Gen-Searcher trains a search-augmented image generation agent via supervised and reinforcement learning, yielding about 16-point gains on knowledge-intensive benchmarks.
Equilibrium Matching learns implicit energy landscapes for optimization-based sampling, surpassing diffusion models with 1.90 FID on ImageNet 256x256 while supporting denoising, OOD detection, and composition.
Equilibrium Forcing removes noise conditioning from video diffusion to enable adaptive closed-loop inference that improves generation quality and consistency.
AutoSpec is a neural framework that discovers iterative spectral algorithms via self-supervised prediction of recurrence coefficients, yielding order-of-magnitude speedups over classical baselines.
Spectral Feedback iteratively selects protein tokens to re-mask and resample using sparse Fourier edit-set value functions, improving inverse-folding stability by up to 32.3% at test time.
AlphaQ allocates MoE quantization bits without calibration using heavy-tailed spectral analysis, outperforming calibration-based methods and achieving near full-precision accuracy at 3.5-bit average precision.
ICWBench reveals current LLMs fail at in-context watermarking, and self-distillation with reinforcement learning raises watermark detectability near perfect while preserving quality.