Vidya Srinivas

PhD Student @uwcse, prev. @UMich | Human-Centered Systems, Speech and Voice, Conversational AI

[Vidya Srinivas @vysrini]

What if your voice agent could use real-time knowledge from a frontier model—database lookups, tool calls, and all—and weave it into conversation without compromising on millisecond-level responsiveness?

Check out our paper!
🔗 https://t.co/wHTLgF8axb
💻 https://t.co/Ozvte3qz6m

[Zachary Englhardt @zenglhardt1]

What if your real-time voice agent could reason like a frontier model?

Conversational Infill lets a tiny on-device model talk while a frontier model thinks in the background: fast and responsive, grounded in frontier reasoning.

w/ @vysrini
📄 https://t.co/anfJCk6vrB
🧵👇

[Vidya Srinivas @vysrini]

To support complex, evolving conversations and original insights, LLMs must be designed for engagement, not just answers.

🔗Paper: https://t.co/lIlvq9n2Ed

(5/5)
Thanks to co-authors @timalthoff, @danmcduff, @xliucs, @Orson_Xu, @kmr_ayush, Issac Galatzer-Levy, Shwetak Patel

[Vidya Srinivas @vysrini]

Excited to share our paper, Substance over Style: Evaluating Proactive Conversational Coaching Agents, accepted to ACL 2025 Main Track!

LLMs do well when given clear context up front.

But what if that context doesn't exist yet?
(1/5)

🔗Read more: https://t.co/lIlvq9n2Ed

[Vidya Srinivas @vysrini]

In this paper, we explore five coaching agent styles across real user conversations.

We find that users value core functionality over stylistic elements—and that third-person evaluations often miss what truly drives progress.

(4/5)

🔗 Read more: https://t.co/lIlvq9n2Ed

[Vidya Srinivas @vysrini]

Thanks to my co-authors
@timalthoff
@danmcduff
@Orson_Xu
@xliucs
@kmr_ayush
, Shwetak Patel, and Isaac Galatzer-Levy! Had a great time completing this work during my time
@GoogleAI

[Vidya Srinivas @vysrini]

🚨What do users really want from #ConversationalAI? 🚨

Imagine a coach who truly gets you—understands and adapts to your goals, challenges, and preferences. Now imagine that coach is always there to check in on you. What would that take?

🔗https://t.co/lIlvq9n2Ed

[Vidya Srinivas @vysrini]

Furthermore, LM and expert ratings didn’t match real user feedback, revealing the complexity of evaluating #HumanCenteredAI.
Building agents that truly understand and adapt is challenging, much remains to be done!
🔗 Read our paper here: https://t.co/lIlvq9n2Ed (4/4)

[Vidya Srinivas @vysrini]

Users expected conversations that pushed them to think, explore, and set meaningful goals. Style and tone mattered, but without substance, they were frustrating. Simple prompting wasn’t enough; elaborate thought chains were too slow. (3/4)