Publications
RLPE: Teaching LLMs to Prompt with Reinforcement Learning
Training a 9B-parameter model with reinforcement learning to generate prompts for other models, outperforming frontier models on 2/3 benchmarks.
Rahul Singhal, Karime Maamari, Pradyumna Tambwekar
NeurIPS 2026, Pre-to-Post Workshop
Paper · Blog post coming soon
PrefPO: Pairwise Preference Prompt Optimization
Optimizing prompts through pairwise preferences, achieving performance competitive with SOTA prompt optimizers without labeled examples while reducing attempts to game automated evaluations.
Rahul Singhal, Pradyumna Tambwekar, Karime Maamari
COLM 2026
HARDEN: Constrained Evolutionary Search for Harder, Answer-Preserving Evaluation Cases
Using constrained evolutionary search to make evaluation cases more realistic and challenging while preserving correct answers, exposing model failures that existing benchmarks miss.
Aditya Kumaran, Rahul Singhal, Karime Maamari, Amine Mhedhbi, Pradyumna Tambwekar
NeurIPS 2026, EvoRobust Workshop