here are my recent publications
A framework for improving LLM reasoning through adversarial self-play where a Setter generates challenging problems and a Solver attempts to solve them, achieving up to +10% on MATH and +8% on MBPP with cross-domain transfer.
A self-questioning framework that trains a VLM to decompose compositional visual questions into sub-questions using GRPO, without any reasoning demonstrations. Applied to a 3B-parameter model on CLEVR and A-OKVQA, both self-questioning and standard RL substantially improve accuracy over the untrained model.
Investiga ting how large language models can reason about and generate code in esoteric programming languages through in-context learning, without requiring fine-tuning.
A multimodal deep learning approach integrating radar and satellite imagery to detect illegal selective logging activities with high accuracy.
Teaching models to identify and recover from flawed reasoning chains using reinforcement learning techniques to improve mathematical problem-solving capabilities.