5 appearances on the backlist front page in the last 30 days.
DistIL starts from a simple weakness in RLVR: most of the signal is still one bit at the end. Instead, it uses richer feedback: execution traces, tool outputs, expert corrections, ground-truth solutions, or model critiques. It replaces re
Using a generative flow model to solve a difficult signal-processing optimization problem and output deployable FIR filters. Nice. Paper: https:// arxiv.org/abs/2606.04570
MetaPoint is a clean fix for spatial control in image generation: make the coordinate itself a token. It uses the modelβs existing positional encoding instead of new architecture, large coordinate vocabularies, or custom attention masks.
Prompt-free self-training can work when the synthetic corpus is compatible with the student. The setup is intentionally minimal: BOS-only text sampled from a base model, then used for continued training without prompts, verifiers, reward m
HyperParallel-MoE is an Ascend-specific scheduling system for MoE training. Ascend A3 exposes separate AIC matrix units and AIV vector/communication units, but standard MoE execution still runs Dispatch, GMM, SwiGLU, and Combine as seriali