Ayush Jain

We show that language models can explicitly self-localize errors in incorrect reasoning when it is structured as discrete, semantically coherent thought steps.

We propose an inverse RL method for learning from constrained demonstrators and finding shorter trajectories to the goal.

We propose reset-based policy optimization methods that improve credit assignment in multi-step language model reasoning by resampling continuations from intermediate reasoning states.

Certain tasks produce much larger gradients during RL post-training of multi-task LLMs, biasing model updates without translating to greater learning gains.

Language models improve by reinforcing their own debate consensus across diverse reasoning paths.

Actor-free Q-learning in continuous action spaces by learning a wire-fitted Q-function.

We identify that TD3 gets stuck in local optima in tasks with complex Q-functions and propose a new actor architecture to find better optima.

We introduce behavior-sharing for efficient multitask reinforcement learning, complementary with parameter-sharing and data-sharing.

For optimal decision-making under a varying action space, we learn relations between available actions using a graph-attention policy architecture.

Our RL framework enables agents to solve sequential decision-making tasks even when available actions have not been seen before.

This work investigates the extent to which word order choices in Hindi are influenced by the drive to minimize information variance in a sentence.