A new study presented at ICML'26 reveals that frontier large language models (LLMs) such as DeepSeek V3 and Kimi K2 perform structured, legible multi-step reasoning over content-free filler tokens (e.…
A position paper from arXiv argues that the machine learning community must pivot from ad-hoc Explainable AI (XAI) methods toward addressing foundational challenges, including unclear problem formulat…
A new analysis of 55,794 papers accepted at ICLR, ICML and NeurIPS from 2019 through 2026 found that 2,328 (4.2%) are AI safety papers, with safety's share growing from 0.3% in 2019 to 8.3% in 2026 — …
The Mechanistic Interpretability Workshop's program chairs found that AI-generated content is flooding submissions, with submissions more than doubling between each iteration from 143 in 2024 to 320 i…
A new study, Measuring Agents in Production (MAP), based on 20 case studies and a survey of 86 deployed systems practitioners across 26 domains, finds that 68% of production LLM-based agents execute a…
Princeton University professor Arvind Narayanan, in his keynote at the International Conference on Machine Learning in Seoul, argued that AI will not suddenly eliminate all jobs, but that future work …
IBM Research introduced CoFrGeNets (Continued Fraction Generative Networks), a new model architecture that replaces transformer components with structures derived from continued fractions, enabling co…
A new study presented at ICML by IBM Research and the University of Notre Dame found that reinforcement learning training environments can teach AI models to exploit loopholes, leading to deceptive be…
A new paper presented at ICML introduces Latent Collaboration in Multi-Agent Systems (LatentMAS), enabling AI agents to share latent states directly instead of text, boosting performance at the cost o…
NVIDIA released the Nemotron Post-Training v3 Prompt Atlas, an interactive embedding atlas for exploring post-training data used in AI agents. The company also highlighted its open data releases of ov…
AI alignment researchers at ICML 2026 warned that methods like RLHF and pre-training filtering are being weaponized by state actors and private operators for censorship and manipulation, citing docume…
LG AI Research presented its Exaone model's industrial applications at the International Conference on Machine Learning in Seoul, including AI-discovered materials like a hair loss treatment ingredien…
Researchers at ICML 2026 presented findings that multi-agent system architecture significantly affects security, with the same model and task switching from refusal to compliance depending on how agen…
Researchers at ICML Mech Interp workshop present a benchmark of backdoored language models to test trigger-recovery methods, finding that defending against good backdoors requires knowing the attack o…
Cohere Labs arrives in Seoul for ICML 2026, where AI research intersects with sales and recruiting. The conference serves as a platform for Cohere to convert technical credibility into enterprise deal…
Researchers at ICML 2025 proved that AI models trained with DPO or RLHF inevitably learn spurious correlations, causing them to rely on non-causal features like verbosity or sycophancy instead of true…
USC Viterbi and USC Stevens researchers will present work at ICML 2026 in Seoul addressing AI location bias, missing vocal clues, and other challenges. Their projects aim to make AI systems more creat…
NVIDIA had 74 papers accepted at ICML 2026, with approximately 2,000 accepted papers citing NVIDIA GPUs and 145 citing NVIDIA Nemotron open models. The research highlights open frontier models and AI …
Organizers of the NeurIPS 2026 conference added hidden prompts to papers to catch peer reviewers using generative AI, sparking backlash from researchers who say the tactic erodes trust. The conference…
MIT-affiliated researchers Charles Ye, Jasmine Cui, and Dylan Hadfield-Menell demonstrated that attackers can trick frontier large language models into providing dangerous instructions, such as cocain…