Arslan Chaudhry

Research Scientist @GoogleDeepMind | Ex @UniofOxford, @AIatMeta, @GoogleAI

[Arslan Chaudhry @arslan_mac]

Check out our new work on combining General Preference Modeling (GPMs) with GRPO style methods. Our proposed GRPL algo prevents reward hacking and improves reasoning quality in many open-ended tasks compared to the baselines.

My contributions were only advisory on this project.

[Ahmed Mohsin @ahmedmohsin7338]

[1/3] Happy to share our latest work, "General Preference Reinforcement Learning"

Link: https://t.co/ZLSn2dVVzU

[Arslan Chaudhry @arslan_mac]

@_rockt @Recursive_SI Congratulations, Tim! Looking forward to the exciting results.

[Arslan Chaudhry @arslan_mac]

Checkout our new work on exploring different inductive biases of causal vs bi-directional language modeling objectives vis-à-vis latent generalization (reversal curse). This work was led by @juliancodaforno as a student researcher in @GoogleDeepMind

[Julian Coda-Forno @juliancodaforno]

🎉 Excited to share our new paper from my SR work @GoogleDeepMind : "The Illusion of Latent Generalization: Bi-directionality and the Reversal Curse" . It was accepted at the ICLR 2026 Workshop on Representational Alignment (Re-Align)!

[Arslan Chaudhry @arslan_mac]

A few colleagues and friends are organizing a workshop at ICML in Seoul on the continuous adaptation of frontier models. Consider submitting your works to the workshop and/ or attend it if you are going to ICML.

Link: https://t.co/8yAblYtyLA

[Arslan Chaudhry @arslan_mac]

[8/8] There are a lot of other interesting details and experiments in the paper. Check it out: https://t.co/bYt6K2rSI3) and provide us with any feedback.

Joint work with @Sridhaar96 @AndrewLampinen

[Arslan Chaudhry @arslan_mac]

[1/8] Our new paper on using test-time compute to improve latent learning: https://t.co/bYt6K2rSI3

Thinking models have been shown to improve performance on maths, reasoning and coding tasks. In this work, however, we look into how thinking models improve latent learning.

[Arslan Chaudhry @arslan_mac]

[7/8] Yet, pure reversals lack intermediate reasoning chains, making them difficult for thinking models; here, ICL remains the best baseline. This suggests scaling compute alone cannot solve pure reversals, implying a need for architectural changes.

[Yann LeCun @ylecun]

@haider1 When pondering about the effects of technological revolutions on the labor market, listen to economists who have studied the effects of past technological revolutions on the labor market.
Never listen to computer scientists.

[Tim Rocktäschel @_rockt]

Wise words.

[Arslan Chaudhry @arslan_mac]

@realJessyLin @AIatMeta @LukeZettlemoyer @gargighosh @scottyih @aramHmarkosyan @vinceberges @barlas_berkeley Really nice work! Just a quick clarification, in your full-finetuning and LoRA baselines, do you also use (co-train with) those 1000 DCLM batches for a fair comparison of forgetting in those methods?

[Tim Althoff @timalthoff]

I’m excited to share our new @Nature paper 📝, which provides strong evidence that the walkability of our built environment matters a great deal to our physical activity and health.

Details in thread.🧵

https://t.co/omO3YcHrvG

[Arslan Chaudhry @arslan_mac]

@SharonYixuanLi Congrats, Sharon! 🎉

[Arslan Chaudhry @arslan_mac]

[4/4] The starting date is flexible but I would prefer it if you can start in July or August. The duration of the project would be six months.

[Arslan Chaudhry @arslan_mac]

[1/4] I am hiring a student researcher to work on topics related to continual learning, knowledge acquisition, and science of learning in LLMs this summer. This will be an in-person position and you will be based in Mountain View. https://t.co/jlgdZelgQG

[Arslan Chaudhry @arslan_mac]

[3/4] If you are interested, please send me your resume, brief snippet of your research, and what you would want to explore during the internship at [email protected]. Add [SRP 2025 Candidate] in your email subject.