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...

A Survey of Reinforcement Learning for Large Reasoning Models Five sections: - Foundational Components - Foundational Problems - Training Resources - Applications - Future Directions

When reading AI reasoning text (aka CoT), we (humans) form a narrative about the underlying computation process, which we take as a transparent explanation of model behavior. But what if our narratives are wrong? We measure that and find it usually is. Now on arXiv: arxiv.org/abs/2508.16599

Great interview with @stevenstrogatz.com with a lot of discussion of research advising. Parts reminded me of @eegilbert.org and @informor.bsky.social's (excellent) guides to PhD mentorship, with a big focus on ideation. Eric's: docs.google.com/document/d/1... Mor's: s.tech.cornell.edu/phd-syllabus/

We try to avoid self-promoting too much, but we (with @sjgreenwood.bsky.social) built a personalized feed with posts about papers from your network. Many people say it's the closest they can get to old academic twitter, and I hope you enjoy it and share with others too! bsky.app/profile/pape...

🤖 But wait! There's more! You can check out @shiraamitchell.bsky.social 's most recent update on the details of Calibration, posted yesterday! statmodeling.stat.columbia.edu/2025/08/12/s...

#acl2025 anyone get a good quote of phil resnik's last comment? context: (some?all?) panelists & him agree the field needs more deep, careful research on smaller models to do better science. everyone is frustrated with impossibility of large-scale pretraining experiments

@kennyjoseph.bsky.social , Kenny check this thread out

What are your favorite recent papers on using LMs for annotation (especially in a loop with human annotators), synthetic data for task-specific prediction, active learning, and similar? Looking for practical methods for settings where human annotations are costly. A few examples in thread ↴

UB's new Department of AI and Society is hiring faculty across ranks (Assistant, Associate, Full Professor). We’re looking for transdisciplinary scholars interested in building AI by society, for society. Start dates begin Fall 2025. More info: www.ubjobs.buffalo.edu/postings/57734

Took me a second, but I knew I'd seen something related to this recently: arxiv.org/abs/2505.120...

Check out our take on Chain-of-Thought. I really like this paper as a survey on the current literature on what CoT is, but more importantly on what it's not. It also serves as a cautionary tale to the (apparently quite common) misuse of CoT as an interpretable method.

Hi everyone. I'm excited to announce that I will be organizing a 2nd Summit on Responsible Computing, AI, and Society rcais.github.io October 27-29, 2025. We will explore the future of computing for health, sustainability, human-centered AI, and policy. Please consider submitting a 1-page abstract

So ICWSM concluded today and it was a blast, was a great honor to attend @icwsm.bsky.social at Copenhagen and present my work with @kennyjoseph.bsky.social and other colleagues. The paper link is here : ojs.aaai.org/index.php/IC...,

This is great! The idea is somewhat obvious (good!), and I'm sure many have toyed with the connection to learning-to-rank. However, no work had developed it. This should be relevant for constructing valid PIs from just preferential feedback. openreview.net/pdf?id=ENJd3...

🚨 New preprint! 🚨 Phase transitions! We love to see them during LM training. Syntactic attention structure, induction heads, grokking; they seem to suggest the model has learned a discrete, interpretable concept. Unfortunately, they’re pretty rare—or are they?

Generative language systems are everywhere, and many of them stereotype, demean, or erase particular social groups.

Alright, people, let's be honest: GenAI systems are everywhere, and figuring out whether they're any good is a total mess. Should we use them? Where? How? Do they need a total overhaul? (1/6)

🧵 1/ Las redes están llenas de odio. ¿Puede la inteligencia artificial ayudarnos a detectarlo… ⚖️ sin discriminar, 🚫 sin reforzar estereotipos, 🔁 y sin aprender a odiar? Esa es la gran pregunta de mi tesis. 👇 Te lo cuento en este #HiloTesis @crueuniversidades.bsky.social @filarramendi.bsky.social

As we go through a lot of excitement about RL recently with lots of cool work/results, here is a reminder that RL with a reverse KL-regularizer to the base model cannot learn any new skills that were not already present in the base model. It can only amplify the weak skills. 👇

My path into AI The sort of small wins that accumulate into a real career in AI. When I started grad school AI prof's didn't have space for me in their group and when I ended I had no papers at NeurIPS/ICLR/ICML, yet the process can still work. www.interconnects.ai/p/my-path-in...

I posted this on LinkedIn too and it has over 600 reactions there, with the caveat that I don't know how many are from bots.

Out of all the mountains of work on "debiasing" word2vec, BERT, and LLMs, what anecdotes or evidence do we have that debiasing techniques have *actually been used in practice*, in industry or research? (not referring to "cultural alignment" techniques) (bonus if used and *to good effect*)

Training with more data = better LLMs, right? 🚨 False! Scaling language models by adding more pre-training data can decrease your performance after post-training! Introducing "catastrophic overtraining." 🥁🧵👇 arxiv.org/abs/2503.19206 1/10

US tech firms - IME - rely heavily on grads of higher ed in general, and US higher ed in particular... And yet I've not heard much noise from the Pichais, Nadellas, Zuckerbergs, Cooks, Benioffs, Sus, Jassys, Huangs, etc. about attacks on higher education... Am I missing it or does it not exist?

Not a bad turnout for a Friday afternoon

RL is so back! (well, for some of us, it never really left) awards.acm.org/about/2024-t...

🚨🚨 New preprint 🚨🚨 Ever wonder whether verbalized CoTs correspond to the internal reasoning process of the model? We propose a novel parametric faithfulness approach, which erases information contained in CoT steps from the model parameters to assess CoT faithfulness. arxiv.org/abs/2502.14829

Rare that a paper these days uses the original literature of "Outcome reward model" and not just doing bradley-terry model on right/wrong labels. Nature is healing. Exploring the Limit of Outcome Reward for Learning Mathematical Reasoning Lyu et al arxiv.org/abs/2502.06781

Artem Kirsanov, Chi-Ning Chou, Kyunghyun Cho, SueYeon Chung The Geometry of Prompting: Unveiling Distinct Mechanisms of Task Adaptation in Language Models https://arxiv.org/abs/2502.08009

How should we train models for optimal test-time compute scaling? Unsurprisingly, the standard RLHF framework is sub-optimal!

𝐛𝐞𝐬𝐭-𝐨𝐟-𝐧 is a strong baseline for - improving agents - scaling inference-time compute - preference alignment - jailbreaking models How does 𝐁𝐨𝐧 work? and why is it so strong? Find some answers in the paper we wrote over two Christmas breaks!🧵

WebSci'25 is a cool venue with great people (speaking from experience :) ), please heed this call

🚨 Big news! We're thrilled to announce that #ICWSM 2026 will take place in sunny Los Angeles, California! 🌴☀️ More details coming soon! ✨

My commentary on the do's and don'ts of cognitive evaluations in LLMs is now out in Nature Human Behavior: doi.org/10.1038/s415... posting here with a figure that didn't make it into the final draft and is now instead a boring table :P #CogSci #LLMs #AI

Has anyone that has successfully gotten access to the Youtube Researcher API willing to share some tips? Stuck in "we'll get back to you" limbo and hoping there's some way to at least get a decision one way or the other at some point.

#neuhier #introduction #newhere #newkidontheblock Hi everyone at Bluesky, we just started here and would like to introduce us briefly: As the largest European infrastructure institute for the social sciences GESIS offers advice, expertise and services at all stages of scientists' research projects.

I've posted the practice run of my LSA keynote. My core claim is that LLMs can be useful tools for doing close linguistic analysis. I illustrate with a detailed case study, drawing on corpus evidence, targeted syntactic evaluations, and causal intervention-based analyses: youtu.be/DBorepHuKDM

We are looking for volunteers for reviewing and AC roles! Please sign up here: forms.gle/rZp67YvMn1hn...

Excited to share 𝐈𝐧𝐟𝐀𝐥𝐢𝐠𝐧! Alignment optimization objective implicitly assumes 𝘴𝘢𝘮𝘱𝘭𝘪𝘯𝘨 from the resulting aligned model. But we are increasingly using different and sometimes sophisticated inference-time compute algorithms. How to resolve this discrepancy?🧵

Transformer LMs get pretty far by acting like ngram models, so why do they learn syntax? A new paper by sunnytqin.bsky.social, me, and @dmelis.bsky.social illuminates grammar learning in a whirlwind tour of generalization, grokking, training dynamics, memorization, and random variation. #mlsky #nlp

Reporting in from the workshops at #NeurIPS2024! Been Kim gave a wonderful keynote at the Workshop on Socially Responsible Language Modeling Research. Her talk reflected on experiences where interpretability research delivers surprising and seemingly illogical results and how to think about this.

Meet OLMo 2, the best fully open language model to date, including a family of 7B and 13B models trained up to 5T tokens. OLMo 2 outperforms other fully open models and competes with open-weight models like Llama 3.1 8B — As always, we released our data, code, recipes and more 🎁

Read this very nice work by @andersgiovanni.com on best practices on using LLM tools for Computational Social Science arxiv.org/pdf/2408.01346

examples! CoMPosT: Characterizing and Evaluating Caricature in LLM Simulations arxiv.org/abs/2310.11501 This Prompt is Measuring <MASK>: Evaluating Bias Evaluation in Language Models arxiv.org/abs/2305.12757