Our research on easy-to-hard generalization will be supported by the OpenAI Superalignment Fast Grant. Congratulations to the team and stay tuned!😎

Excited to announce the AI for Math Workshop at #ICML2024 ICML Conference! Join us for groundbreaking discussions on the intersection of AI and mathematics. 🤖🧮

📅 Workshop details: sites.google.com/view/ai4mathwo…

📜 Submit your pioneering work: sites.google.com/view/ai4mathwo…

🏆 Take on our…

Calling motivated students interested in pursuing MS/PhD in ML/AI, specifically privacy & generative AI!

The research group I'm starting at IIT Madras has openings!

Apply by *Mar 31* directly to Dept. of Data Science & AI, IIT Madras or IIT Madras CSE Dept. at research.iitm.ac.in!

🔥I will be joining Carnegie Mellon University Language Technologies Institute | @CarnegieMellon this upcoming Fall, working with Graham Neubig and Sean Welleck on evaluating LLMs & improving them with (human) feedback!

Can't wait to explore what lies ahead during my Ph.D. journey☺️

As a side product, we developed `gpt-accelera`, which supports batched inference with torch.compile during RL, and 2-D parallelism (tp + fsdp) training that can scale to 34b models. We perform all the training in full fine-tuning with this codebase. (n/n)

github.com/Edward-Sun/gpt…

Whoa this looks neat.

It's often said that 'evaluation is easier than generation'...

We go one step further: strong evaluators enable generalizing to harder problems! New paper led by Zhiqing Sun and Longhui Yu

Using supervision only on easy problems, 52.5 on MATH with Llemma-34b + re-ranking

Wanna know gpt-3.5-turbo's embed size? We find a way to extract info from LLM APIs and estimate gpt-3.5-turbo’s embed size to be 4096. With the same trick we also develop 25x faster logprob extraction, audits for LLM APIs, and more!
📄 arxiv.org/abs/2403.09539
Here’s how 1/🧵

𝗛𝗼𝘄 𝗰𝗮𝗻 𝘄𝗲 𝗯𝘂𝗶𝗹𝗱 𝗺𝗼𝗿𝗲 𝗿𝗲𝗹𝗶𝗮𝗯𝗹𝗲 𝗟𝗠-𝗯𝗮𝘀𝗲𝗱 𝘀𝘆𝘀𝘁𝗲𝗺𝘀? Our new position paper advocates for retrieval-augmented LMs (RALMs) as the next gen. of LMs, exploring the promises, limitations, and a roadmap for wider adoption.
arxiv.org/abs/2403.03187 🧵

Another day, another math LM bootstrapping their data work off of the work done by OpenWebMath and Llemma teams. This makes the three in the past week!

Open data work is 🔥 Not only do people use your data, but high quality data work has enduring impact on data pipelines.

Self-RAG in LlamaIndex 🦙

We’re excited to feature Self-RAG, a special RAG technique where an LLM can do self-reflection for dynamic retrieval, critique, and generation (@AkariAsai et al.).

It’s implemented in LlamaIndex 🦙 as a custom query engine with…

If you're interested in this direction, check out our paper Inference-Time Policy Adapters (IPA🍺).

IPA guides a frozen LLM such as GPT-3 during decoding time through a lightweight policy adapter trained to optimize an arbitrary user objective with RL.

Teaching a new course on Neural Code Generation with Daniel Fried!

cmu-codegen.github.io/s2024/

Here is the lecture on pretraining and scaling laws:
cmu-codegen.github.io/s2024/static_f…

Proud of this work. Here's my 22min video explanation of the paper: youtube.com/watch?v=TuZhU1…

Wrote a summary of my thoughts on the plane back from NeurIPS: albertqjiang.github.io/thoughts/

That's the serious stuff. Will do a thread of silly things later.

Thesis Review throwback (May 2021):

Tomas Mikolov on word2vec, RNN-LMs, and more

Had a great time at the NeurIPS instruction workshop and we are honored to get the Best paper Honorable Mention Award!

Check the details of the paper here:
selfrag.github.io