Saurabh Dash

Bottomless pit supervisor. ML @CohereAI , PhD Student @GeorgiaTech. Previously @Apple, @IITkgp. Opinions expressed are my own

[Saurabh Dash @TheyCallMeMr_]

@ChenHenryWu We did something similar too :)

https://t.co/IBWqkhATho

[Cohere Labs @Cohere_Labs]

Announcing new research on Self-Verified RL using Soft-Rewards 📋, led by @TheyCallMeMr_, Pierre Clavier, @johnamqdang, @mgalle, @mziizm, @ahmetustun89, and @beyzaermis.

📜Read the paper: https://t.co/niWrDFmhOu

[Marzieh Fadaee @mziizm]

Really proud to see this paper coming together. Congrats to the authors, specially @TheyCallMeMr_ and @beyzaermis leading this great work.

[Marzieh Fadaee @mziizm]

RL for language models works best when answers can be verified exactly, like math or code. We argue that most real-world tasks are only partially verifiable and shows that turning prompts into checklists gives models a much richer learning signal than a single pass/fail judgment

[Hieu Pham @hyhieu226]

https://t.co/y0tc2tyjgE 😂

[Saurabh Dash @TheyCallMeMr_]

@bilaltwovec Time for my bio to shine

[Saurabh Dash @TheyCallMeMr_]

7/ Read the full paper here: https://t.co/IBWqkhATho

Work with Pierre Clavier, @johnamqdang, @mgalle, @mziizm, @ahmetustun89, and @beyzaermis.

[Saurabh Dash @TheyCallMeMr_]

1/ RLVR has driven big gains in math and code because many outputs admit reliable automatic checks: an answer matches the expected result, or a program passes tests.

But many real tasks are not like that. Code can be functionally correct but qualitatively terrible or a response may satisfy 4 syntactic constraints but fail 1 semantic constraint.

[Saurabh Dash @TheyCallMeMr_]

6/ We also study self-verification: using the same model as generator and verifier.

Naive self-verification collapses: measured reward rises, but IFEval drops from 73.9 → 55.1 as the verifier learns to always say “yes.”

Soft-SVeRL stabilizes this with verifier co-training, aggregating parallel verifier calls, and an anti-collapse penalty.

[Dhruv Kuchhal @kuchhal_dhruv]

Excited for this initiative! https://t.co/0hJ8BR1Zg6

[Saurabh Dash @TheyCallMeMr_]

@willccbb > my instinct is you basically need to model the world
I am increasingly subscribing to this camp with the caveat of the need to co-train the generators and verifiers / world-models.

[Nick Frosst @nickfrosst]

@leftylabourtech Man it’s my company?

[Adib @adibvafa]

how is this app free

[Saurabh Dash @TheyCallMeMr_]

Wordcel in the streets
Shape-rotator in the sheets

[Goodfire @GoodfireAI]

Neural networks do math by rotating shapes.

We found a shape-rotating calculator hidden inside an LLM – and it’s used for more than just math! (1/6)

[Cohere @cohere]

Excited to share our work on production-ready W4A8 inference, now integrated in vLLM! By combining 4-bit weights (low memory) with 8-bit activations (high compute), we hit the sweet spot for both decoding and prefill — up to 58% faster TTFT and 45% faster TPOT vs W4A16 on Hopper.

[Saurabh Dash @TheyCallMeMr_]

@anmol01gulati All the best, Anmol!