E2S Finetuning: From Expert Demonstrations to On-Policy Learning
E2S learns an amortized GFlowNet sampler to turn off-policy expert responses into more on-policy SFT data, offline for dataset creation or online for post-training.
Notes & essays
Blog
Notes on reasoning, learning, and building AI systems.
Unified Multimodal Reasoning: Can Different Modalities Share One Thought Space?
Different modalities do not need to become the same. Uni-LaDiR explores a shared latent interface for reasoning, shaped by what future steps need and generated through diffusion.