Tsinghua AI

Uncovering 1,834 OAuth-based Authentication misuses in mini-programs reveals critical flaws that could allow attackers to impersonate users across multiple platforms.

Harness VLA boosts the performance of frozen VLA models by 38.6 percentage points on challenging manipulation tasks without the need for finetuning.

Frequency usage in transformers is not random; it’s intricately tied to the data’s dependency structure, revealing a data-driven mechanism behind RoPE's emergent behavior.

Multimodal unlearning could revolutionize how we handle sensitive data in AI, enabling targeted removal without sacrificing model performance.

Over 40% improvement in analytical efficiency could revolutionize how researchers conduct trajectory inference in single-cell transcriptomics.

Current LLMs falter in complex deliberative collaboration tasks, revealing critical gaps in their reasoning capabilities even when aided by external tools.

Annotation noise in vascular CT scans can be detected with a novel method that reveals systematic biases, improving training robustness dramatically.

Existing text-to-image models struggle to capture individual aesthetic preferences, but PIPBench reveals critical gaps in their performance that could redefine personalized image generation.

Outperforming previous methods, UniLM-Nav achieves zero-shot last-mile navigation by effectively integrating multimodal reasoning and task context.

Achieving a staggering 98.75% success rate in dexterous manipulation tasks, LAMP redefines how we approach real-world learning in robotics.

CompactionRL enables LLMs to effectively manage long-horizon tasks by summarizing context, leading to substantial performance gains in coding benchmarks.

Surpassing human performance in gaze estimation, PaGE closes the human-AI gap by over 60% while remaining lightweight for real-world applications.

CuRe transforms video captioning reward design by shifting from holistic evaluations to precise claim-level verification, significantly boosting factual accuracy and diversity in generated captions.

ShadowProbe uncovers hidden algorithmic risks in codebases, revealing vulnerabilities that traditional methods miss, with significant implications for software reliability.

ProCon achieves unprecedented anomaly detection accuracy without the need for training or pseudo-anomaly supervision, redefining the capabilities of memory-based methods.

TL-ANDI transforms how Tabular Foundation Models handle transfer learning by optimizing context selection to prevent negative transfer.

NPUs can waste up to 40% of energy due to suboptimal configurations, but a new profiling tool reveals how to cut this waste significantly.

Cortex outperforms traditional models by enabling zero-shot execution of complex long-horizon tasks, bridging the gap between high-level planning and low-level execution.

ResearchStudio-Reel not only automates research dissemination but does so with unprecedented quality, outperforming both traditional methods and leading LLMs in aesthetic appeal and information accuracy.

Current avatar systems are more diverse than ever, yet foundational prior learning is often overlooked in discussions of photorealistic digital humans.

HALO-WA boosts robotic manipulation success rates from 26.4% to 87.1% by effectively adapting to real-world errors in just over an hour of training.

State-of-the-art planners falter in long-tail scenarios, revealing critical gaps in autonomous driving safety and effectiveness.

Many robotic policies that seem successful in manipulation tasks actually compromise safety, with SoftVTBench revealing a stark contrast between goal completion and physical safety metrics.

Automatically constructed data can dramatically enhance the temporal localization abilities of audio models, overcoming the limitations of manual annotation.

ACE achieves a remarkable 70% success rate in constraint retrieval tasks without any task-specific retraining, showcasing the power of zero-shot workflow reasoning in robotic manipulation.

UI-MOPD achieves a remarkable balance between retaining existing capabilities and adapting to new platforms, with task success rates that challenge conventional approaches in GUI agent learning.

Users can now generate and control high-quality videos in real-time using just their voice, revolutionizing interactive content creation.

Bias in LLM judges can be corrected to improve ranking accuracy, lifting recall rates significantly in noisy environments.

Unconstrained egocentric video generation now achieves unprecedented fidelity and control by disentangling hand and camera motion with a novel 3D-aware representation.

WorldSample achieves a 28% boost in policy success rates while slashing training steps by nearly 60% through innovative real-synthetic data integration.

AutoMIA can generate complex 3D illusions in under 80 seconds, revolutionizing the intersection of art and computational design.

CheckRLM cuts error accumulation in reasoning chains by correcting factual inaccuracies in real-time, outperforming traditional approaches.

MiShield outperforms leading moderation tools by effectively identifying harmful semantics in multi-image content that appear benign in isolation.

MG-RWKV achieves state-of-the-art TFL performance with a groundbreaking O(T) complexity, redefining efficiency in audio-visual content authenticity verification.

Transforming image quality assessment from a single score to a nuanced diagnosis of multiple quality issues could revolutionize smartphone ISP tuning.

Positional leakage in 3D masked autoencoders can be mitigated, leading to significantly improved semantic representation quality.

Achieving a 22.3% word error rate with just 240 ms latency, LipsFlow redefines the capabilities of Visual Speech Recognition in challenging multi-speaker environments.

BRIDGE slashes circuit execution time by up to 1000x while achieving unprecedented fidelity in neutral atom quantum computing.

A single BDDL specification can drastically enhance the efficiency and effectiveness of embodied task planners, achieving a 25.9% performance improvement over existing baselines.

Current interactive world models fall short, with none passing the rigorous tests of WorldRoamBench designed to assess long-horizon stability across action, vision, physics, and memory.

DynFly achieves a remarkable 4.69 improvement in navigation performance, showcasing how dynamic-aware trajectory generation can transform UAV navigation in complex urban environments.

Physics-aware constraints in a neural operator can drastically enhance the accuracy of temperature field predictions in fusion devices, outperforming conventional methods.

OVOW achieves unprecedented accuracy and speed in reconstructing 4D scenes from a single video, making it a game-changer for physics simulation in AI.

Late visual-token updates can be safely ignored, leading to a 33.7% reduction in computational load without sacrificing performance.

LLMs show significant vulnerability to logical fallacies, with distinct profiles of resilience that could inform future model training strategies.

CIPE-Dance is a game-changer, providing the largest dataset for dance video generation and enabling OmniDance to set new benchmarks in multimodal video synthesis.

Online imitation learning can outperform offline methods, but only when the student can effectively represent the expert—realizability is key.

Thousands of AI-Apps are leaking sensitive credentials and harboring vulnerabilities that could lead to arbitrary code execution, revealing a critical security crisis in the AI ecosystem.

CI-MSE dramatically improves the correlation between offline validation and real-world performance, making it a game-changer for robot policy evaluation.

Token-level learning dynamics, not just model size, dictate scaling laws in language models, revealing actionable insights for training optimization.

TACO reveals that agentic models can learn to optimize tool usage without external judges, achieving higher accuracy and efficiency in multimodal tasks.

Even the most advanced LLMs struggle with consistent rubric verification, revealing substantial noise in scoring outputs across complex agentic scenarios.

Urban facade reconstruction can achieve superior geometric accuracy by integrating lightweight structural supervision, overcoming common pitfalls of traditional methods.

DRIFT not only sets a new state-of-the-art in self-improvement for language models but also redefines how we can dynamically adapt learning strategies based on problem difficulty.

Flow Splatting achieves superior image quality and faster rendering speeds by efficiently modeling dynamic scenes with 4D Gaussian representations.

STEAM redefines how robots learn from mixed-quality data, achieving up to 59% higher success rates in real-world tasks by effectively identifying reliable progress.

COSM achieves a remarkable 2.8x improvement in PIM throughput while keeping CPU performance degradation under 2.0%.

GUICrafter achieves superior GUI agent performance with just a fraction of the data, revolutionizing the way we think about training in data-scarce environments.

Off-policy distillation fails in multi-task settings, but a two-phase approach combining it with on-policy refinement can achieve single-task expert performance across multiple tasks.

GeoEdit achieves unprecedented geometric accuracy and identity fidelity in object editing, overcoming the limitations of existing diffusion-based methods.

UniGP reveals that joint training of controllable generation and dense prediction can significantly enhance performance without the need for complex designs, outperforming specialized models.

Effective keyframe extraction can boost MLLM performance by over 2% on complex video-guided tasks, revealing a critical link between video understanding and procedural learning.

Evaluating LLM agents in microservice failure diagnosis reveals that traditional outcome-based benchmarks miss critical reasoning processes, which these new datasets effectively capture.

ProMSA achieves superior accuracy in KB-VQA by dynamically selecting retrieval strategies, outperforming traditional fixed pipelines.

Pressure integration in humanoid motion imitation significantly enhances accuracy and stability, revealing the limitations of traditional vision-based methods.

A transparent probe-success rule boosts robot policy selection success rates by over 14 percentage points, revealing the hidden power of pre-deployment evaluations.

Effective service zone design can outperform battery upgrades in profitability, especially under varying demand conditions.

PhysEditWorld reveals that explicit control over physical parameters can transform how game world models interact with their environments, leading to more realistic and manipulable simulations.

EGG achieves a remarkable 2.13x speedup in GPU kernel generation, setting a new benchmark for performance in automated optimization.

HarmVideoBench reveals that existing benchmarks miss critical layers of harmful video understanding, while a new method boosts model accuracy by over 20%.

ViQ achieves a groundbreaking balance between semantic richness and detail in visual representations, enabling efficient multimodal training without sacrificing quality.

DeformGen transforms the landscape of deformable manipulation by enabling effective policy learning through innovative state augmentation and trajectory adaptation techniques.

SVP-IL boosts success rates on ambiguous language tasks by over 60% with minimal training data, revolutionizing data efficiency in robotic manipulation.

Despite high benchmark scores, SOTA semantic code clone detectors falter in real-world scenarios, revealing a reliance on shortcut learning over genuine semantic equivalence.

ASSCG cuts inference latency by 60% while boosting performance scores in autonomous driving systems, redefining how LLMs can be efficiently integrated into fast-slow planning architectures.

Internal biological constraints can dramatically reduce errors in hand pose estimation, enabling robust tracking in metric space.

Evolving hardware-aware compression techniques can outperform human designs, achieving unprecedented efficiency in deploying massive AI models.

Expressiveness preservation in speech-to-speech translation remains a significant hurdle, with systems scoring poorly on emotional and nonverbal fidelity despite achieving high translation accuracy.

Teacher-forcing consistency models can accelerate autoregressive video generation by ten times, revolutionizing the training landscape for streaming applications.

Achieving a 14-point boost in grounding accuracy, VistaRef redefines how we approach spatial orientation in AR and human-robot interaction.

Agents can now escape the Self-Confirmation Trap, leading to more reliable experience learning and improved self-evolution.

PreciseDoc achieves unprecedented precision in grounding critical document elements, transforming how LMMs can interpret complex text-rich environments.

MambaRaw achieves a remarkable 1.4 dB increase in PSNR at low metadata bitrates while slashing coding latency by nearly 9%, setting a new benchmark in raw image reconstruction.

Embedding geometric intelligence into segmentation models can dramatically enhance the recovery of small vascular structures and improve overall topology fidelity.

WVM outperforms existing models by accurately assessing task progressions and improving robotic manipulation from both expert and suboptimal data.

ReMMD-Agent achieves a remarkable 41.80% accuracy in detecting misinformation across complex multilingual and multi-image scenarios while slashing verification costs by up to 80%.

PoinTriE achieves state-of-the-art performance in point cloud video tasks while slashing memory requirements and annotation costs.

A unified runtime boundary and time-aware execution can boost LLM agent accuracy by over 2% in long-horizon tasks, revealing a critical leverage point for enhancing agent stability.

Mistakes in human demonstrations can enhance robot learning when properly harnessed, revealing a new dimension of value estimation that traditional methods overlook.

Training on SignNet-1M boosts sign language model robustness by improving generalization across diverse real-world conditions without sacrificing performance.

PointVG-R achieves a groundbreaking 15.86-point boost in mIoU by integrating geometric reasoning into visual grounding tasks, reshaping our approach to spatial interpretation in models.

Isolated assessments may mask biases, but comparative evaluations can unleash hidden discrimination in LLMs, especially as model sizes grow.

Counterfactual controllability in video generation could be the key to creating self-evolving world models that understand and adapt to their actions.

RaDaR can identify rare diseases 1.87 months earlier than traditional methods, revolutionizing diagnostic timelines for patients.

Achieving a 5x speedup in document parsing without sacrificing accuracy could redefine efficiency benchmarks in Vision-Language Models.

Reinforcement learning enables video-LLMs to re-watch and refine answers without the costly overhead of chain-of-thought training, achieving better performance with less computation.

REDI-Match not only sets a new benchmark in dense feature matching but also accelerates inference speed, achieving 41 FPS on a single GPU.

No single memory architecture is best for all tasks; performance hinges on how well memory structures align with specific workload challenges.

Quantum pseudorandom states can only be stretched to a limited extent, revealing a stark contrast with classical counterparts.

Current vision-language models falter in streaming interaction understanding, with alarming mis-calibration leading to confidently incorrect predictions.