Sandesh Ghimire

Staff Researcher/Engineer at Qualcomm working on ML, CV, Gen AI (edge). Prev. Researcher @Northeastern Uni, Research Intern @IBM. PhD from @RITtigers. LoveMaths

[Sandesh Ghimire @drsandeshgh]

1/n In our NeurIPS spotlight paper, 'Reliable Estimation of KL Divergence using a Discriminator in Reproducing Kernel Hilbert Space', we try to analyze estimation of KL divergence from samples: https://t.co/wAcLHIknxe

[Sandesh Ghimire @drsandeshgh]

@karlmehta Windows for LLM, right?

[Sandesh Ghimire @drsandeshgh]

Few-shot flow models consider immediate match between shortcut and teacher model; akin to immediate reward in RL. We propose to use cumulative reward from future. Our work, with
@paribeshregmi
will be presented this Saturday at ICLR 2026. Link: https://t.co/2vYIWzk9Va

[Sandesh Ghimire @drsandeshgh]

@nmboffi Great work Nicholas. Feels like the right direction for few step Diffusion LLM. One question I have is did you also try self distillation and it was worse than distilling from FLM? I can’t help but think it would be so much wonderful if self distillation works directly.

[Sandesh Ghimire @drsandeshgh]

@VictorTaelin The more I read papers along this line, the more I see connection with Diffusion models, and feedback in control system. May be diffusion type recursion and reasoning will eventually meet and voila!

[Sandesh Ghimire @drsandeshgh]

A must read indeed!

[Francis Bach @BachFrancis]

What if AI isn’t about building solo geniuses, but designing social systems?
Michael Jordan advocates blending ML, economics, and uncertainty management to prioritize social welfare over mere prediction.
A must-read rethink.
https://t.co/HUJh97pq5N

[Sandesh Ghimire @drsandeshgh]

True words!

[François Chollet @fchollet]

Key to research success: ambition in vision, but pragmatism in execution. You must be guided by a long-term, ambitious goal that addresses a fundamental problem, rather than chasing incremental gains on established benchmarks. Yet, your progress should be grounded by tractable short-term metrics/tasks that force you to constantly make contact with reality.

[Sandesh Ghimire @drsandeshgh]

Anyone who has read the original paper would appreciate this one. The original paper didn’t think too deeply what they wanted to say and what the take home message was.

[ℏεsam @Hesamation]

bro sh*t just got so real. Claude Opus published a response paper to Apple’s paper, criticizing their experiment design, putting models under token limit constraints, and having them solve unsolvable problems.

[Sandesh Ghimire @drsandeshgh]

@iScienceLuvr I wonder why this is called a Diffusion model!

[Sandesh Ghimire @drsandeshgh]

Really liked @ArnaudDoucet1 ‘s talk unifying diffusion, flow matching and schrodinger bridge into a common framework and the equivalence with OT and regularized OT.

[Arnaud Doucet @ArnaudDoucet1]

The slides of my NeurIPS lecture "From Diffusion Models to Schrödinger Bridges - Generative Modeling meets Optimal Transport" can be found here: https://t.co/RrgPIrmkdx

[Sandesh Ghimire @drsandeshgh]

The first surprise in @NeurIPSConf was the long spiraling line where we had to wait for 30min just to get our badge. The impact of scale in real life #NeurIPS2024

[Sandesh Ghimire @drsandeshgh]

@gabrielpeyre The crux of how diffusion models work lies in these two equations.

[Sandesh Ghimire @drsandeshgh]

Interesting!

[Aran Komatsuzaki @arankomatsuzaki]

Generative Reward Models

Presents GenRM, an iterative algorithm that trains an LLM on self-generated reasoning traces, leading to synthetic preference labels matching human preference judgments

https://t.co/f7oS828PwS

[Sandesh Ghimire @drsandeshgh]

Neat idea! More works are likely to follow leveraging best of both diffusion and 3D Gaussian

[Matthias Niessner @MattNiessner]

📢𝐋𝟑𝐃𝐆: 𝐋𝐚𝐭𝐞𝐧𝐭 𝟑𝐃 𝐆𝐚𝐮𝐬𝐬𝐢𝐚𝐧 𝐃𝐢𝐟𝐟𝐮𝐬𝐢𝐨𝐧📢 #SIGGRAPHAsia

We propose a generative diffusion model for 3D Gaussians. Key is a learnt latent space which substantially reduces the complexity of the diffusion process, thus facilitating room-scale scene generation!

https://t.co/5XwWbZP2m4
https://t.co/ClMN95c5PR

Great work by @barbara_roessle in with @normanisation, @angelaqdai, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder!

[Sandesh Ghimire @drsandeshgh]

@ofirnachum There are always many ways to think about everything. A way of thinking cannot be absolutely right or wrong, simpler or complex. It’s just a question of what tool is better fit for which job. You should think in that way which helps you do what you want to do

[Sandesh Ghimire @drsandeshgh]

Linearization of attention block lies at the heart of several efforts optimizing Transformers efficiency. One early paper in this topic was Transformers are RNN. In this blogpost, I try to explain this idea with elementary concepts. Enjoy reading :)

https://t.co/iQvE4NWNya

[Sandesh Ghimire @drsandeshgh]

@karpathy @maharshii That’s true, but also things depend on new ideas furling the field. After a while people tend to reach a steady state technicalizing only as much as needed. However, if newer ideas still keep coming from mathy topics like Optimal TP,the thought process also needs to be explained.

[Sandesh Ghimire @drsandeshgh]

@dwarkesh_sp @fchollet Makes me wonder if we could design an adversarial strategy to find a problem where LLM would flatten out even if we increase data.

[Sandesh Ghimire @drsandeshgh]

@karpathy Any pointers on the creative ways to combine pretraining and finetuning data/training?

[Davin Hill @davinhill_]

Interested in robustness and uncertainty in XAI? Come check out our posters at #AISTATS2024 this week!

Thursday: https://t.co/YibVd2kc2T

Saturday: https://t.co/ewJuGaA5Yg

@zq90 @drsandeshgh @MaxTee1996 @JoshBone18 @mllabneu

[Sandesh Ghimire @drsandeshgh]

Sounds super interesting to me. Bringing diffusion ideas into LLM can give a boost to language generation from algorithmic side.

[Aaron Lou @aaron_lou]

Announcing Score Entropy Discrete Diffusion (SEDD) w/ @chenlin_meng @StefanoErmon.

SEDD challenges the autoregressive language paradigm, beating GPT-2 on perplexity and quality!

Arxiv: https://t.co/EA9FpO1ieo
Code: https://t.co/0J3kKtHTgO
Blog: https://t.co/pCsDYBy1Dw

🧵1/n