My talk at Samsung AI Forum yesterday www.youtube.com/watch?v=L2nA...

In the year since LRMs ("reasoning models") hit the scene, we have been trying to understand, analyze and demystify them.. Here are our efforts to date--conveniently all in one place..👇 www.linkedin.com/posts/subbar...

𝐏𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐭𝐢𝐯𝐞 𝐓𝐡𝐢𝐧𝐤𝐢𝐧𝐠? The anthropomorphization of LRM intermediate tokens as thinking begat a cottage industry to "get efficiency by shortening thinking." We ask: 𝗜𝘀 𝗖𝗼𝗧 𝗹𝗲𝗻𝗴𝘁𝗵 𝗿𝗲𝗮𝗹𝗹𝘆 𝗮 𝗿𝗲𝗳𝗹𝗲𝗰𝘁𝗶𝗼𝗻 𝗼𝗳 𝗽𝗿𝗼𝗯𝗹𝗲𝗺 𝗵𝗮𝗿𝗱𝗻𝗲𝘀𝘀 𝗼𝗿 𝗶𝘀 𝗶𝘁 𝗺𝗼𝗿𝗲 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝘃𝗲? 👉 www.linkedin.com/posts/subbar...

Rejecting papers in #AI Conferences because of "resource constraints" is shooting ourselves in the foot as a community; use Findings.. #SundayHarangue 👇 x.com/rao2z/status...

Proofs are not reasoning traces & I/O Format Language shouldn't be much of an issue for LLMs + other things #SundayHarangue (Special IMO edition). 🧵 👇 x.com/rao2z/status...

Both LLMs and LRMs are upper bounded by humanity's knowledge closure. True scientific discoveries are, by definition, outside of that closure. Ergo, LLMs/LRMs are great force multipliers to us; but don't support "Nobel this weekend" hype.. 👉 www.linkedin.com/posts/subbar...

Computational Complexity is the wrong measure for LRMs (as it was for LLMs)--think distributional distance instead #SundayHarangue (yes, we're back!) 👉 x.com/rao2z/status...

A̶̶̶I̶̶̶ ̶ ̶ ̶ ̶(̶A̶r̶t̶i̶f̶i̶c̶i̶a̶l̶ ̶I̶n̶t̶e̶l̶l̶i̶g̶e̶n̶c̶e̶)̶ ̶̶̶A̶̶̶G̶̶̶I̶̶̶ ̶(̶A̶r̶t̶i̶f̶i̶c̶i̶a̶l̶ ̶G̶e̶n̶e̶r̶a̶l̶ ̶I̶n̶t̶e̶l̶l̶i̶g̶e̶n̶c̶e̶)̶ ̶̶̶A̶̶̶S̶̶̶I̶̶̶ ̶(̶A̶r̶t̶i̶f̶i̶c̶i̶a̶l̶ ̶S̶u̶p̶e̶r̶ ̶I̶n̶t̶e̶l̶l̶i̶g̶e̶n̶c̶e̶) ASDI (Artificial Super Duper Intelligence) Don't get stuck with yesterday's hypeonyms! Dare to get to the next level! #AIAphorisms

For anyone interested, here are the videos of the three ~50min each lectures on the reasoning/planning capabilities of LLMs/LRMs that I gave at #ACDL2025 in Riva Del Sole resort last week. 1/ www.youtube.com/playlist?lis...

...it basically confirmed what is already well-established: LLMs (& LRMs & "LLM agents") have trouble w/ problems that require many steps of reasoning/planning. See, e.g., lots of recent papers by Subbarao Kambhampati's group at ASU. (2/2)

An AGI-wannabe reasoning model whining that it couldn't handle a problem because its context window isn't big enough is like a superman-wannabe little kid protesting that he couldn't add those numbers because he doesn't have enough fingers and toes.. #AIAphorisms

"our counter-intuitive results demonstrate ways in which common interpretations of Large Reasoning Models may be anthropomorphizations or simplifications" arxiv.org/abs/2505.13775

The transformer expressiveness results are often a bit of a red herring as there tends to be a huge gap between what can be expressed in transformers, and what can be learned with gradient descent. Mind the Gap, a new paper with Lucas Saldyt dives deeper into this issue 👇👇 x.com/SaldytLucas/...

Anthropomorphization of intermediate tokens as reasoning/thinking traces isn't quite a harmless fad, and may be pushing LRM research into questionable directions.. So we decided to put together a more complete argument. Paper 👉 arxiv.org/pdf/2504.09762 (Twitter thread: x.com/rao2z/status...)

This RLiNo? paper (arxiv.org/abs/2505.13697) lead by Soumya Samineni and Durgesh_kalwar dives into the MDP model used in the RL post-training methods inspired by DeepSeek R1, and asks if some of the idiosyncrasies of RL aren't just consequences of the simplistic structural assumptions made

Do Intermediate Tokens Produced by LRMs (need to) have any semantics? Our new study 👇 Thread 👉 x.com/rao2z/status...

Delighted to share that Siddhant Bhambri & Mudit Verma's critical evaluation and refutation of the reasoning claims of ReACT has been accepted to #TMLR (Transactions on Machine Learning) 👉https://openreview.net/forum?id=aFAMPSmNHR

Solving Single Agent Fully Observable Deterministic (SAFODP) Problems with Dec-POMDP approaches #SundayHarangue #allegory x.com/rao2z/status...

IMHO, the whole idea of connecting "length of intermediate tokens" produced by LRMs to inference time compute is a mind-boggling demonstration of circular reasoning--that comes from the assumptions about MDP model and reward model.. 👇 x.com/rao2z/status...

It ain't "The Bitter Lesson" if you are in the loop curating the training data for your LLM, y'all.. Pick your lesson, will ya? #SundayHarangue (h/t @kstechly.bsky.social)

Don't use summarizers for the papers by @rao2z.bsky.social because the reasoning traces therein are, unlike the LRMs & LLMs under investigation, substantively meaningful, semantically well-ordered, and stylistically compelling and engaging! #AI #LLMs #CoT arxiv.org/abs/2504.09762

Here is a recording of my talk at @msftresearch.bsky.social last week titled "(How) Do LLMs Reason/Plan?" (Also gave a version of it at as a distinguished lecture at Oracle today..) www.youtube.com/watch?v=0u2h...

A preprint available at arxiv.org/abs/2504.09762

Our invited commentary for the Annals of New York Academy of Sciences titled "(How) Do reasoning models reason?" is now online 👉 nyaspubs.onlinelibrary.wiley.com/doi/epdf/10.... It is a written version of my recent talks (and #SundayHarangues) on the recent developments in LRMs..

Woo hoo.. Our first #TMLR paper!🤗 On the planning and scheduling abilities of LRMs o1 & R1 (w/ Karthik, Kaya, Atharva) 👉 openreview.net/forum?id=FkK... Even a jaded researcher like me has to admit that Transactions on Machine Learning Research is a veritable oasis among #AI publication venues! 🙏

AI Hype: The phenomenon where experts without expertise hype up imminent arrival of expertise without experts. #AIAphorisms

Test-time-scaling, Post-training and Distillation are just compiling the verifier signal into the LLM at different phases #SundayHarangue See 👉 x.com/rao2z/status... Or 👉 www.linkedin.com/posts/subbar...

Intermediate tokens being dubbed as "Reasoning Traces" is the new anthropomorphization fashion.. See this video #SundayHarangue 👉https://youtube.com/watch?v=CQ5JS3v61Ns&list=PLNONVE5W8PCRbf3WmbcqgXPToJuA2NUfP&t=3787s that wonders whether LRMs should instead be called LMMs--Large Mumbling Models..

RL is great; but RL envy in LLMs may not be.. (or R1's SFT vs. RL is more like Batch vs. SGD)#SundayHarangue (Special Turing edition 😋) There has been a tendency in the LLM literature to dress up simplistic ideas in RL garb to gain additional respectability... 👉 x.com/rao2z/status...

Thanks to everyone who came to my #AAAI2025 award talk on Saturday. I was gratified to see so many of the people I know and respect in the audience.. 🙏🤗 Here is a recording if you are interested.. 👉 www.youtube.com/watch?v=CQ5J...

Thank you AAAI & EAAI for this honor..🙏 (my talk at #AAAI2025 will be on Saturday 2pm..)

One reason I am so looking forward to #AAAI2025 is for a chance to commiserate with my homies about the abrupt shift of the Overton Window..

No, Virginia, those Rorschach musings by LRMs may not be "reasoning traces" in any meaningful way.. #SundayHarangue I have been saying for sometime now that the intermediate tokens/"mumblings" that LRMs tell themselves are not to be seen as "reasoning traces" 👉 x.com/rao2z/status...

"Test Time scaling" is being used to refer to anything with variable length answers--whether or not there is adaptive computation. If R1 can be said to be doing test time scaling, then GPT3.5 was doing test time scaling too--given that it too output variable length essays!

On the MDP formulation of LLMs used in R1 [Not quite #SundayHarangue] You know DeepSeek R1 uses RL--but did you grok the strange MDP formulation it uses for this? We had a fun 3hr discussion about this in our group meeting last Friday; here is a summary 👉 x.com/rao2z/status...

Qn for the #AcademicBirds: The IDC that universities charge is the same for all federal grants--NIH, NSF etc. iirc. So does the NIH IDC tweet also apply across the board or we wait for individual tweets?

Spoke with @nitasha.bsky.social of @washingtonpost.com about reasoning models like DeepSeek R1 for her article, which accurately reflects my view about the Rorschach nature of the so called "reasoning traces" 👉https://www.washingtonpost.com/technology/2025/02/08/deepseek-ai-chatbot-reasoning-china/

For anyone who saw my MLST interview from December and wanted to make sense of Deepseek R1 from that prompt augmentation perspective, here is an R1 delta addendum you might enjoy: 👉 x.com/rao2z/status...

In this 2nd outing on MLStreetTalk, I chatted with Tim Scarfe about reasoning models like o1/R1; CoT & prompt augmentations; our evaluation of o1; my speculations on how o1 works; RL post training vs. Inference Time scaling; yummy stone soup; and ofc Monty Python! www.youtube.com/watch?v=2xFT...

In case you interested, performance of Deepseek's R1 model on PlanBench (thanks to Karthik Valmeekam) Explanatory thread here: x.com/rao2z/status...

BREAKING NEWS: AGI launch is being delayed on account of AGI benchmark and test data not being available in time for pre-training. #LLaMAGI #AIAphorisms

On the seedy optics of "Building an AGI Moat by Corralling Benchmark Creators" #SundayHarangue [Thoughts on OpenAI/Frontier Math benchmark story] x.com/rao2z/status...

2025 IS GOING TO BE A HUGE YEAR FOR AI AGENTS!!! Russell & Norvig recently published this great textbook on AGENTS!! 𝘏𝘦𝘳𝘦 𝘪𝘴 𝘸𝘩𝘢𝘵 𝘪𝘴 𝘪𝘯𝘤𝘭𝘶𝘥𝘦𝘥 AI Complete! 𝘞𝘩𝘦𝘳𝘦 𝘤𝘢𝘯 𝘺𝘰𝘶 𝘭𝘦𝘢𝘳𝘯 𝘢𝘣𝘰𝘶𝘵 𝘪𝘵 𝘢𝘭𝘭? Your neighborhood Intro #AI course (eg. rakaposhi.eas.asu.edu/cse471 ) See also 👉 x.com/rao2z/status...

Happy New Year from Kailash.. 😍 #Ellora

A meta list of all my 2024 #SundayHarangue's Not quite sure why, but I apparently wrote sixteen long #AI related Sunday Harangues in 2024.. 😅. Most were first posted on twitter. 👉https://x.com/rao2z/status/1873214567091966189

LLMs and "True" RL #SundayHarangue With RL back in favor, there is some heated discussion on #AI Twitter/Sky? on whether or not LLMs use any "true" RL. Unfortunately, IMHO, some of the discussion confounds some basic elements of RL. 👉 x.com/rao2z/status...

👀

Is it fair to say ARC got Winogranded?🤔 Demise of ARC reminded me of this 👇from three years back..

Inference Time Computation is NOT new--we wanted to get rid of it, but are letting it back in out of necessity #SundayHarangue OAI folks argue that including inference time computation was a relatively newer idea in games (which they have brought to LLMs). Not really.. 👉 x.com/rao2z/status...

That's a wrap for me at #NeurIPS2024.. Met more than 16 cool people in an ocean of 16,000+ And some of them even seemed to *already* know frontier concepts such as A* search (c.f. x.com/rao2z/status....