Dual-Stream Transformer Automatically Detects Mutual Gaze and Joint Attention
A research team has proposed an efficient dual-stream Transformer architecture that can automatically detect mutual gaze…
📂 Research
Latest articles in Research
New Study Decouples the True Contributions of Subword Tokenization to Large Language Model Training
A latest arXiv paper systematically disentangles the specific contributions of Subword Tokenization to large language mo…
Proactive Dialogue Models: Ending Redundant Interactions with Intent Prediction
A new study proposes a proactive dialogue model based on Temporal Bayesian Networks that predicts users' upcoming intent…
New Research on LLM Reasoning Controllability: The Tug-of-War Between Compliance and Sensibility
A new arXiv paper explores whether fundamental reasoning modes such as deduction, induction, and abduction can be decoup…
When 2D Tasks Meet 1D Serialization: LLMs' Structural Friction Problem
A new arXiv paper introduces the concept of "Serialization Friction," revealing the performance degradation that occurs …
PLE Architecture: Solving the Reasoning Leakage Problem in Hybrid Thinking Models at Its Root
A latest arXiv paper proposes the Path-Lock Expert (PLE) method, which employs an architecture-level separation mechanis…
The Semantic Structure of LLM Feature Spaces Turns Out to Be Highly Aligned with Human Psychology
A new study reveals that the geometric relationships of semantic features in the hidden states of large language models …
New Study Uses Minimal Translation Pairs to Precisely Diagnose the Linguistic Capabilities of Sign Language AI Models
A latest arXiv paper proposes using minimal translation pairs to systematically evaluate how well sign language models c…
Selective Augmentation Method Improves Universal Automatic Phonetic Transcription Accuracy
Researchers propose a "Selective Augmentation" bootstrapping method that selectively transfers distinctive features acro…
Safe but Useless? New Benchmark Exposes the LLM Alignment Dilemma
A research team has introduced CarryOnBench, the first benchmark to systematically evaluate whether large language model…
Length Value Model: A New Framework for Token-Level Length Modeling
Researchers propose the Length Value Model (LenVM), which for the first time models the generation length of autoregress…
Exploring the Limits of Pruning: Task-Specific Neurons and Model Collapse
A new study systematically explores the limits of neuron pruning in large language models, revealing for the first time …