Tsinghua AI

Consolidating specialist LLMs into a single, more accurate generalist is possible with *zero* gradient updates, relying solely on forward passes and input statistics.

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.

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.

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.

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.

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.

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.

Experiments on three open-weight LLMs show that SimPrint reliably recovers the owner signature in both clean and modified settings, remains robust under fine-tuning, pruning, quantization, model merging, and serving-time perturbations, and maintains comparable downstream utility.

DART, a training-free method that combines low-rank coordinate transport with target-schedule response calibration with forward evaluations and no source training videos is proposed, and Adapter-level results reveal positive functional effects for some adapters and strong attenuation with reduced negative functional effects for others.

A reward-compatible bounded mixing mechanism for $\gamma\mathrm{OPD}$ that balances verifiable outcome feedback with the discounted OPD advantage to move beyond purely teacher-dependent optimization is developed.

Comparing full-precision and quantized forward passes, and identifying two mechanisms that characterize pretrained quantization robustness, these mechanisms explain why quantization error that passes through numerous layers can still produce only small output changes, and are verified across models and quantization settings.

LLMs can parse complex clinical and security logs using just 1% to 2% of the original context length without sacrificing predictive performance or explanation faithfulness.

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.

Chain-of-thought context can be halved without sacrificing accuracy by using the geometric trajectory of hidden states to selectively compress exploratory reasoning steps into continuous latent tokens.

On-policy distillation is massively data-overfed: a single training prompt recovers most full-dataset performance gains, while just 16 prompts saturate 98.9% of reachable state space to match full-data distillation.

Group-Calibrated On-Policy Distillation boosts long-context reasoning performance by reconciling teacher guidance with verifier feedback, achieving up to a 12-point increase in benchmark scores.

TileMix achieves a breakthrough in LLM inference by enabling mixed-precision attention that boosts throughput while maintaining long-context quality.

Pruning 50% of channels in RGB-infrared object detectors can actually boost performance by 0.6% mAP, challenging conventional wisdom about redundancy.

BAMU achieves a remarkable MOS improvement in speech quality by dynamically allocating quantization resources based on frame complexity, outperforming traditional fixed-depth codecs.

Hard prompt compressors can leave critical context gaps, leading to a staggering 60% of examples suffering from referential dangling, which severely impacts accuracy in multi-hop question answering.

Accepting selective mismatches in autoregressive decoding can boost throughput by over 15% without any additional training or model adjustments.

PhyAI achieves up to 4.65x speedup in Physical AI tasks by unifying disparate inference processes into a single, efficient runtime.

Substantial IP leakage risks in autonomous multi-agent systems are revealed through a novel framework that extracts harness capabilities dynamically during inference.

Evolving contexts can transform the stability and effectiveness of on-policy distillation, leading to superior performance in open-ended tasks.

Pruning 80.2% of vision tokens without sacrificing accuracy could revolutionize the efficiency of multimodal large language models.

Evolving Cache Schedules can slash action-generation time by over 8x without sacrificing performance, revolutionizing real-time deployment of diffusion policies.

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

Weak models can teach strong ones how to act better, boosting performance without the heavy lifting of direct RL training.

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

Achieving 100% task success in closed-loop execution, Embodied.cpp revolutionizes how embodied AI models are deployed across diverse hardware platforms.

Hypic slashes time-to-first-token by 2.45x and doubles throughput for hybrid-attention LLMs, all while preserving near-full accuracy.

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

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.

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

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

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

MOCAP slashes LLM inference latency by over 76% while boosting throughput more than threefold, redefining efficiency for long-context processing on wafer-scale chips.

Token-oriented inference optimizations can cut production costs and boost efficiency, transforming large model services from merely callable to fully operable.

d-OPSD enables dLLMs to learn from their own future outputs, drastically improving sample efficiency and performance in reasoning tasks.

CoreMem achieves a remarkable +4.51 percentage point improvement in open-domain reasoning accuracy while operating within a strict 8 GB VRAM budget.

TetherCache slashes quality drift in long-form video generation from 7.84 to 1.33, ensuring stability and coherence over extended sequences.

SKIM compresses procedural skills in LLMs by 30-60% without sacrificing performance, revolutionizing how we manage reusable natural language skills.

Local MoE inference can now match cloud-level performance, achieving 1,800 tokens/s on consumer GPUs while maintaining original model precision.

ReasonAlloc reallocates KV cache resources in real-time, achieving superior reasoning efficiency with minimal overhead.

LC-QAT achieves superior performance in 2-bit quantization with just a fraction of the training data, setting a new standard for data-efficient model optimization.