Athul Paul Jacob

I am a Ph.D. student in AI, multi-agent systems and, NLP at MIT CSAIL, advised by Jacob Andreas. I received my master's degree from MIT EECS in 2022, where my thesis was co-advised by Noam Brown and Jacob Andreas.

In 2019, I completed my bachelor's in computer science, combinatorics and optimization at the University of Waterloo, where I was advised by Pascal Poupart. From 2016 until 2018, I was also a visiting student researcher at Mila where I worked under the supervision of Yoshua Bengio.

I have been fortunate to intern at Facebook AI Research several times in 2018, 2020 and 2021 as a researcher, where I was mentored by Noam Brown, Kyunghyun Cho and Mike Lewis. I have also previously worked as a research intern at Microsoft Research in fall 2017 and winter 2018 with Alessandro Sordoni and Adam Trischler. I have also spent time at Google, MIT-IBM Watson AI Lab and General Catalyst Venture Partners as an AI Fellow.

Email  /  Google Scholar  /  LinkedIn  /  Twitter

* indicates equal first author contributions

We study the problem of modeling a population of agents pursuing unknown goals subject to unknown computational constraints. We introduce latent inference budget model (L-IBM) that models agents’ computational constraints explicitly, via a latent variable (inferred jointly with a model of agents’ goals) that controls the runtime of an iterative inference algorithm. In three modeling tasks—inferring navigation goals from routes, inferring communicative intents from human utterances, and predicting next moves in human chess games—we show that L-IBMs match or outperform Boltzmann models of decision-making under uncertainty

We present a model of pragmatic language understanding called ReCo, where utterances are produced and understood by searching for regularized equilibria of signaling game. Across several datasets capturing real and idealized human judgments about pragmatic implicatures, ReCo matches or improves upon predictions made by best response and rational speech act models of language understanding

We introduce a training-free, game-theoretic procedure for language model decoding that improves performance across a number of NLP tasks.

We introduce Cicero, an AI agent that demonstrates human-level performance in the mixed-motive 7-player strategic board game Diplomacy that involves natural language negotiation, strategic coordination and persuation.

We introduce Diplodocus, an AI that achieves expert human-level performance in No-press Diplomacy, a challenging 7-player strategy game that involves both cooperation and competition.

A new algorithm for detecting nonsensical responses in complex applications involving dialogues.

We show that by regularizing search towards a human policy, we get state-of-the-art human prediction accuracies in chess, Go and no-press Diplomacy while being significantly stronger.

We prove that multitask training eliminates semantic drift in a well-studied family of signaling games, and show that multitask training of neural latent language policies (LLPs) in a complex strategy game reduces drift and while improving sample efficiency.

A novel constituency parsing scheme free from compounding errors, while being faster and easier to parallelize.

A hierarchical model for sequential data that learns a tree on-the-fly. The model adapts its structure and reuses recurrent weights in a recursive manner by creating adaptive skip-connections that ease the learning of long-term dependencies.

A principled method for training generative adversarial networks on discrete data that uses the estimated difference measure from the discriminator to compute importance weights for generated samples, providing a policy gradient for training the generator of the network.

A generative adversarial network capable of jointly generating images and their labels.

We introduce several ways of regularizing the GAN training objective, which can dramatically stabilize the training of these models.

Design courtesy of Jon Barron