Chi-Heng Lin 貓貓🐱

Meta / 布瑠部由良由良 🐱🐱🐱🐱🐱

[Chi-Heng Lin 貓貓🐱 @ChiHengHenry]

🔥🔥Thrilled to share our work @iclr_conf🔥🔥

🤔 Low-rank decomposition for LLM compression adds significant overhead by splitting 1 matrix into 2.

🔑 Key contribution:
✅ Introduction of decomposition-compression duality: By leveraging the duality, we compress two adjoint matrices within a module, eliminating parameter overhead and achieving state-of-the-art performance!

🚀: w/@jamessealesmith @TuliShikhar
📜: https://t.co/R9Zm7hd9AZ

[Binfeng Xu @billxbf]

Today we give Superintelligence back to its owners.
Introducing Skill2Env 🎨 , the most aligned and diverse dataset to fuel modern Agentic RL . 🧵

[Pedro Domingos @pmddomingos]

I’ve never seen an LLM come up with a new concept.

[Pedro Domingos @pmddomingos]

RL only works if the pretrained model can already solve the problem (otherwise there's no reward signal). So RL can't solve any hard problems, and when it appears to it's just thinly disguised brute force.

[Binfeng Xu @billxbf]

Surprised how opinions converge at NeurIPS.
Now it’s (sadly) all just data game.
We are all excited about scientific discovery.
Long term, we all need (new) architecture for continuous learning/reasoning.
I feel lucky to see great minds like @doyend @YejinChoinka @vwxyzjn who know the drill yet don’t get trapped by it.

[Jason Lee @jasondeanlee]

Gd dominates ridge regression but not sgd! Surprising that such simple things aren't known for linear regression

[Xinyu Yang @Xinyu2ML]

Thank @_akhaliq for introducing our Parallel-R1! 😀

Parallel thinking has recently emerged as a promising direction for enhancing model performance. However, most existing inference-time approaches rely on heuristic rules to aggregate results, without enabling the model to truly understand or utilize parallel thinking itself. In contrast, our Parallel-R1 introduces the first RL framework that empowers the model to actively explore and internalize parallel thinking. We hope that our work will lead to further studies in this novel direction.

[Chi-Heng Lin 貓貓🐱 @ChiHengHenry]

@billxbf The guy definitely had read death note….right?😅

[Tri Dao @tri_dao]

I played w it for 1h. Went through my usual prompts (math derivations, floating point optimizations, …). It’s a good model, feels comparable to the best frontier models

[hardmaru @hardmaru]

Text-to-LoRA: Instant Transformer Adaption

https://t.co/FXTZUFd0Jo

Generative models can produce text, images, video. They should also be able to generate models! Here, we trained a Hypernetwork to generate new task-specific LoRAs by simply describing the task as a text prompt.

[Turing Post (Ksenia Se) @TheTuringPost]

Paper: https://t.co/J24Kk3Gxk8
Code: https://t.co/3lGvWZZgXG

@MIT, @CarnegieMellon, @togethercompute

[Chi-Heng Lin 貓貓🐱 @ChiHengHenry]

@TivadarDanka Modern proof technique: solved by major ity votes

[Tivadar Danka @TivadarDanka]

I'm a mathematician and an AI optimist. Please explain to me how do you "solve math". What's the solution? What's the problem that is solved?

[joowon @n0w00j]

she said my TC is too low

[Chi-Heng Lin 貓貓🐱 @ChiHengHenry]

@PolSciEnjoyer @miniapeur Yeah that’s true. They taste almost like completely different stuffs

[Chi-Heng Lin 貓貓🐱 @ChiHengHenry]

@behrouz_ali ok, click "I acknowledge that the authors do not want to be acknowledged for being ghosted."

[AK @_akhaliq]

Meta presents Multi-Token Attention

demonstrate that MTA achieves enhanced performance on a range of popular benchmarks. Notably, it outperforms Transformer baseline models on standard language modeling tasks, and on tasks that require searching for information within long contexts, where our method's ability to leverage richer information proves particularly beneficial.

[Chi-Heng Lin 貓貓🐱 @ChiHengHenry]

@zainhas 😀👍👍

[Zain @zainhas]

they tested sota LLMs on 2025 US Math Olympiad hours after the problems were released

Tested on 6 problems and spoiler alert!

They all suck -> 5%

[Chi-Heng Lin 貓貓🐱 @ChiHengHenry]

@sang_yun_lee Nice pic 😆

[Chi-Heng Lin 貓貓🐱 @ChiHengHenry]

@Andr3jH Wow he did look aged lolll