First student author
Geo-RL for long-form image captioning
Contributed to problem formulation, geometric diversity reward design, credit assignment, implementation, evaluation, and ablations. Geo-RL uses a CLIP-weighted RBF-DPP log-determinant reward, pairwise repulsion, and PPO to balance semantic coverage with visual fidelity. Leave-one-out marginal advantages are computed from one Cholesky factorization. On DeCapBench, compared with LLaVA-1.5 7B MLE, Distinct-2 increased from 0.66 to 0.84 and SVS from 2.20 to 3.55, CLIPScore rose from 0.74 to 0.81, and CHAIRi decreased from 8.3% to 7.1%.
