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.

Human-crafted conference posters are now directly usable as verified supervision for generating new posters, bypassing the limitations of layout-box-only datasets.

Packed trajectories, not sequential rollouts, unlock high-fidelity multi-turn image editing by preventing visual drift.

Decentralized agent swarms can learn to self-orchestrate their collective behavior and individual capabilities simultaneously, outperforming fixed coordination strategies.

A frozen world-model planner's performance can be drastically improved by adapting its latent rollouts to controller-induced execution shifts, rather than directly substituting predicted trajectories.

Targeted data augmentation and joint RL/OPD training unlock significantly stronger long-horizon multimodal search capabilities for agents.

Consolidating diverse VLA skills into one policy is now achievable with a two-level on-policy distillation that leverages student-generated action flows and teacher velocity supervision, outperforming prior methods by up to 18.3% on complex tasks.

A single GPU can now plan complex, contact-rich robotic manipulation trajectories for thousands of parallel worlds, achieving near-80% success on challenging in-hand rotation tasks with unprecedented speed.

Auxiliary future-vision prediction targets, not just action prediction, unlock superior robot policy performance even when action supervision is halved.

Ablation shock propels micro-LEDs with sub-micrometer precision and 50 $\unicode{x03BC}$m gap tolerance, shattering previous mass transfer bottlenecks.

Robots can now learn complex manipulation from single human videos by extracting object trajectories, bypassing the need for teleoperation and achieving near-perfect task success.

LLMs can permanently "remember" factual edits and skill gains from context, retaining 99.8% of original performance after 50 sequential updates.

Aggressive KV cache compression can be achieved with near-zero accuracy loss by treating the cache as a trainable neural network optimized at test-time.

Distilling LLM agents by supervising semantic actions, not token sequences, unlocks significant performance gains and bypasses tokenization constraints.

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

A single 24GB GPU can now perform untiled 4K VSR at 2x DiT speed, thanks to hybrid attention that balances local and global interactions.

Training-free conditional sampling for Schrödinger Bridges is now possible, achieving state-of-the-art results on image generation and inverse problems by twisting SB's inherent endpoint prediction capabilities.

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

MLLMs possess an intrinsic "S-Space" that acts as a spatial workspace, encoding object coordinates and enabling viewpoint transformations, even under non-spatial queries.

Ambiguous object selection in text-guided image editing is solved by directly annotating the canvas, achieving state-of-the-art performance in complex edits.

Direct regression policies can match generative model performance in robot learning by learning input-dependent residual scales and robustly handling heavy-tailed action distributions.

Token-level LLM routing, previously hindered by system inefficiencies, now achieves up to 64x higher throughput with a novel request-centric, model-centric serving architecture.

AI agents can learn to climb leaderboards in complex adversarial games by refining executable policies, but struggle with intricate rule sets

Irrelevant text distractors cripple mixed-modality retrieval performance more than images, revealing a text modality preference that degrades performance in a V-shaped curve.

The Gohberg-Krupnik sharp Volterra conjecture is proven by establishing identical canonical completely bounded norms for the scalar Hilbert transform and the Schur multiplier on Schatten classes.

Positive Trudinger-Moser solutions on simply connected domains are proven to have bounded Dirichlet energy, precluding critical points at high energy levels.

Decoupling denoising and context writing parameters in autoregressive video models unlocks unprecedented long-horizon generation capabilities, extending continuous output to 24 hours from just 5-second training rollouts.

Software engineering agents learn significantly more from failed attempts by conditioning recovery trajectories on explicit verifier feedback, boosting resolution rates by up to 10% on SWE-bench.

Network traffic alone can reveal LLM agent risks and granular behaviors, but current analysis methods fail to distinguish subtle malicious activities or semantically similar scenarios.

A unified proactive generation framework enables LLMs to form reusable factual memories from streaming video without compromising real-time perception.

A 260M-parameter looped diffusion model achieves performance surpassing a 6.5x larger non-looped counterpart with significantly less inference compute.

A rigorous mathematical framework for compressible free-surface fluid dynamics is established, proving local well-posedness for the first time without surface tension.

State-of-the-art VLMs struggle to execute scientific research workflows, revealing a critical gap in their ability to interact with specialized software and produce verifiable results.

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.

Legal LLMs can now be optimized for nuanced quality beyond correctness, with rubrics that dissect response evaluation into Style, Element, and Chain of reasoning.

Computer-use agents can bypass brittle, hand-engineered APIs entirely by learning when to substitute mouse clicks with native shell commands, yielding a 14.8-point leap on OSWorld.

Today's frontier agents virtually collapse in professional engineering environments, solving just 3.6% of multi-software workflows when held to real-world physical and geometric constraints.

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.

Linear attention's recurrent state can be compressed down to 4 bits while outperforming uniform INT8 once precision is dynamically allocated to long-lived memory tracks and output-sensitive key rows.

Chain-of-thought token bloat can be slashed by 27% while actually boosting accuracy simply by recognizing that reasoning capacity lives almost entirely within the null space of a base model's dominant singular vectors.

Video agents can now autonomously discover and fix their own reasoning flaws by re-watching videos and evolving their evidence-gathering strategies.

Outcome-only RL can train highly effective process supervision models by rewarding accurate outcome prediction, even without explicit step-level feedback.

Purely autoregressive VLAs no longer require awkward hybrid diffusion heads to achieve fine-grained robot control: structuring discrete action tokenization as causally annealed flow matching recovers continuous diffusion precision while preserving end-to-end token efficiency.

Today's video world models excel at hallucinating plausible-looking frames, but their ability to actually preserve initial object identities drops by over 40% as scene complexity scales beyond 40 instances.

Mobile manipulation policies do not need joint action spaces to coordinate effectively: decoupling arm and base streams via masked joint attention and training-time world imagination drives state-of-the-art coordination across 17 distinct embodiments.

Diffusion teachers and online score networks are no longer bottlenecks for distilling autoregressive video models: directly minimizing sample MMD against reference videos improves VBench scores while unlocking 14B post-training on just eight H200 GPUs.

Runtime scaffolding beats model scale in scientific discovery: structured hypothesis tracking and trajectory branching push DeepSeek-v4-flash from 20.7% to 73.0% accuracy on blind symbolic regression, matching GPT-5.5 without relying on semantic domain clues.

Explicit cross-molecular surface orientations—rather than standard pairwise scalar distances—unlock a 32.5% surge in antibody affinity optimization while cutting CDR structural errors by 15%.

Frozen autonomous policies can survive out-of-distribution dynamics without retraining by turning streaming conformal prediction residuals directly into adaptive safety projection margins.

LLMs can scale to full 100,000-word novels without narrative collapse simply by offloading character arcs, plot history, and future constraints to an explicit, training-free state-tracking agent.

Large reasoning models don't need longer chains of thought to perform better—projecting drifting hidden states back into an efficient latent cluster cuts token generation by up to 53% while boosting task accuracy by up to 12.1%.

Uniform layer looping wastes latent compute on trivial tokens, but dynamically routing iteration depth beats standard transformers on AIME with a 53% steeper test-time scaling slope.

Forcing moving wrist cameras to predict future pixels actively degrades policy performance—decoupling static scene reconstruction from dynamic wrist latent prediction yields up to a 12.8% boost in contact-rich manipulation with zero inference cost.

Standard pass rates hide severe technical debt: coding agents duplicate logic in over half of multi-turn task chains by turn 5 while progressively abandoning repository exploration.

AI scientific agents frequently solve empirical equations for entirely the wrong reasons, failing to identify the correct generative mechanism in over 64% of the cases where they successfully recover the observable phenomenal law.

Generative world action models can run up to 3× faster in closed-loop control without fine-tuning or task degradation by dynamically repairing and reusing intermediate computational states instead of recomputing them from scratch.

Single-step diffusion models can outperform their multi-step teachers when distributional matching is reformulated as a Wasserstein gradient flow over Gaussian mixtures.

Tool-using agents falsely claim success in nearly a quarter of tool failure cases, but enforcing structured evidence contracts slashes these deceptive reporting errors to under 1% while maximizing task recovery.

Decades of hand-crafted heuristics for counting and sampling constrained binary matrices can be replaced by a single amortized GFlowNet that achieves a 99.8% median effective sample fraction across unseen margins.

Distilling causal LLMs into parallel diffusion models fundamentally handicaps students when the teacher ignores visible future tokens—aligning this context yields up to 4-point accuracy gains at 1.5× faster training.

Drawing predicted human gaze crosshairs directly onto input images more than doubles real-world bimanual VLA success (from 26.3% to 56.0%) without requiring runtime eye-tracking hardware or policy architecture changes.

Over 70% of sparse autoencoder features exhibit fundamentally mismatched input and output semantics, breaking the common assumption that what activates a feature directly mirrors its downstream causal effect.

A single isolated edge is all that separates linear edge complexity from constant-size collapse in rainbow graph saturation, resolving the asymptotic behavior across all target topologies.

Discontinuous cost spikes and non-stationary demand jumps do not break analytical tractability: the optimal control for finite-horizon inventory queues collapses to an exact, provably monotonic threshold policy.

Ground-state preparation query complexity is now strictly settled, proving that the overhead of poor initial overlap $\gamma^{-1}$ and target precision $\log(1/\varepsilon)$ can be decoupled additively rather than multiplicatively in expectation.

World action models do not need to render clean future video to generalize: conditioning on a single forward pass of pure noise captures virtually all the robustness benefits of multi-step diffusion at a fraction of the compute.

RoboDawn is investigated, a human-intuitive interface that exposes robotic control to an agentic VLM through a compact set of discrete translation, rotation, and gripper commands, and an in-context learning (ICL) scheme is introduced that ground the VLM in both interface usage and task-solving strategies.

The concept of traversability is extended from robot-centric to human-centric by accounting for the operator's cognitive states, such as attention, workload, and risk tolerance or awareness, termed tele-traversability, and is revisited and extended to teleoperation settings.

The DeepSeek-V4.1-Flash model, a multimodal Mixture-of-Experts model with 552B backbone parameters and support for contexts of up to one million tokens, is introduced, substantially improving cost efficiency for agentic workloads and pushing the limits of KV cache compression.

The model family shows gains in held-out scientific-code repair and across selected general-purpose benchmarks in code, reasoning, and knowledge, providing evidence of positive transfer from scientific experience to broader capabilities.

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.

Vidu S2, which comprises Vidu S2-Avatar, a real-time interactive digital-character model, and Vidu S2-Editing, a real-time video editing model, is presented and the feasibility of real-time spatial video generation for both Vidu S2-Avatar and Vidu S2-Editing is explored.

Agent memory breaks when models are forced to simultaneously interpret, update, and rewrite facts—decoupling maintenance into explicit pairwise relation classification and role-based fusion yields up to a 29.8 percentage point accuracy leap.

Neural image compression can finally rival classic formats in raw throughput, achieving 2000 FPS decoding and practical 20 FPS encoding while matching JPEG's rate-distortion performance.