Skip to main content
Search

Ashok Prasad Neupane

2 related articles

Subscribe to updates
  1. arXiv

    iADD uses early-timestep updates and Feynman-Kac branching to improve reward alignment, diversity, and rare-prompt success in image and 3D scene generation

    The authors propose iADD, a reinforcement-learning post-training method for discrete-time diffusion models that theoretically shows updating early denoising timesteps preserves diversity better than updating only late timesteps, and combines an incremental sparse-timestep curriculum with discrete-time Feynman-Kac path pruning and branching; across rare-prompt image generation, vanishing-point correction, and 3D indoor scene synthesis, iADD improves reward, Inception Score, rare-event success, and AUC over DDPO, B2-DiffuRL, and related baselines, and its components compose with GRPO-style optimization.

  2. arXiv

    Foresight lets a frozen Qwen3-VL-8B plan its own future perception in streaming video, reaching 23.0 joint F1 on OmniPro Online and beating the strongest trained baseline by 9.5%

    The authors introduce Foresight, a training-free dual-stream architecture with two Siamese LLMs sharing weights, input encoders, and a KV cache, where one copy continuously ingests the stream while the other runs ahead over the same causal state to write a plan deciding when to reason next, what to check then, and how densely to sample; with a frozen Qwen3-VL-8B backbone it reaches 23.0 mean joint F1 on OmniPro Online (versus 13.5 for the strongest trained baseline, MiniCPM-o 4.5), improves the backbone by 6.7 points on StreamingBench with a best overall score, and by 15.4 points on OVO-Bench, with the largest gain of 18.7 points on Forward Active Responding where evidence arrives later in the stream.

You’re all caught up