Riti Agarwal

@CompositeAI | @caltech

[Riti Agarwal @riti_agarwal084]

Interesting that more history didn’t necessarily help with better intent detection!

[Charlie Deane @KicksCharlie]

Prompting is temporary. We believe the future of computer use is completely unprompted.

Introducing our latest research: a browser agent that predicts your next action from only the current page and your recent history.

[Riti Agarwal @riti_agarwal084]

@servo_chignon insane cook!

[Riti Agarwal @riti_agarwal084]

@itsKrishM Lets go!!

[Riti Agarwal @riti_agarwal084]

Every decision can now be made quickly and based on insights! Congratulations to @itsKrishM and the @ListenLabs team! Super excited for this!

[Krish Mehta @itsKrishM]

Feels good to finally be able to talk about what you spent the last half of the year building

[Riti Agarwal @riti_agarwal084]

@itsKrishM Holy!! So cool that every decision can now be validated! No more decision fatigue!

[Riti Agarwal @riti_agarwal084]

@itsKrishM So cool!

[Riti Agarwal @riti_agarwal084]

@YangFanYun Lets go!!! @t2chuang the goat

[Riti Agarwal @riti_agarwal084]

@KicksCharlie I wonder what portion of this 22% are internal company websites? Makes it hard to measure how models generalize across these websites because i assume they aren't as prevalent in training data and evals?

[Riti Agarwal @riti_agarwal084]

Super interesting to see all the web patterns! You can enter in any domain and see how people got there and where they go next!

[Tim Chuang @t2chuang]

Across a large panel of real browsing sessions, we studied how work moves from page by page to see where people go next. Every jump tells a story about how work really gets done.

Here’s what 700 of the most-used sites told us 🧵 (1/5)

[Riti Agarwal @riti_agarwal084]

@t2chuang so interesting that @GeminiApp comes in higher than @claudeai in total active time!

[Riti Agarwal @riti_agarwal084]

@tamarajtran Yes! We ran a lot of evals internally! Blog post on this coming soon actually!

[Riti Agarwal @riti_agarwal084]

Interesting that more history didn’t necessarily help with better intent detection!

[Charlie Deane @KicksCharlie]

Prompting is temporary. We believe the future of computer use is completely unprompted.

Introducing our latest research: a browser agent that predicts your next action from only the current page and your recent history.

[Riti Agarwal @riti_agarwal084]

@tamarajtran @KicksCharlie We found the last ~15 actions usually reveal the broader goal, if they don’t, the user has likely moved on. A 128-action history adds unrelated browsing and stale intents, diluting the signal.

[Charlie Deane @KicksCharlie]

How it works:
- Fine-tuned on real browser sessions
- Uses a custom DOM representation
- Predicts sequentially, in real time
- Learned confidence through abstention training

[Charlie Deane @KicksCharlie]

The key insight: inferring intent from a short history is much harder than selecting the right element.

Find our methods and learnings below:

https://t.co/EUADtfMoCM

[Riti Agarwal @riti_agarwal084]

@rronak_ @KicksCharlie You still have to press tab to complete!

[Charlie Deane @KicksCharlie]

Prompting is temporary. We believe the future of computer use is completely unprompted.

Introducing our latest research: a browser agent that predicts your next action from only the current page and your recent history.

[Riti Agarwal @riti_agarwal084]

@itsKrishM Each SFT epoch took 13.5 hrs GPU hrs, absolute-reward RL run took 18 hrs, group relative RL run took 35 hrs (all numbers for B200s)

[Krish Mehta @itsKrishM]

Pretty cool stuff @riti_agarwal084

[Riti Agarwal @riti_agarwal084]

@itsKrishM Thank you Krish!

[Riti Agarwal @riti_agarwal084]

@Veer429791 @KicksCharlie Means a lot coming from best intern!

[Riti Agarwal @riti_agarwal084]

@arjunsriram2002 @KicksCharlie haha the credit card information is there you should try using it