Reshinth

Research | Language Models

[Reshinth @reshinth_]

How to define Diversity in the context of CodeLMs and Programming Languages ?
1. Diversity is positively correlated with Performance in solving a problem.
2. Shortcomings of diversity in small codeLMs.
3. Code Embedding models don't capture semantics.
https://t.co/aNGhSpxz48

[Reshinth @reshinth_]

ChatGPT users, don't y'all feel the model is too verbose ? How to fix it ? I couldn't stop but notice after using Claude for a while.

[Reshinth @reshinth_]

@eisokant where can I read more about "Model Factory" ?

[Reshinth @reshinth_]

@herbiebradley Woah !

[Reshinth @reshinth_]

EOD

[Reshinth @reshinth_]

👀

[Reshinth @reshinth_]

I wonder if an extension that measures the bits of information consumed in twitter might be useful. It sounds useful atleast to me.🤨

[humans& @humansand]

At humans&, we train models from the long-term impacts of their interactions with people. This requires prioritizing long-horizon multi-agent RL. We've developed and are excited to share an open-source, hardware-native 4-bit RL recipe, significantly accelerating training

[Reshinth @reshinth_]

I wonder if an extension that measures the bits of information consumed in twitter might be useful. It sounds useful atleast to me.🤨

[Jane Zhang @jjanezhang]

My June essay is out now!🐟

It is titled: It’s time to do the harder things

I discuss how media companies like the Discovery Channel once captured beautiful wildlife shots but now create unscripted reality TV shows like Deadliest Catch because it is cheaper and faster to produce

But with AI, I argue that now it is time for us to go back to doing the harder things

[changyi yang @YangYi17826]

Weight sync in RL post-training looks RDMA-bound. But >99% of BF16 weights are byte-identical step to step — ship only the delta: ~100–1000× less traffic, bit-exact. Hardware bottleneck → software problem: RL over commodity ethernet.
🧵 https://t.co/ivniFyzPo1

[Reshinth @reshinth_]

On that note, I have a blogpost coming on a research topic 🔜.

[Reshinth @reshinth_]

👀

[Reshinth @reshinth_]

These are small scale GRPO runs with Qwen-1.7B-Base on MATH task wherein I consistently observed optimizers with no velocity like Muon, Signum work comparably. Is this something studied in scale ? Are there public info on such large scale runs ?

[Reshinth @reshinth_]

👀

[CiaraRowles @CiaraRowles1]

We're excited to share Stable-Layers!

We train Qwen-Image-Layered further with RL for improved layerization,
using only feedback from a VLM — no paired supervision required!

Paper: https://t.co/WktmmXGNLh
Project Page: https://t.co/WnEV76afQp

[Varun Jampani @jampani_varun]

🎬 Introducing Stable Cinemetrics, to be presented at NeurIPS 2025.

We present the first taxonomy of professional controls to systematically study and control video generative models through the lens of filmmaking.

Interactive webpage with paper link: https://t.co/Eh4Hw3hBZl
🧵

[Reshinth @reshinth_]

@tylerangert Happy Birthday 🎉

[Xiao Liang @MasterVito0601]

🙋‍♂️ Can RL training address model weaknesses without external distillation?

🚀 Please check our latest work on RL for LLM reasoning!

💯 TL;DR: We propose augmenting RL training with synthetic problems targeting model’s reasoning weaknesses.

📊Qwen2.5-32B: 42.9 → SwS-32B: 68.4

[Reshinth @reshinth_]

@evanthebouncy Congrats Evan.

[Reshinth @reshinth_]

With CodeLMs scaling actually solved models intrinsically learning internal structural syntactical & semantic information.