🚀 New Paper Alert! 🧠 Simple Denoising Diffusion Language Models (SDDLMs) We simplify the complex ELBO objectives in Uniform-State Diffusion Models with a simple denoising loss, making training more scalable — while matching or surpassing baseline generation quality. (1/2)

Want to get an LLM agent to succeed in an OOD environment? We tackle the hardest case with SPA (Self-Play Agent). No extra data, tools, or stronger models. Pure self-play. We first internalize a world model via Self-Play, then we learn how to win by RL. Like a child playing

I will attend #ICLR2025 to present our papers on LLM alignment, RL, and RLHF. Looking forward to meet you at our posters! On a Connection Between Imitation Learning and RLHF 📍 Hall 3 + Hall 2B #223 🕙 Saturday, Apr 26 at 10:00 (+08) SimPER: A Minimalist Approach to Preference

Huge thanks to my incredible collaborators: @1t4chiii @learn2dropout, Mingxiao Li, Shangsong Liang, @zhaochun_ren, and @vhonavar! This is a huge milestone, and we’re thrilled to see SimPER ( https://t.co/ClL23HoefG) contributing to cutting-edge advancements in LLM reasoning.

🚀 Incredible news! Our SimPER accepted by ICLR 2025 has been used as the training algorithm in a recent paper EXAONE Deep: Reasoning Enhanced Language Models ( https://t.co/mTst9UN4h5) from LG AI—and their open model achieved better and competitive performance compared to QwQ

🚀 Breaking News! We’re thrilled to introduce #EXAONEDeep, a next-generation AI model designed to enhance reasoning capabilities—Evolving into ‘Agentic AI‘ for real-world industry solutions! 🧠 Specialized in math, science, and coding tasks, EXAONE Deep pushes the boundaries of

🚀 Is RLHF really RL… or barely RL? 🤔 Excited to share our #ICLR2025 paper: "On a Connection Between Imitation Learning and RLHF" 🎯 🔍Our work suggests that RLHF is more closely related to imitation learning than RL, offering new insights into the alignment process. In this

❓ Can we align large language models without hyperparameters? The answer is SimPER! 🚀 Excited to share our ICLR paper: SimPER: A Minimalist Approach to Preference Alignment without Hyperparameters! 🎯 SimPER eliminates the need for complex tuning and reference models,

In our EMNLP2024 paper, "How to Leverage Demonstration Data in Alignment for Large Language Models? A Self-Imitation Learning Perspective" ( https://t.co/quniEQRdiN), we also propose GSIL, a imitation learning (IL) approach that eliminates the need for complex adversarial training

🚀 Excited to share that our new paper, Cal-DPO, focused on LLM alignment, has been accepted to #NeurIPS2024! 🎉 We empirically and theoretically demonstrate that substantial improvements over DPO can be achieved by calibrating implicit rewards to align with absolute reward

Extensive experiments show that GSIL consistently and significantly outperforms baselines in many challenging benchmarks, such as coding, mathematical reasoning and instruction-following benchmark. Code will be public available at