperforming traditional methods in experimental settings.

CAER redefines how we train world models by ensuring that only the most relevant tokens drive learning, leading to substantial gains in video generation quality.

A closed-loop world model that learns from both successes and failures could revolutionize dexterous manipulation in robotics.

Recalibrating supervision in world model training can drastically enhance interaction fidelity and physical plausibility without the need for external representations.

Achieving minute-long real-time interactive world generation while maintaining geometric consistency and dynamic subject identity is now a reality with Matrix-Game 3.5.

Riemann-1.0 transforms embodied intelligence by unifying robot policy execution and world simulation, achieving unprecedented success rates in real-world manipulation tasks.

Achieving a 25% reduction in prediction error, MOSH-WM redefines the landscape of object-centric video forecasting by grounding state representations in visual support.

Command-inconsistent memory can lead to a 15.6% increase in collision rates during autonomous driving, but MomADv2 effectively filters this noise for safer long-horizon planning.

Achieving a 99.8% success rate in reconnecting initially disconnected convex regions transforms the feasibility of GCS-based motion planning.

RuleMaze reveals that separating perception, execution, and rule verification can dramatically enhance MLLMs' ability to follow complex natural-language instructions in spatial planning tasks.

DECOWAM achieves a 21.7% reduction in action prediction error while maintaining robust task performance, showcasing the power of embodiment-aware factorization in mobile manipulation.

GS-Voxel revolutionizes 3D scene generation by allowing scalable, fitting-free structured latents that adapt to the complexity of the scene without the need for per-scene optimization.

The empirical analysis reveals that larger models may excel in generating narratives but often fail to maintain coherence and depth, exposing a critical trade-off in LLM storytelling capabilities.

Visual perturbations can significantly alter predictions in world models, but ACPC offers a quantifiable way to diagnose and mitigate these effects.

Bypassing RGB entirely, Latent-to-4D achieves significant improvements in 4D scene generation while maintaining a reusable framework across different video models.

Robots can now predict not just immediate actions but also the next stages of complex tasks, leading to more efficient manipulation strategies.

ChemWorld allows researchers to isolate the impact of hidden chemical laws on agent behavior, enabling unprecedented control and replayability in autonomous chemistry experiments.

JEPA-WAM achieves a remarkable 79.2% on LIBERO-Plus without large-scale pretraining, setting a new benchmark for efficient robot control.

Achieving a 14.6% improvement in planning accuracy while slashing communication costs by over half, DH-VLM redefines the potential for cooperative autonomous driving.

Simulation traces can transform LLMs into powerful tools for diagnosing and improving complex scheduling policies, achieving unprecedented performance gains.

FactorDrive redefines autonomous driving by seamlessly integrating spatial-physical evidence into adaptive reasoning, achieving state-of-the-art planning performance.

Achieving state-of-the-art performance in both 4D reconstruction and point tracking, Uni4R leverages continuous velocity fields to model dynamics at any timestamp, breaking free from traditional limitations.

Revisiting the same locations across time in Sekai2 enables the learning of persistent scene representations, a game-changer for interactive world modeling.

SLIM achieves state-of-the-art performance in robot manipulation with just 0.5B parameters, outperforming larger models while slashing GPU memory usage and inference latency.

EffectLearner achieves unprecedented video object removal quality by effectively reasoning about complex object-induced effects in dynamic scenes.

Early trajectory decoding from video diffusion models can cut planning latency by nearly 50% without sacrificing decision quality.

Ignoring future cross-variable dependencies can lead to significant forecasting errors, but a new structural regularizer, CvLoss, bridges this gap and boosts model accuracy.