Dilek Hakkani-Tur

The regular submission deadline for SIGdial 2025 has now passed...

But we are still welcoming submissions through ACL Rolling Review 🎉

ARR Commitment Deadline: June 6th

Acceptance notifications will be on June 20th, and then SIGdial will be held in Avignon, France: August 25th - 27th

🚨 Paper Alert: “RL Finetunes Small Subnetworks in Large Language Models”

From DeepSeek V3 Base to DeepSeek R1 Zero, a whopping 86% of parameters were NOT updated during RL training 😮😮
And this isn’t a one-off. The pattern holds across RL algorithms and models.
🧵A Deep Dive

LLMs are all around us, but how can we foster reliable and accountable interactions with them??

To discuss these problems, we will host the first ORIGen workshop at @colmweb.org! Submissions welcome from NLP, HCI, CogSci, and anything human-centered, due June 20 :)

origen-workshop.github.io

Thrilled to announce our new survey that explores the exciting possibilities and troubling risks of computational persuasion in the era of LLMs 🤖💬
📄Arxiv: arxiv.org/pdf/2505.07775
💻 GitHub: github.com/beyzabozdag/...

🚀Our ICML 2025 paper introduces "Premise-Augmented Reasoning Chains" - a structured approach to induce explicit dependencies in reasoning chains.

By revealing the dependencies within chains, we significantly improve how LLM reasoning can be verified.

🧵[1/n]

Incredibly proud of my students @adadtur.bsky.social and Gaurav Kamath for winning a SAC award at #NAACL2025 for their work on assessing how LLMs model constituent shifts.

Agents like OpenAI Operator can solve complex computer tasks, but what happens when users use them to cause harm, e.g. spread misinformation?

To find out, we introduce SafeArena (safearena.github.io), a benchmark to assess the capabilities of web agents to complete harmful web tasks. A thread 👇

While persuasive models are promising for social good, they can also be misused towards harmful behavior. Recent work by @beyzabozdag.bsky.social and @shuhaib.bsky.social aims to assess LLM persuasiveness and susceptibility towards persuasion.

Data selection for instruction fine-tuning of LLMs doesn't need to be computationally costly. Great work by @wonderingishika.bsky.social! @convai-uiuc.bsky.social

Instruction data can also be synthesized using feedback based on reference examples. Please check our recent work for more information. Thanks to @shuhaib.bsky.social, Xiusi Chen, and Heng Ji!

AI over-reliance is an important issue for conversational agents. Our work supported mainly by the DARPA FACT program proposes introducing positive friction to encourage users to think critically when making decisions. Great team-work, all!
@convai-uiuc.bsky.social @gokhantur.bsky.social

Neurips24 Simulating Users.pptx 1 Acknowledgements Approach Results Motivation Introduction Simulating User Agents for Embodied Conversational AI Daniel Philipov, Vardhan Dongre, Gokhan Tur, Dilek Hakkani-Tür University of Illinois,...

Attending NeurIPS 2024? Learn about simulating users for embodied agents!
Catch our work, "Simulating User Agents for Embodied Conversational AI," at the Open World Agents Workshop on Dec 15.
Poster: tinyurl.com/yc42h4ud
Audio: tinyurl.com/mr2hd35a
@dilekh.bsky.social @gokhantur.bsky.social