The blog
Two parallel tracks for two kinds of readers: deep dives for engineers, accessible explainers for everyone. Pick the one that matches where you are today.
Notes from Day 2's first oral session at ICLR 2026. Five papers on learning rate schedules, MoE sparsity, curriculum data, softmax expressivity, and pre-training under infinite compute.
Notes from the second oral session at ICLR 2026. Six papers on reward hacking detection, multi-turn reliability (the Best Paper), micro-benchmarking, value-difference generation, preference interpretability, and mutual-judgment alignment.
Notes from the first oral session at ICLR 2026. Five papers on active reasoning, long-context memory, CoT verification, retriever training, and model merging.
Opening keynote at ICLR 2026 from USC and Google DeepMind's Maja Mataric — a case that intelligence without a body misses the point, and that the robots we actually need aren't the ones trying hardest to look human.
Tune in to the details
Notes from Day 2's first oral session at ICLR 2026. Five papers on learning rate schedules, MoE sparsity, curriculum data, softmax expressivity, and pre-training under infinite compute.
Notes from the second oral session at ICLR 2026. Six papers on reward hacking detection, multi-turn reliability (the Best Paper), micro-benchmarking, value-difference generation, preference interpretability, and mutual-judgment alignment.
Notes from the first oral session at ICLR 2026. Five papers on active reasoning, long-context memory, CoT verification, retriever training, and model merging.
Opening keynote at ICLR 2026 from USC and Google DeepMind's Maja Mataric — a case that intelligence without a body misses the point, and that the robots we actually need aren't the ones trying hardest to look human.
A two-track ML blog: reproducible engineering deep dives for practitioners, and clear explainers for everyone.
