Lightnews — Scholar-powered news

Apoorv Khandelwal @apoorvkh.com · Jul 28

aclanthology.org/2025.conll-1...

What is an “Abstract Reasoner”? Revisiting Experiments and Arguments about Large Language Models

Tian Yun, Chen Sun, Ellie Pavlick. Proceedings of the 29th Conference on Computational Natural Language Learning. 2025.

aclanthology.org

Apoorv Khandelwal @apoorvkh.com · Jul 28

Will be at ACL this week! #ACL2025 #ACL2025NLP

Presenting Tian Yun’s paper on abstract reasoners at CoNLL on Thursday.

I’ve been investigating how LLMs internally compose functions lately. Happy to chat about that (among other things) and hang out in Vienna!

1

Apoorv Khandelwal @apoorvkh.com · May 29

Curious how many papers were assigned to reviewers on average! Review quality seems better than average from my small sample size. Wondering if that correlates with a lower reviewer load? E.g. I only received 2 papers to review.

Reposted by Apoorv Khandelwal

Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · Mar 31

Tests on USAMO immediately after problems were posted yield surprisingly bad model performance. Suggests there's much more training on test than expected.
arxiv.org/abs/2503.219...

Scores of R1, Flash-thinking, Claude 4.7, QwQ, o1-pro, o3-mini on USAMO 2025. Scores less than 5% of max score.

7 8 29

Reposted by Apoorv Khandelwal

Dr Sasha Luccioni @sashamtl.bsky.social · Mar 19

Just read that AI’s energy consumption in data centers is nothing to be worried about because most of the hyperscale datacenters running AI are "powered by renewable energy or low-carbon nuclear power."

Let's debunk that, shall we?

2 12 30

Reposted by Apoorv Khandelwal

Naomi Saphra @nsaphra.bsky.social · Mar 19

If you're in the northeastern US and you're submitting a paper to COLM on March 27, you should absolutely be sending its abstract to New England NLP on March 28.

New England NLP Meeting Series

nenlp.github.io

3 7

Apoorv Khandelwal @apoorvkh.com · Mar 11

+ No system pre-reqs, multi-stage PyTorch workflows in one script, CLI integrations, catching system failures as exceptions, SLURM support, better logging, and so much more!

Additional fine-tuning examples in our docs with:
@pytorch.org, Deepspeed, @lightningai.bsky.social, HF Accelerate

Apoorv Khandelwal @apoorvkh.com · Mar 11

A cool side-effect: fine-tune any LLM (from
@huggingface
transformers) on any text dataset *with multiple nodes* in just *one command*.

torchrun.xyz/examples/tra...

1 2

Apoorv Khandelwal @apoorvkh.com · Mar 11

It's a replacement for CLI tools, like "torchrun".

Most basic usage: specify some (SSH-enabled) machines you want to parallelize your code on. Then launch a function onto that configuration.

All from inside your Python script!

1 1

Apoorv Khandelwal @apoorvkh.com · Mar 11

We made a library (torchrunx) to make multi-GPU / multi-node PyTorch easier, more robust, and more modular! 🧵

github.com/apoorvkh/tor...
Docs: torchrun.xyz

`(uv) pip install torchrunx` today!

(w/ the very talented, Peter Curtin, Brown CS '25)

GitHub - apoorvkh/torchrunx: Easily run PyTorch on multiple GPUs & machines

Easily run PyTorch on multiple GPUs & machines. Contribute to apoorvkh/torchrunx development by creating an account on GitHub.

github.com

1 1 4

Reposted by Apoorv Khandelwal

William Merrill @lambdaviking.bsky.social · Mar 7

✨How does the depth of a transformer affect its reasoning capabilities? New preprint by myself and @Ashish_S_AI shows that a little depth goes a long way to increase transformers’ expressive power

We take this as encouraging for further research on looped transformers!🧵

Paper: A Little Depth Goes a Long Way: The Expressive Power of Log-Depth Transformers

1 2 12

Reposted by Apoorv Khandelwal

Sonia Murthy @soniakmurthy.bsky.social · Feb 10

(1/9) Excited to share my recent work on "Alignment reduces LM's conceptual diversity" with @tomerullman.bsky.social and @jennhu.bsky.social, to appear at #NAACL2025! 🐟

We want models that match our values...but could this hurt their diversity of thought?
Preprint: arxiv.org/abs/2411.04427

2 10 63

Apoorv Khandelwal @apoorvkh.com · Feb 7

I started a blog! First post is everything I know about setting up (fast, reproducible, error-proof) Python project environments using the latest tools. These methods have saved me a lot of grief. Also a short guide to CUDA in appendix :)

blog.apoorvkh.com/posts/projec...

Managing Project Dependencies

blog.apoorvkh.com

3

Apoorv Khandelwal @apoorvkh.com · Jan 25

I think typing my code and using a linter (ruff) + static type checker (pyright) saves me a lot of grief.

1

Reposted by Apoorv Khandelwal

James Tompkin @jamestompkin.bsky.social · Jan 10

Can GANs compete in 2025? In 'The GAN is dead; long live the GAN! A Modern GAN Baseline', we show that a minimalist GAN w/o any tricks can match the performance of EDM with half the size and one-step generation - github.com/brownvc/r3gan - work of Nick Huang, @skylion.bsky.social, Volodymyr Kuleshov

3 14 69

Apoorv Khandelwal @apoorvkh.com · Jan 10

A couple sources for academic talks that I really like!

Cohere For AI (www.youtube.com/playlist?lis...)

Simons Institute (www.youtube.com/@SimonsInsti...)

Simons Institute

The Simons Institute for the Theory of Computing is the world's leading venue for collaborative research in theoretical computer science. Established on July 1, 2012, the Institute is housed in Calvin...

www.youtube.com

2

Reposted by Apoorv Khandelwal

Naomi Saphra @nsaphra.bsky.social · Dec 18

Let he who hath not \usepackage[subtle]{savetrees}

1 1 13

Reposted by Apoorv Khandelwal

Jennifer Hu @ COLM (recruiting PhDs and postdocs!) @jennhu.bsky.social · Dec 11

Slides from the tutorial are now posted here!

neurips.cc/media/neurip...

neurips.cc

7 17

Reposted by Apoorv Khandelwal

Alexander Doria @dorialexander.bsky.social · Dec 5

“They said it could not be done”. We’re releasing Pleias 1.0, the first suite of models trained on open data (either permissibly licensed or uncopyrighted): Pleias-3b, Pleias-1b and Pleias-350m, all based on the two trillion tokens set from Common Corpus.

11 85 250

Apoorv Khandelwal @apoorvkh.com · Nov 27

I am an ex-Paperpile user and am liking Zotero lately! Free storage from the university helps.

Reposted by Apoorv Khandelwal

Ben Lipkin @benlipkin.bsky.social · Nov 25

Lots of folks talking about scaling LLM inference over this last year

Internally, I’ve been developing and using a library that makes this extremely easy, and I decided to open-source it
Meet the decoding library: github.com/benlipkin/de...

1/7

GitHub - benlipkin/decoding: Composable inference algorithms with LLMs and programmable logic

Composable inference algorithms with LLMs and programmable logic - benlipkin/decoding

github.com

1 5 26

Apoorv Khandelwal @apoorvkh.com · Nov 26

“Turn” a decoder into an encoder with LLM2Vec (github.com/McGill-NLP/l...). Seen at COLM 2024 :)

If you want the naive, training-free / model-agnostic approach: their related work section says it is most common to using the final token’s last hidden state.

GitHub - McGill-NLP/llm2vec: Code for 'LLM2Vec: Large Language Models Are Secretly Powerful Text Encoders'

Code for 'LLM2Vec: Large Language Models Are Secretly Powerful Text Encoders' - McGill-NLP/llm2vec

github.com

1 14

Reposted by Apoorv Khandelwal

Joe Stacey @joestacey.bsky.social · Nov 24

Okay genius idea to improve quality of #nlp #arr reviews. Literally give gold stars to the best reviewers, visible on open review next to your anonymously ID during review process.

Here’s why it would work, and why would you should RT this fab idea:

3 5 27

Apoorv Khandelwal @apoorvkh.com · Nov 21

Thanks and great! Hope you are likewise doing well!

1 1

Apoorv Khandelwal @apoorvkh.com · Nov 21

Would be great to join, thanks!

1 1