You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$2,700 left to reach the funding goal.
Many AI agent benchmarks use LLMs to simulate users. But these simulated users don't always follow their instructions.
We studied this in our paper, UserProxyBench. We found that 24.4% of successful episodes had user instruction violations. We also found that changing the user simulator, while keeping the same agent, could change benchmark scores. This is a problem because we rely on these scores to understand how well agents work.
Our paper was accepted at the AABA4ET workshop at NeurIPS 2026 in Sydney.
Another paper, MIPS, was accepted at the EvoRobust workshop. In this work, we study mode collapse in reinforcement learning. We propose a method to help models find different high-quality solutions instead of focusing on just a few.
I'm hoping to attend NeurIPS to present both papers and discuss our work with other researchers. I'm looking for some support to cover my flight and stay in Sydney.
Papers:
UserProxyBench: https://arxiv.org/abs/2609.38043
There are still a few questions we want to explore with UserProxyBench.
Right now, we use models from the same family to write and grade our evaluation rubrics. A separate model reviews them. We'd like to find better ways to check our results other than simply using seperate models (ongoing work). We also want to see what happens with longer conversations and more realistic users: how do we design the setup for such a study reliably.
The AABA4ET workshop is sponsored by Snorkel AI so it would be great to learn how companies run agent evaluations at a larger scale. I'd like to discuss this with people who work on them.
For MIPS, we'd like to see where our method could be useful beyond the tasks we tested. We tested this in scientific setups like phylogenetic trees and molecule discovery, but I would be interested to find what sort of RL applications other researchers might be looking at.
The main goal is to present both papers. Other than that, It would also be great to find researchers interested in working on some of these questions together and to come back with a clearer idea of which experiments are worth trying next.
I'm requesting $2,700 for:
Round-trip flight from Boston to Sydney: $1,800
Accommodation for 4 nights (December 10–14): $900
These are estimates based on current prices. The minimum goal of $1,800 would cover the flight, and the full amount would cover my hostel stay as well.
I've already paid for my Australian visa and NeurIPS registration myself. I'll also cover food, local transportation, and other expenses during the trip.
UserProxyBench is joint work with Ashish Jain from Sarvam AI.
MIPS is joint work with Abhijeet Sinha and Dianbo Liu from the National University of Singapore. The project started during my research internship at NUS, and the full paper is currently under review at ICLR 2027.
I'm currently pursuing my master's in Computer Science at UMass Amherst. I received a $600 Rapid Grant from BlueDot Impact for some compute earlier for work on RL which was published at COLM 2026. This was for research, not travel.
The biggest challenge is the cost of getting to Sydney. If I can't raise enough for the trip, I may have to miss the in-person presentations.
The papers are already public and I will continue my research either way, but I would miss is the opportunity to discuss the results in person, and potentially start new collaborations.
I haven't received any funding for NeurIPS travel so far.
There are no bids on this project.