Yuda Song

Training with more data = better LLMs, right? 🚨

False! Scaling language models by adding more pre-training data can decrease your performance after post-training!
Introducing "catastrophic overtraining." 🥁🧵👇

arxiv.org/abs/2503.19206

1/10

1.5 yrs ago, we set out to answer a seemingly simple question: what are we *actually* getting out of RL in fine-tuning? I'm thrilled to share a pearl we found on the deepest dive of my PhD: the value of RL in RLHF seems to come from *generation-verification gaps*. Get ready to 🤿:

super happy about this preprint! we can *finally* perform efficient exploration and find near-optimal stationary policies in infinite-horizon linear MDPs, and even use it for imitation learning :) working with @neu-rips.bsky.social and @lviano.bsky.social on this was so much fun!!

What are the minimal supervised learning primitives required to perform RL efficiently?

New paper led by my amazing intern Dhruv Rohatgi:

Necessary and Sufficient Oracles: Toward a Computational Taxonomy for Reinforcement Learning

arxiv.org/abs/2502.08632

1/

Mind the Gap: Examining the Self-Improvement Capabilities of Large Language Models | alphaXiv View 3 comments: Delete the space?

Models can self-improve🥷 by knowing they were wrong🧘‍♀️ but when can they do it?

Across LLM families, tasks and mechanisms
This ability scales with pretraining, prefers CoT, non QA tasks and more in 🧵

alphaxiv.org/abs/2412.02674
@yus167.bsky.social @shamkakade.bsky.social
📈🤖
#NLP #ML

On Saturday I will present our LLM self-improvement paper in the workshop on Mathematics of Modern Machine Learning (M3L) and the workshop on Statistical Foundations of LLMs and Foundation Models (SFLLM).
bsky.app/profile/yus1...

The Importance of Online Data: Understanding Preference Fine-tuning via Coverage Learning from human preference data has emerged as the dominant paradigm for fine-tuning large language models (LLMs). The two most common families of techniques -- online reinforcement learning (RL) ...

Arxiv link for HyPO: arxiv.org/abs/2406.01462

I will present two papers at #NeurIPS2024!

Happy to meet old and new friends and talk about all aspects of RL: data, environment structure, and reward! 😀

In Wed 11am-2pm poster session I will present HyPO-- best of both worlds of offline and online RLHF: neurips.cc/virtual/2024...

Mind the Gap: Examining the Self-Improvement Capabilities of Large Language Models Self-improvement is a mechanism in Large Language Model (LLM) pre-training, post-training and test-time inference. We explore a framework where the model verifies its own outputs, filters or reweights...

There are many more intriguing results that I can not fit into one post! For more details, please check out our paper: arxiv.org/abs/2412.02674. This is joint work with amazing collaborators Hanlin Zhang, Carson Eisenach, @shamkakade.bsky.social , Dean Foster and @ughai.bsky.social. (9/9)

We also dive deep into the similarity and difference between different verification mechanisms. We observed the consistency, distinction and ensemble properties of the verification methods (see the summary image). (8/9)

In iterative self-improvement, we observe the gap diminishes to 0 in a few iterations, resembling many previous findings. We discovered that one cause of such saturation is the degradation of the "effective diversity" of the generation due to the imperfect verifier. (7/9)

However, self-improvement is not always possible on all tasks. We do not observe significant self-improvement signal on QA tasks like Natural Questions. Also, not all models can self-improve on sudoku, a canonical example of "verification is easier than generation". (6/9)

Our first major result is an observational scaling law: with certain verification methods, the relative gap increases monotonically (almost linear) to the log of pretrain flops, on tasks like GSM8K and MATH. (5/9)

We propose to use the performance difference between the reweighted and original responses (2-1) -- the "generation-verification gap". We also study the relative gap -- gap weighted by the error rate. Intuitively, improvement is harder if the model makes fewer mistakes. (4/9)

While previous works measure self-improvement using the performance difference between the models (3-1), we found out that step 3 (distillation) introduces confounders (for example, the models can just be better at following certain formats). (3/9)

We study self-improvement as the following process:
1. Model generates many candidate responses.
2. Model filters/reweights responses based on its verifications.
3. Distill the reweighted responses into a new model.
(2/9)

LLM self-improvement has critical implications in synthetic data, post-training and test-time inference. To understand LLMs' true capability of self-improvement, we perform large-scale experiments with multiple families of LLMs, tasks and mechanisms. Here is what we found: (1/9)

Yuda Song, Hanlin Zhang, Carson Eisenach, Sham Kakade, Dean Foster, Udaya Ghai
Mind the Gap: Examining the Self-Improvement Capabilities of Large Language Models
https://arxiv.org/abs/2412.02674

I think the main difference in terms of interpolation / extrapolation between DPO and RLHF is that the former only guarantees closeness to the reference policy on the training data, while RLHF usually tacks on an on-policy KL penalty. We explored this point in arxiv.org/abs/2406.01462.

(1/n) 💡How can we speed up the serial runtime of long pre-training runs? Enter Critical Batch Size (CBS): the tipping point where the gains of data parallelism balance with diminishing efficiency. Doubling batch size halves the optimization steps—until we hit CBS, beyond which returns diminish.

I created a starter pack for people who are or have been affiliated with the Machine Learning Department at CMU. Let me know if I missed someone!

go.bsky.app/QLTVEph

#AcademicSky

Ojash Neopane, Aaditya Ramdas, Aarti Singh
Logarithmic Neyman Regret for Adaptive Estimation of the Average Treatment Effect
https://arxiv.org/abs/2411.14341

Intro 🦋

I am a final-year PhD student from CMU Robotics. I work on humanoid control, perception, and behavior in both simulation and real life, using mostly RL:

🏃🏻PHC: zhengyiluo.com/PHC
💫PULSE: zhengyiluo.com/PULSE
🔩Omnigrasp: zhengyiluo.com/Omnigrasp
🤖OmniH2O: omni.human2humanoid.com

Hi Bsky people 👋 I'm a PhD candidate in Machine Learning at Carnegie Mellon University.
My research focuses on interactive AI, involving:
🤖 reinforcement learning,
🧠 foundation models, and
👩‍💻 human-centered AI.

Also a founding co-organizer of the MineRL competitions 🖤 Follow me for ML updates!