Soundarya Krishnan

MSML @ CMU
Previously Physics+CS @bitspilanigoa; Research intern @MSFTResearch; @MIT @medialab, @Uber, @Mitacs

Interested in #explainableAI #Causality #ML4H

[Soundarya Krishnan @SoundaryaKris16]

Last week I presented our work (done by me, @rishabkhincha, and others) at the @MICCAI2020 conference, MIL3ID workshop!
In our work, we examined the efficacy of domain-specific transfer learning for lesion classification, especially in the regime of less training data.

[Soundarya Krishnan @SoundaryaKris16]

Excited to see this work out! Huge thanks to @Le_Zhang7 for leading the project and to everyone involved. Check it out at CVPR!

[Le Zhang @ CVPR2026 @Le_Zhang7]

Humans don't just see in space—we build cognitive maps of what everything is FOR.

Can MLLMs do the same? We built SFI-Bench to find out. (CVPR 2026)

Most spatial benchmarks stop at perception: "How many chairs?" or "Is the table left of the couch?"

SFI-Bench asks harder questions:
→ "Find the maximum number of same-brand bottles on the cabinet" (conditional counting)
→ "Which objects share a functional relationship across these views?" (affordance reasoning)
→ "Diagnose this device failure using the scene + a web search" (causal troubleshooting)

We tested Gemini-3, GPT-5, Qwen3-VL-235B, LLaVA-Video-72B...

Results reveal a critical gap:
📷 Models are decent at local perception
📷 They fail at maintaining global spatial memory
📷 They struggle to link objects to their real-world function
📷 Web search helps — but only if the model can reason strongly enough to use it

SFI-Bench: 1,555 questions · 134 real-world scans · 6 cognitive tasks

project: https://t.co/dkss3hqn9N
Poster at Fri, Jun 5, 2026 • PM – PM PDT
ExHall A & F 453

#CVPR2026 #Agent #MultimodalLLM #SpatialIntelligence

[Le Zhang @ CVPR2026 @Le_Zhang7]

Humans don't just see in space—we build cognitive maps of what everything is FOR.

Can MLLMs do the same? We built SFI-Bench to find out. (CVPR 2026)

Most spatial benchmarks stop at perception: "How many chairs?" or "Is the table left of the couch?"

SFI-Bench asks harder questions:
→ "Find the maximum number of same-brand bottles on the cabinet" (conditional counting)
→ "Which objects share a functional relationship across these views?" (affordance reasoning)
→ "Diagnose this device failure using the scene + a web search" (causal troubleshooting)

We tested Gemini-3, GPT-5, Qwen3-VL-235B, LLaVA-Video-72B...

Results reveal a critical gap:
📷 Models are decent at local perception
📷 They fail at maintaining global spatial memory
📷 They struggle to link objects to their real-world function
📷 Web search helps — but only if the model can reason strongly enough to use it

SFI-Bench: 1,555 questions · 134 real-world scans · 6 cognitive tasks

project: https://t.co/dkss3hqn9N
Poster at Fri, Jun 5, 2026 • PM – PM PDT
ExHall A & F 453

#CVPR2026 #Agent #MultimodalLLM #SpatialIntelligence

[Deep Tejas Karkhanis @deepTKarkhanis]

Excited to unveil the LLM - Giraffe! 🦒

Worked with @siddartha_naidu & Arka Pal on this.

Extended context to 4K and 16K, tackling SOTA open-source LLM limitations.

Llama-1 based, Llama-2 soon. Eager for its impact on real-world AI! #AI #LLM 🚀

Repo: https://t.co/bImOOZJ0Wi

[Andrej Karpathy @karpathy]

Dear Apple I am not able to keep track of and get back to conversations across 10 apps. Needs some OS-level help to sort notifications into fyis and todos that you can sort through, mark as “unread” and deal with when you’re able. Sad as the concept is.

[Divy Thakkar @divy93t]

We are excited to open applications for female and genders underrepresented in CS undergrads in India and Southeast Asia to spend a summer with a faculty member to work on exciting problems with comp stipends, supported by Google Research.

Apply now!

https://t.co/wdnNwmY62T

[Aayushi Sanghi (she/her) @aayushisanghi]

currently at THAT point in the semester where even being incredibly organized about deadlines + trying to have a daily to-do list which tracks unfinished work STILL culminates in UTTER CHAOS

[SAiDL @SforAiDL]

New dates are out! We will be having the AI Symposium on the weekend of 2nd and 3rd October. Registrations are free! For more details:

Website: https://t.co/GYpguaJpFy
Registration: https://t.co/0YgfHd2dcK
Mailing List: https://t.co/xpLyTzrDDv

See you there!

[ACMI Lab (CMU) @acmi_lab]

"Does Pretraining for Summarization Require Knowledge Transfer?"

Short answer: most of the gains of T5 over a random init baseline can be realized absent any knowledge (by pretraining on procedurally generated babble)!

Long answer: see thread...

https://t.co/TMBtXIgkLr

[Rahul Chauhan @__rahulchauhan]

Scenes at the BITS Goa gate at 10:30PM

[michaela okland @MichaelaOkla]

Accidentally didn’t respond to any of the people I dearly love for 3 days

[Mr. Nomer @Trivial__]

my advisor: makes sense, right?

me, lying: yea of course

[Rishab Khincha @rishabkhincha]

3. those that think EOD means anywhere on earth so next day 5.30pm IST

[Years Progress @YearsProgress]

▓▓▓▓▓▓▓▓░░░░░░░ 50%

[Soundarya Krishnan @SoundaryaKris16]

@vinayaksn That’s true!

[Soundarya Krishnan @SoundaryaKris16]

anyone else feeling completely burnt out this year?

[Hritika @Hritikaa_]

Me 2 hours after taking the vaccine vs me after 12 hours

[Soundarya Krishnan @SoundaryaKris16]

PS 1 🤮🤮

[pastor micheal @dexxe]

you've heard of unpaid internship but have you heard of reverse financed internships?

[Soundarya Krishnan @SoundaryaKris16]

Writing Math on dark paper on goodnotes is the only happiness I’m getting in this lockdown

[Emma Pierson @2plus2make5]

If someone had given me an iPad in year 1 of PhD, I would've done way more methods work just to get a chance to scribble math on this thing

[𝙷𝚒𝚖𝚊 𝙻𝚊𝚔𝚔𝚊𝚛𝚊𝚓𝚞 @hima_lakkaraju]

If you have less than 3 hours to spare & want to learn (almost) everything about state-of-the-art explainable ML, this thread is for you! Below, I am sharing info about 4 of our recent tutorials on explainability presented at NeurIPS, AAAI, FAccT, and CHIL conferences. [1/n]