Resources

Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?

Shiyu Ni

December 16, 2025

The existing mechanism and limitations of RLVR, as well as how to use RLVR training to develop models that surpass human-level performance.

View Slides View Blog
OneRec Technical Report

Lulu Yu

October 28, 2025

OneRec-V1, OneRec-V2 and OneRec-Think.

View Slides View Blog
Llama See, Llama Do: A Mechanistic Perspective on Contextual Entrainment and Distraction in LLMs

Minghao Tang

October 14, 2025

Llama See Llama Do, an outstanding paper at ACL 2025, finds a novel mechanistic phenomenon termed contextual entrainment, which LMs assign higher logits and probabilities to tokens that appear in the context.

View Slides View Blog
Why Language Models Hallucinate

Jiahan Chen

September 16, 2025

View Slides
Evaluating Implicit Bias in Large Language Models by Attacking From a Psychometric Perspective

Yuchen Wen

September 05, 2025

A psychometrics-inspired attack framework is proposed to reveal implicit biases in large language models(LLMs) via three attack methods, Disguise, Deception, and Teaching.

View Slides View Blog
Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation

Minzhu Tu

July 25, 2025

监控推理模型的不当行为及强化学习中隐蔽奖励作弊的风险

View Slides View Blog
CLIP-Based Fine-Grained Image Retrieval Methods

Jiahan Chen

July 17, 2025

基于CLIP的细粒度图像检索方法

View Slides View Blog
Massive Values in Self-Attention Modules are the Key to Contextual Knowledge Understanding

Shiyu Ni

July 06, 2025

自注意力模块中的大值是理解上下文知识的关键

View Slides View Blog
Adversarial Attack Survey

Yingshuo Wang

April 08, 2025

View Slides
Parametric Retrieval Augmented Generation

Minghao Tang

March 26, 2025

Parametric RAG proposes a novel paradigm integrating external knowledge as parameters into the feed-forward networks of LLMs.

View Slides
Evaluating Retrieval Quality in Retrieval-Augmented Generation

Minghao Tang

September 24, 2024

eRAG, SIGIR 2024 best short paper, proposes a new evaluation approach for RAG, using LLM to process each retrieved document individually and evaluates outputs by downstream task labels.

View Slides