Tim Vieira

When you say, "This is a reinforcement learning problem," you should say it with the same excitement as "This is NP-hard."

@roydanroy BTW if you get stuck in this configuration, you can push the "reset" button to get a random initialization.

@xtimv I think I broke your optimizer.

I built an interactive JavaScript thingy to study the two faces of KL divergence. timvieira.github.io/blog/interacti… I have wanted this since 2009. Thank you, Claude Code, for helping me get there!

@roydanroy Yeah, sometimes KL(q||p) optimization diverges. I thought about constraining the variance parameters to mitigate this, but then realized that it might be instructive to see it happening. Got any suggestions for improvements? Implementing suggestions is cheap these days!

It's pretty addictive to play with. You can drag the bumps around and +/- them. It currently only supports single-model approximators.

Congratulations to our researchers (including @leoduw and alum @xtimv) on winning an Outstanding Paper Award at the most recent @COLM_conf! Learn more about their award-winning paper here: cs.jhu.edu/news/hopkins-c…

Claude Chic: A Claude drop-in with updated UX and multi-agent support. Also what I've been obsessively building the last two weeks 🙂. Enjoy! matthewrocklin.com/introducing-cl…

@yisongyue @gneubig No way - what years, @gneubug? I was there 2004–2009?

Today I learned that @gneubig and I both went to UIUC for undergrad at the exact same time. But we didn't know each other. #OnlyAtCOLM

Congrats to @adveisner and @leoduw on their Outstanding Paper at COLM 2025! 🎉 Extra shoutout to @xtimv and Ryan — both proud @jhuclsp alums — for co-authoring this amazing work led by @ben_lipkin and Ben LeBrun. 👏

Good morning @COLM_conf! Excited to present our poster on Self-Steering LMs (#50, 11AM-1PM). If you’re thinking about codegen, probabilistic inference, or parallel scaling, stop by for a chat!

Having a hard time controlling your LM? Want a fast approach to get high-quality constraint-satisfying generations? Come to COLM tomorrow morning to learn about our new decoding algorithm! Oral spotlight @ 10:15-10:30am Poster #49 @ 11:00am-1:00pm

Outstanding paper 🏆 1: Fast Controlled Generation from Language Models with Adaptive Weighted Rejection Sampling openreview.net/forum?id=3BmPS…

Thanks to everyone at COLM for the awesome discussions so far and to the amazing team that made this paper happen :) Ben LeBrun @postylem @JoaoLoula @DRMacIver @leoduw @adveisner Ryan Cotterell @vmansinghka Tim O'Donnell @alexanderklew @xtimv

Excited to share that this paper has been accepted to #NeurIPS2025 main track 🎉!

Vibe coding be like

@suzyahyah Thanks! Yeah, this seems like a special case of the halo effect.

@xtimv maybe you're looking for "halo effect" or "authority bias". Its widely used in marketing, when they get actors in lab coats to present a product

Is there a name for the following effect? There exists a cognitive bias whereby the perceived reliability of an argument increases with its presentation quality, leading to a decrease in verification effort.

@arntzenius Sounds like Baeza-Yates intersection

Intersection of 2 sorted lists A, B: Wlog |A| <= |B|. Let x = A[|A|/2] and binary search for x in B. This splits A into halves and B into two parts. Recursively intersect A1,B1 and A2,B2. Is this a well-known sorted list intersection algorithm? What is its worst-case complexity?

Has anyone tried running AI models (CNNs/LLMs, ViTs/ Diffusion) on weird chips? Edge: Qualcomm AR1, Ambarella, TensTorrent Cloud: Trainium, Inferentia, AMD Or even just porting Ampere → Hopper → Blackwell? Curious: how painful was it? Did it kill your project before it started?

📣 "Understanding and Meeting Practitioner Needs When Measuring Representational Harms Caused by LLM-Based Systems" is forthcoming at #ACL2025NLP---and you can read it now on arXiv! 🔗: arxiv.org/pdf/2506.04482 🧵: ⬇️

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