@kjha02.bsky.social
CS PhD Student @University of Washington, CSxPhilosophy @Dartmouth College Interested in MARL, Social Reasoning, and Collective Decision making in people, machines, and other organisms kjha02.github.io
Can't wait to present this work @iclr-conf.bsky.social this year!!! Looking forward to hearing everyone's thoughts on the paper and learning more about peoples' research!
Thanks again to my collaborators for all of their help on this project!
Correctly tagging @aydanhuang265.bsky.social !
For more analyses and insights, check out the paper and code: shorturl.at/siUYI
Can’t thank my collaborators @aydan_huang265, @EricYe29011995, @natashajaques.bsky.social , @maxkw.bsky.social enough for all the help and support!!!
The big takeaway: framing behavior prediction as a program synthesis problem is an accurate, scalable, and efficient path to human-compatible AI!
It allows multi-agent systems to rapidly and accurately anticipate others' actions for more effective collaboration.
ROTE doesn’t sacrifice accuracy for speed!
While initial program generation takes time, the inferred code can be executed rapidly, making it orders of magnitude more efficient than other LLM-based methods for long-horizon predictions.
What explains this performance gap? ROTE handles complexity better. It excels with intricate tasks like cleaning and interacting with objects (e.g., turning items on/off) in Partnr, while baselines only showed success with simpler navigation and object manipulation.
We scaled up to the embodied robotics simulator Partnr, a complex, partially observable environment with goal-directed LLM-agents.
ROTE still significantly outperformed all LLM-based and behavior cloning baselines for high-level action prediction in this domain!
A key strength of code: zero-shot generalization.
Programs inferred from one environment transfer to new settings more effectively than all other baselines. ROTE's learned programs transfer without needing to re-incur the cost of text generation.
Can scripts model nuanced, real human behavior?
We collected human gameplay data and found ROTE not only outperformed all baselines but also achieved human-level performance when predicting the trajectories of real people!
Introducing ROTE (Representing Others’ Trajectories as Executables)!
We use LLMs to generate Python programs 💻 that model observed behavior, then uses Bayesian inference to select the most likely ones. The result: A dynamic, composable, and analyzable predictive representation!
Traditional AI is stuck! Predicting behavior is either brittle (Behavior Cloning) or too slow with endless belief space enumeration (Inverse Planning).
How can we avoid mental state dualism while building scalable, robust predictive models?
Forget modeling every belief and goal! What if we represented people as following simple scripts instead (i.e "cross the crosswalk")?
Our new paper shows AI which models others’ minds as Python code 💻 can quickly and accurately predict human behavior!
shorturl.at/siUYI%F0%9F%...
Still catching up on my notes after my first #cogsci2025, but I'm so grateful for all the conversations and new friends and connections! I presented my poster "When Empowerment Disempowers" -- if we didn't get the chance to chat or you would like to chat more, please reach out!
Our new paper is out in PNAS: "Evolving general cooperation with a Bayesian theory of mind"!
Humans are the ultimate cooperators. We coordinate on a scale and scope no other species (nor AI) can match. What makes this possible? 🧵
www.pnas.org/doi/10.1073/...
Really pumped for my Oral presentation on this work today!!! Come check out the RL session from 3:30-4:30pm in West Ballroom B
You can also swing by our poster from 4:30-7pm in West Exhibition Hall B2-B3 # W-713
See you all there!
I'll be at ICML next week! If anyone wants to chat about single/multi-agent RL, continual learning, cognitive science, or something else, shoot me a message!!!
Oral @icmlconf.bsky.social !!! Can't wait to share our work and hear the community's thoughts on it, should be a fun talk!
Can't thank my collaborators enough: @cogscikid.bsky.social y.social @liangyanchenggg @simon-du.bsky.social @maxkw.bsky.social @natashajaques.bsky.social
For the paper and code, see kjha02.github.io/publication/...
Thank you so much to my collaborators @cogscikid.bsky.social @liangyanchenggg @
@simon-du.bsky.social @maxkw.bsky.social @natashajaques.bsky.social for making the first publication of my PhD a fun one!!!
The big takeaway: Environment diversity > Partner diversity
Training across diverse tasks teaches agents how to cooperate, not just whom to cooperate with. This enables zero-shot coordination with novel partners in novel environments, a critical step toward human-compatible AI.
Our work used NiceWebRL, a Python-based package we helped develop for evaluating Human, Human-AI, and Human-Human gameplay on Jax-based RL environments!
This tool makes crowdsourcing data for CS and CogSci studies easier than ever!
Learn more: github.com/wcarvalho/ni...
Why do humans prefer CEC agents? They collide less and adapt better to human behavior.
This increased adaptability reflects general norms for cooperation learned across many environments, not just memorized strategies.
Human studies confirm our findings! CEC agents achieve higher success rates with human partners than population based methods like FCP and are rated qualitatively better to collaborate with than the SOTA approach (E3T) despite never having seen the level during training.
Using empirical game theory analysis, we show CEC agents emerge as the dominant strategy in a population of different agent types during Ad-hoc Teamplay!
When diverse agents must collaborate, the CEC-trained agents are selected for their adaptability and cooperative skills.
The result? CEC agents significantly outperform baselines when collaborating zero-shot with novel partners on novel environments.
Even more impressive: CEC agents outperform methods that were specifically trained on the test environment but struggle to adapt to new partners!
We built a Jax-based procedural generator creating billions of solvable Overcooked challenges.
Unlike prior work studying only 5 layouts, we can now study cooperative skill transfer at unprecedented scale (1.16e17 possible environments)!
Code available at: shorturl.at/KxAjW
We introduce Cross-Environment Cooperation (CEC), where agents learn through self-play across procedurally generated environments.
CEC teaches robust task representations rather than memorized strategies, enabling zero-shot coordination with humans and other AIs!
Current AI cooperation algos form brittle strategies by focusing on partner diversity in fixed tasks.
I.e. they might learn a specific handshake but fail when greeted with a fist bump.
How can AI learn general norms that work across contexts and partners?