🏢 Peking University

LoRA-Inlaid: a novel multi-task LLM serving system boosts throughput by 1.58x, latency by 1.76x, and job completion time by 2x, while improving SLO attainment by 10x, all while maintaining model quali…

Du-IN: Revolutionizing speech decoding from intracranial neural signals with discrete units-guided mask modeling, achieving state-of-the-art performance on a challenging Chinese word-reading sEEG dat…

DGNet enhances weakly supervised point cloud segmentation by aligning feature embeddings to a mixture of von Mises-Fisher distributions, achieving state-of-the-art performance.

Cross-environment Intra-class Alignment (CIA) and its label-free variant, CIA-LRA, significantly improve node-level OOD generalization on graphs by aligning representations and eliminating spurious fe…

DeltaDock: a novel two-stage framework revolutionizes molecular docking by improving accuracy and reliability, achieving a 300% increase in success rate compared to the prior state-of-the-art in blind…

Delta-CoMe: Training-free mixed-precision delta compression boosts LLM deployment efficiency.

Spiking neural network effectively segments mixed-motion event streams via spike-based Bayesian computation, achieving efficient real-time motion decoupling.

This work unveils a novel algorithm for contextual decision-making with knapsacks, achieving significantly improved regret bounds beyond worst-case scenarios, thereby offering a more practical and eff…

Statistical Flow Matching (SFM) uses information geometry to create a new flow-matching framework for generating discrete data, achieving superior sampling quality and likelihood compared to existing …

GNNs enhance LLM-based task planning by improving the ability to process task graphs, surpassing existing solutions even without training.

Buffer of Thoughts (BoT) boosts Large Language Model reasoning by storing and reusing high-level ’thought-templates’, achieving significant accuracy and efficiency gains across diverse tasks.

Geometric Diffusion Bridge (GDB) accurately predicts geometric state evolution in complex systems by leveraging a probabilistic approach and equivariant diffusion processes, surpassing existing deep l…

AutoSurvey automates comprehensive literature survey creation using LLMs, overcoming challenges of context limitations and knowledge constraints via a novel, efficient, and rigorously evaluated method…

Boosting AI’s math skills, this paper introduces a novel framework for autoformalizing mathematical statements, improving accuracy by 0.22-1.35x via symbolic equivalence and semantic consistency check…

ARKVALE boosts LLM inference efficiency by intelligently evicting and recalling key-value pairs from cache, improving latency and throughput without significant accuracy loss.

EMDiffusion trains clean diffusion models from corrupted data using an expectation-maximization algorithm, achieving state-of-the-art results on diverse imaging tasks.

Aligner efficiently aligns LLMs by learning to correct initial responses, achieving significant improvements in helpfulness and harmlessness across various models with resource efficiency.

Adversarial Representation Engineering (ARE) offers a unified, interpretable approach for editing large language models (LLMs) by using a representation sensor as an editing oracle, enhancing model sa…

MetaDebias tackles hidden confounding in recommender systems using heterogeneous observational data, achieving state-of-the-art performance without expensive RCT data.

Trifle: Tractable inference for Offline RL achieves state-of-the-art results by using tractable generative models to overcome the inference-time suboptimality of existing sequence modeling approaches.