iiinova — where your digital self keeps evolving

Valentina Pyatkin @valentinapy.bsky.social · 27/10/2025

I will be giving a talk at @eth-ai-center.bsky.social next week, on RLVR for verifiable instruction following, generalization, and reasoning! 📢 Join if you are in Zurich and interested in hearing about IFBench and our latest Olmo and Tülu works at @ai2.bsky.social

060

Valentina Pyatkin @valentinapy.bsky.social · 10/10/2025

Next up we had @tsvetshop ‘s Yulia Tsvetkov talk about ethics, safety, and reliability of LLMs in the health domain.

020

Valentina Pyatkin @valentinapy.bsky.social · 10/10/2025

💡We kicked off the SoLaR workshop at #COLM2025 with a great opinion talk by @michelleding.bsky.social & Jo Gasior Kavishe (joint work with @victorojewale.bsky.social and @geomblog.bsky.social ) on "Testing LLMs in a sandbox isn't responsible. Focusing on community use and needs is."

1154

Valentina Pyatkin @valentinapy.bsky.social · 10/08/2025

On my way to Oxford: Looking forward to speaking at OxML 2025

070

Valentina Pyatkin @valentinapy.bsky.social · 03/07/2025

Additionally, we wrote new training constraints and verifier functions and suggest a good recipe for IF-RLVR training for improved generalization. We find that IF-RLVR generalization works best on base models and when you train on multiple constraints per instruction!

110

Valentina Pyatkin @valentinapy.bsky.social · 03/07/2025

💡Beyond math/code, instruction following with verifiable constraints is suitable to be learned with RLVR. But the set of constraints and verifier functions is limited and most models overfit on IFEval. We introduce IFBench to measure model generalization to unseen constraints.

1295

Valentina Pyatkin @valentinapy.bsky.social · 17/06/2025

Interested in shaping the progress of responsible AI and meeting leading researchers in the field? SoLaR@COLM 2025 is looking for paper submissions and reviewers! 🤖 ML track: algorithms, math, computation 📚 Socio-technical track: policy, ethics, human participant research

181

Valentina Pyatkin @valentinapy.bsky.social · 12/05/2025

📢 The SoLaR workshop will be collocated with COLM! @colmweb.org SoLaR is a collaborative forum for researchers working on responsible development, deployment and use of language models. We welcome both technical and sociotechnical submissions, deadline July 5th!

1166

Valentina Pyatkin @valentinapy.bsky.social · 11/12/2024

this year it’s a thermos and not a mug 🥲

3100

Valentina Pyatkin @valentinapy.bsky.social · 01/12/2024

Michael will present his work on "Diverging Preferences" at the Pluralistic Alignment workshop!

170

Valentina Pyatkin @valentinapy.bsky.social · 01/12/2024

@drjingjing.bsky.social will present her SafetyAnalyst work at the SoLaR workshop:

191

Valentina Pyatkin @valentinapy.bsky.social · 01/12/2024

@shocheen.bsky.social and co will be at the Thursday poster session to present our paper on "Contextual Noncompliance"

1101

Valentina Pyatkin @valentinapy.bsky.social · 01/12/2024

On Thursday, @hamishivi.bsky.social will present our work on "Unpacking DPO and PPO"

180

Valentina Pyatkin @valentinapy.bsky.social · 21/11/2024

Open Post-Training recipes! Some of my personal highlights: 💡 We significantly scaled up our preference data! 💡 RL with Verifiable Rewards to improve targeted skills like math and precise instruction following 💡 evaluation toolkit for post-training (including new unseen evals!)

0221

Valentina Pyatkin @valentinapy.bsky.social · 18/11/2024

📣 We are looking forward to an interesting line up of invited speakers at the SoLaR workshop at NeurIPS! solar-neurips.github.io Come join us on Saturday, Dec. 14th in Vancouver @peterhenderson.bsky.social, Been Kim, @zicokolter.bsky.social, Rida Qadri, Hannah Rose Kirk

1200

Valentina Pyatkin @valentinapy.bsky.social · 07/11/2024

To address this, we develop distributional reward models which can model annotator preferences while also identifying disagreements. We show these models can be applied to improve evaluations by identifying divisive examples in LLM-as-Judge benchmarks like WildBench. [5/6]

110

Valentina Pyatkin @valentinapy.bsky.social · 07/11/2024

We identify similar behaviors in LLM-as-Judge evaluations, and find biases in how these methods identify winning responses in cases of diverging preferences. For instance, LLM-judges prefer complying responses when annotators disagree on whether refusing is appropriate. [4/6]

110

Valentina Pyatkin @valentinapy.bsky.social · 07/11/2024

Since disagreements caused by split user perspectives, do reward models capture these diverse user views? We find reward models fail to distinguish between diverging and high-agreement preferences, decisively identifying a preferred response even when annotators disagree. [3/6]

100

Valentina Pyatkin @valentinapy.bsky.social · 07/11/2024

Why and when do preference annotators disagree? And how do reward models + LLM-as-Judge evaluators handle disagreements? Michael explored these questions in a new ✨preprint✨ from his @ai2.bsky.social internship with me!

1298

Valentina Pyatkin @valentinapy.bsky.social · 19/01/2024

"Do LLM predictors provide structurally consistent outputs in the zero- and few-shot regime?" Our new work "Promptly Predicting Structures: The Return of Inference" shows that they do not, and we show how to fix it. Paper: arxiv.org/abs/2401.06877 Code: github.com/utahnlp/prom...

020