🏢 Tsinghua University

CRAW4LLM: Efficiently crawls web pages for LLM pretraining by prioritizing influence scores, boosting data quality & cutting crawling waste.

PAFT dynamically adjusts prompts during LLM fine-tuning, improving model robustness and generalization across diverse prompts without sacrificing performance or efficiency.

video-SALMONN-01: An open-source audio-visual LLM enhances video understanding with a novel reasoning-intensive dataset and the pDPO method, achieving significant accuracy gains.

DexTrack achieves highly generalizable neural tracking control for dexterous robot manipulation by iteratively training a controller using high-quality demonstrations refined via homotopy optimization…

New benchmark COUNTERMATH enhances LLMs’ mathematical reasoning using counterexample-driven proofs, revealing current models’ limitations and paving the way for improved mathematical capabilities.

Steel-LLM: A fully open-source, resource-efficient Chinese LLM trained with transparency, achieving competitive performance despite limited resources.

EFFICIENT-VDIT accelerates video generation by 7.8x using sparse attention and multi-step distillation.

Smaller LLMs can outperform larger ones by strategically increasing computation during inference, defying conventional LLM scaling.

Ola: a novel 7B parameter omni-modal language model achieves state-of-the-art performance across image, video and audio tasks using a progressive modality alignment training strategy.

PRIME (Process Reinforcement through IMplicit rEwards) revolutionizes LLM training by efficiently using implicit process rewards from online policy rollouts and outcome labels, significantly boosting …

Human feedback boosts video generation! New VideoReward model & alignment algorithms significantly improve video quality and user prompt alignment, exceeding prior methods.

Pairwise RM, a novel reward model with knockout tournaments, significantly boosts large language model accuracy in test-time scaling by comparing solution pairs, eliminating arbitrary scoring inconsis…

FILMAGENT: A multi-agent framework automates end-to-end virtual film production using LLMs, exceeding single-agent performance in a collaborative workflow.

ComplexFuncBench, a new benchmark, rigorously evaluates LLMs’ complex function-calling abilities across real-world scenarios involving multi-step processes, constraints, and long contexts.

This survey paper explores the exciting new frontier of Large Reasoning Models (LRMs), focusing on how reinforcement learning and clever prompting techniques are boosting LLMs’ reasoning capabilities.

Parameter-Inverted Image Pyramid Networks (PIIP) drastically cut visual model computing costs without sacrificing accuracy by using smaller models for higher-resolution images and larger models for lo…

URSA-7B: A new multimodal model significantly improves chain-of-thought reasoning in mathematics!

This paper unveils critical thresholds for efficient visual autoregressive model computation, proving sub-quartic time is impossible beyond a certain input matrix norm while establishing efficient app…

EpiCoder revolutionizes code generation by using feature trees to create diverse and complex training data, resulting in state-of-the-art performance on various benchmarks.

MotionBench, a new benchmark, reveals that existing video models struggle with fine-grained motion understanding. To address this, the authors propose TE Fusion, a novel architecture that improves mo…