ICML (International Conference on Machine Learning) is an academic conference where researchers publish and discuss machine learning work. Recent coverage includes LittleBit-2, an ICML 2026 follow-up that adds Joint-ITQ initialization to sub-1-bit LLM compression; reports that Nvidia's ICML spotlight paper DreamDojo has code bugs and improved only about 0.5 dB PSNR over Cosmos 2.5; and Nvidia's claim that 145 ICML 2026 papers cited its Nemotron open models. Community threads also discuss paper writing for ICML and AI-generated reviews.
Researchers Turn 109k ML Papers Into Topic PagesA team grouped roughly 109k papers accepted at ICML, NeurIPS, ICLR, CVPR and AAAI since 2016 into about 1,260 topics and extracted results tables from the full text of 79k of them. Topic pages show best benchmark numbers with the source paper, recent ablations and common baselines. It is free to use; the authors warn extracted numbers may not be fully comparable.
Researcher says AI-generated reviews are eroding peer reviewA reviewer for AAAI 2027 and ICLR 2027 says roughly half the reviews they saw appear fully AI-generated or AI-edited, and submitted papers show clear signs of AI writing. The reviewer argues the problem needs far more attention than it currently gets.
Community Discusses How to Write Top-Conference PapersA researcher asks how best to write for NeurIPS, ICLR and ICML audiences and cites Michael J. Black's blog arguing papers should be story-oriented rather than results-oriented, inviting community opinions.
LittleBit Achieves Sub-1-Bit LLM CompressionLittleBit (NeurIPS 2025) factorizes weight matrices into low-rank latent factors, binarizes them, and restores magnitude via lightweight learned scales, enabling 1.0 to 0.1 bits per weight while keeping the original inference architecture. The follow-up LittleBit-2 (ICML 2026) adds Joint-ITQ initialization, improving initialization without adding inference overhead.
Nvidia ICML spotlight paper found to have bugsA Reddit user says Nvidia's robotics world model DreamDojo, an ICML spotlight paper, improved only about 0.5 dB PSNR over Cosmos 2.5 despite using 44k hours of human data and 256 H100s. Released code has bugs reported in GitHub issues affecting pre-training, post-training, and evaluation, making the results suspect.
Nvidia execs to debate open vs closed AI at TechCrunch DisruptNvidia's Nader Khalil and Sydney Sykes will speak at TechCrunch Disrupt 2026 on October 13-15 in San Francisco about how startups should choose between open and proprietary AI models. Nvidia says 145 papers at ICML 2026 cited its Nemotron open models, yet frontier proprietary labs keep advancing, and the choice affects cost, infrastructure and differentiation.
The A.I. Daily