🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

The LMSYS Chat dataset hasn’t been updated in over a year, yet recent models consistently top the leaderboard early on. Are rankings being manipulated? Analyzing model selection trends over time could reveal the right metric for fair LLM evaluation... Wanna check the recent

I even believe that RL (or Regret training) is needed before SFT - or even pertaining phase -- if you worked on RL with LLM, you will know what I am saying. Super nice to read this paper!

One of the biggest cultural differences between East Asians and Americans is the concept of "face" (面子, 체면)—a nuanced idea that encompasses social reputation, honor, and maintaining harmony in interpersonal relationships, with no direct English equivalent. This distinction

Cool definition :)

Amazing... I should use muP..

Pretraining with RL for reasoning. Love this approach.

Nice approach. Good to see that a16z sees the future in social simulations.

I interviewed for an ML research internship at Meta (FAIR) a few years back. Don’t remember every detail now, but a few questions stuck with me. Questions are below.

My mother uses ChatGPT and trades ETH—no surprise there! I really hope South Korea can emerge as a global leader across key tech domains like robotics, AI, crypto, etc. On that note, Korea also offers strong government-backed investment support for startups, which is a great

🚨 New Paper Alert! 🚨 How can we align language models without drowning in prompt engineering or falling into reward hacking traps? We introduce Meta Policy Optimization (MPO)—a new reinforcement learning framework that evolves its own reward model rubrics through meta-level

That is the reason you need an evolving reward function. huggingface.co/papers/2504.20… Check out this paper. -- providing some answers about "curriculum learning / evolving reward / reward hacking" using evaluative thinking. Zae Myung Kim

This is a banger. I read it thoroughly and now understand large model training better.

🤯 We cracked RLVR with... Random Rewards?! Training Qwen2.5-Math-7B with our Spurious Rewards improved MATH-500 by: - Random rewards: +21% - Incorrect rewards: +25% - (FYI) Ground-truth rewards: + 28.8% How could this even work⁉️ Here's why: 🧵 Blogpost: tinyurl.com/spurious-rewar…

Interesting perspective. Misspecification does matter.

Amazing approach on TTT for self-improvement. Looking forward to the followup paper :)