Author | Lightnews

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · 24d

My students will be presenting IRPO arxiv.org/abs/2504.15477 and a paper on Personalized RLHF arxiv.org/abs/2504.07070 on Wed onwards

In-context Ranking Preference Optimization

Recent developments in Direct Preference Optimization (DPO) allow large language models (LLMs) to function as implicit ranking models by maximizing the margin between preferred and non-preferred respo...

arxiv.org

1

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · 24d

My students will be presenting papers next Wed/Thursday so be sure to check those out too

1 2

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · 24d

I'll be at #CoLM2025 and the IVADO agents workshop right before in Montreal. My students will be presenting two papers in the main conf. I'll also do a ws keynote where I'll talk about some of our latest. Come by and say hi next week!

1 7

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · Jun 29

I'm probably mostly going to stop posting on this site. There's close to no engagement and it's not worth the effort to cross post for the amount of time that takes. Find me elsewhere / email me

4 5

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · Jun 19

I recently left Mosaic/Databricks Research. It's been a ride building out the RL team from <4 ppl to 20+ across two companies & acquisition +figuring out RL as a Service in prod. Mosaic had insane talent density

Some "relaxation" while I put out Prof fires for a smol bit then new adventures!

6

Reposted by Prithviraj "Raj" Ammanabrolu

Mark Riedl @markriedl.bsky.social · Jun 17

If you work in the intersection of NLP and games/narrative, then this workshop is for you! wordplay-workshop.github.io/cfp/

Organized by the amazing @laramartin.net and @rajammanabrolu.bsky.social (among others)

/call_for_papers

Official website for the Wordplay Workshop at EMNLP 2025. Exploring interactive narratives, text-adventure games, and AI agents in language-based environments. Join us in Suzhou, China, November 5th-9...

wordplay-workshop.github.io

5 10

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · Jun 16

The thing that feels so off about the core tech world is that every convo is very transactional. Maybe true elsewhere too. "Oh you're an expert in RL, can you answer questions about my new startup?"

Every single (Bay) party. No I do not want to consult. I just wanna hang out.

1 8

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · Jun 10

Of all the labeling startups out there to acquihire, this was... an interesting choice. Says a lot actually

1

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · Jun 9

. @bosungkim.bsky.social will be at #CVPR2025 in Nashville this week to present this and just generally talk about scaling memory for embodied agents!

Catch her at the poster sessions and also the Foundation Models meets Embodied Agents Workshop on Wed

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · May 23

"Foundation" models for embodied agents are all the rage but how to actually do complex looong context reasoning? Can we scale Beyond Needle(s) in the (Embodied) Haystack?

∞-THOR is an infinite len sim framework + guide on (new) architectures/training methods for VLA models

2

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · Jun 9

Yes AI for edu is a thing but almost all vanilla LLMs just railroad students into answers. Complete cognitive offload is not useful for improving learning outcomes

2

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · Jun 9

I've heard this personally from multiple PMs at AI companies. Students are one of the biggest demographics and they need to "break in" and have even more usage to improve their metrics. Classic corporate economic incentives

Mark Riedl @markriedl.bsky.social · Jun 9

AI companies in the US gave access to their systems to students for free during college exams

China disabled access to AI systems during nationwide college exams www.theverge.com/news/682737/...

Feel free to draw your own conclusions

China shuts down AI tools during nationwide college exams

New age problems require new age solutions.

www.theverge.com

1 4 7

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · Jun 6

Tis the era of bringing back every AI benchmark ever but this time by the LLM people and for the LLMs

1 2

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · Jun 5

Had a fun little visit to Cambridge LTL where I talked about a bunch of my lab's latest papers including some still not public with the key takeaway that "RL can absolutely learn new things and is not just resurfacing knowledge"
talks.cam.ac.uk/show/archive...

talks.cam : Language Technology Lab Seminars

talks.cam.ac.uk

1 4

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · Jun 3

That's fair, I guess I should rephrase to "regardless of a possible common prior, it's nearly impossible for different providers to have the same representations pop out of their post trained LLM"

1

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · Jun 3

The moral of the story here is basically that who is making your LLM really matters. Internal use cases critical to their businesses will always influence data distributions and everything downstream of that. This is in contrast to things like Platonic Representation Hypothesis

1 5

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · Jun 3

Interesting tidbit from UCSD's Victor Shih on a podcast talking about Chinese AGI efforts is that Deepseek is good at Chinese govt doc understanding cause that's what affects stock prices most and DS is a hedge fund.
www.youtube.com/watch?v=b1Te...

Xi Jinping’s paranoid approach to AGI, debt crisis, & Politburo politics — Victor Shih

YouTube video by Dwarkesh Patel

youtu.be

1 5

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · Jun 2

The top two scores are 332 not 322 but other than the typo the rest of this list seems legit and consistent across multiple sources
www.indiatvnews.com/education/hi...

x.com/RejaullahmdM...

JEE Advanced 2025 topper list out, Rajit Gupta secures AIR 1 with 332 marks: Check complete list of toppers

IIT Kanpur has released the topper list for JEE Advanced 2025. This year, Rajit Gupta achieved AIR 1 by scoring 332 out of 360 marks. Devdutta Majhi is the leading female candidate with AIR 16. Scroll...

www.indiatvnews.com

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · Jun 2

Looks like Gemini gets AIR 6 in #JEE2025 with a score of 323

Only 5 highschoolers in all India do better than an LLM in the single most important exam of their to get into the IITs

The legacy edu selection systems are now worse than useless

1 4

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · Jun 1

I get prepping for worst case scenarios but a lot of AI Safety debates I somehow end up these days in boil down to "assume you have Machine God in a box, now tell me how to align it"

I could rant for hours but seriously y'all this isn't productive

1 8

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · May 24

Here for the afternoon shift!

1

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · May 23

Paper: arxiv.org/abs/2505.16928
Website/code/data:
pearls-lab.github.io/infini-thor/

Led by @bosungkim.bsky.social who has done a fantastic job on this in the last bit. Full stack from Unity gamedev to Big Model Scaler. Watch out for her in the embodied agent space!

Beyond Needle(s) in the Embodied Haystack: Environment, Architecture, and Training Considerations for Long Context Reasoning

We introduce $\infty$-THOR, a new framework for long-horizon embodied tasks that advances long-context understanding in embodied AI. $\infty$-THOR provides: (1) a generation framework for synthesizing...

arxiv.org

2

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · May 23

But even then agents only perform well up to 130k after which perf sharply decreases and all architectures and additional context extension methods we modified fail after ~400k. None make it to the 3m context sample test set we use let alone infinite. Lots of space for progress!

1 1

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · May 23

And find that when accounting for hardware constraints, only a specific combo of interleaved VLA with a mix of context parallelism + some extension with high pre training context window size works well. We detail the exact architecture and describe potential improvements

1

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · May 23

The first is a static eval called Needle(s) in the Embodied Haystack, which is like QA asking agents to post hoc analyze their trajectories putting many needles together

We then go Beyond this with interactive RL style evals to see how well models interact with a changing env

1 1

Prithviraj "Raj" Ammanabrolu @rajammanabrolu.bsky.social · May 23

So we first extended AI2's THOR embodied sim to continue generating meaningful tasks that are effectively infinite in length. Eg things you do at context len 7k can get reused at 3m

We do a thorough analysis of many types of architectures x training methods on two new evals

1 1