What actually shipped in agent engineering, pulled from the labs, arXiv and Hacker News.
See who we follow →Study compares performance of diffusion-based and autoregressive language models across NLP tasks.
Categorical flow maps enable continuous diffusion models for discrete data generation with faster sampling.
Arbitrage uses advantage-aware speculation to reduce computational cost of long chain-of-thought reasoning in LLMs.
Apple released DeepAmbigQA, a benchmark for testing LLM completeness on ambiguous multi-hop questions requiring disambiguation and reasoning.
Apple proposes Deep Low-Rank Residual Distillation to lock pretrained weights in open-weight language models.
Apple studied outlier tokens in Diffusion Transformers for image generation and their role in attention.
Apple studies preference alignment in multimodal LLMs to reduce hallucination in image understanding tasks.
UMAP's internal k-nearest-neighbor graph encodes high-dimensional data manifold structure before 2D projection distortion.
Apple presents MoMo, a robot learning framework using spatiotemporal action tokenization for manipulation tasks.
Decoupled temporal depth diffusion transformers enable memory-efficient on-device audio synthesis.
One pattern, one tradeoff, one production failure story. A short weekly briefing for people building agentic systems.
Weekly email, one-click unsubscribe. We only use your address to send the briefing.
