Dhananjay Ashok

I’m a Ph.D. student at the University of Southern California, advised by Jonathan May and Jesse Thomason. My research focuses on self-improving agents.

[Dhananjay Ashok @Dhananjay99Ash]

Introducing GameBoyWorlds, a new challenge testbed for self-improvement in embodied video games:

Paper: https://t.co/05Fpom1OM1
Website: https://t.co/qFUZfld80b

[Atharva Kulkarni @athrvkk]

Is the geometry of language model weights really predictive of performance?🔍

Our new work challenges the popular hypothesis that low rank unembedding matrix hurts LLM performance; and the answer is more complicated than you'd think!

https://t.co/0YxcZmNpfb

(1/8)

[Dhananjay Ashok @Dhananjay99Ash]

@JimmyTYYang1 Hi Jimmy. I'm a PhD Candidate working on self-improvement and deploying world models to real robots. Could we chat about the role?

[Dhananjay Ashok @Dhananjay99Ash]

@ThomsonYenTY Covering quality criteria feels quite important. Too many works stick to final accuracy alone.

[Thomson @ThomsonYenTY]

Modern businesses run on spreadsheets. Can AI handle financial spreadsheet construction?

We’re excited to share MBABench 2.0: 101 difficult, long-horizon tasks (take up to 33h) across 6 finance areas and most financial model types.

The best AI agent passes only 16% of the tasks

[calix @calixo888]

How do you build the most diverse UMI dataset in robotics?

In the last 8 weeks, we scaled an international UMI data collection operation from 5 to 90 operators, built out a dataset of over 1M+ unique tasks, and continue to capture over 45k+ new tasks each day.

Robotic World Models require much higher diversity of tasks than imitation learning. Here's the why and how behind our operations. 🧵

[Dhananjay Ashok @Dhananjay99Ash]

Our next steps are to develop a self-improvement algorithm that works across games and create an agent that can acquire knowledge and clear games without relying on pretraining knowledge.

If you'd like to contribute, start here: https://t.co/BKlp0dIbdo

[Dhananjay Ashok @Dhananjay99Ash]

Introducing GameBoyWorlds, a new challenge testbed for self-improvement in embodied video games:

Paper: https://t.co/05Fpom1OM1
Website: https://t.co/qFUZfld80b

[Dhananjay Ashok @Dhananjay99Ash]

We see a drastic difference in progression between classic games and ROM hacks. For example, while in Red, the agent clears several key milestones (obtains a starter, reaches the next town, etc.), the very same agent fails to even receive a starter in Brown.

[Ryan Lee @_ryantlee]

My first PhD paper, 'Sparse Layers are Critical to Scaling Looped Language Models,' is accepted at NeurIPS 2026! Looping & early exits enables adaptive depth; MoE makes it scale. First step toward compute-adaptive LMs.

w/ @jbiloki @edwardjhu @jonathanmay
https://t.co/bjxI5K1siX

[calix @calixo888]

We (@PantheonInc) have been working on a robotics data quality pipeline that uncovered a series of major problems in public robotics datasets, especially for world modeling.

To improve the quality of data available to open-source robotics, we're publishing annotations for four of the most popular datasets. Some examples of issues, and our report 🧵

[Dhananjay Ashok @Dhananjay99Ash]

To see more, check out our NeurIPS 2025 Paper: https://t.co/BH4Ak564UW

[Dhananjay Ashok @Dhananjay99Ash]

🚨 New paper: Language Models Can Predict Their Own Behavior (NeurIPS 2025).
sequence-level

LMs "make up their minds" sooner than we thought. Despite being trained to predict the next token, hidden states are predictive goldmines for a range of sequence level behaviors. 🧵

[Dhananjay Ashok @Dhananjay99Ash]

Excitingly, the probe performance scales with model size. The hidden states of larger models contain clearer signals on how the model is going to behave: