AI Briefing#icml-2026
The latest AI and developer news about #icml-2026, with the original source and a short summary.
The latest AI and developer news about #icml-2026, with the original source and a short summary.
The technique trains small neural networks offline to route and rewrite queries, achieving up to 8x faster search and higher recall than centroid-based indexing on the Natural Questions dataset.
The method identifies that FP4 errors stem from low-entropy tokens and uses trace-aligned training to reach 65.94% AIME-120 accuracy on R1-Qwen-14B.
The method corrects posterior probabilities at test time without retraining, maintaining performance on balanced data while fixing recall collapse in imbalanced scenarios.
The method, presented at ICML 2026, generates observation and action models via LLMs to reconfigure the agent-environment interface, outperforming human-designed baselines in locomotion and manipulation tasks.
Sim-GEAR achieves a 91.1% computational speedup and 62.6% memory reduction compared to GEAR by replacing curvature matching with metric matching.
The official implementation supports compression down to 0.1 bits per weight across models including Llama 3, Phi-4, and Qwen3.
The study reveals how TokenGT models reconstruct graph incidence through node-ID matching to solve tasks like degree calculation and shortest-path distance.
Google Research and UIUC unveiled R4T, which distills RL-trained query fan-out behaviors into a 53.9M parameter diffusion model, reducing search latency by 12–20x.
LG AI Research has developed 'MCFlow,' a single model capable of performing multiple crystal generation tasks—including composition, structure, and atom generation—and has been accepted to ICML 2026.
LG AI Research has developed MCFlow, an integrated model for inorganic material generation capable of element substitution, which has been accepted to ICML 2026.
The ToMoE technique has been released, converting dense LLMs into MoE via dynamic structural pruning to reduce active parameters without performance degradation.
A position paper from ICML 2026 warns that interpreting LLM intermediate tokens as human thought processes is dangerous.
The paper proposes the GRAM technique, which addresses specific knowledge and controls it on a per-module basis during the pretraining stage.
Microsoft has unveiled Memora, a memory framework that separates stored content from retrieval methods to enhance AI agents' ability to perform long-term tasks.
The paper analyzes a 'reversed alignment' phenomenon in which image generation models prioritize their learned aesthetic standards over the user's intent.