ICML 2026 Paper Spotlight: Mobile GUI Agents, Diffusion LLM Decoding, and Graph‑Based Dataset Pruning

Three frontier papers accepted at ICML 2026 are examined, each addressing a key challenge in deploying large models: evaluating mobile GUI agents, optimizing diffusion LLM decoding, and accelerating training through dataset pruning.

VenusBench‑Mobile proposes a challenging online benchmark that defines evaluation tasks by user intent and provides a capability‑diagnostic labeling scheme. Large‑scale tests on current mobile GUI agents reveal a significant performance gap compared with existing benchmarks, with failures mainly due to weak perception and memory and instability under environment changes.

Lookahead Path Likelihood Optimization for Diffusion LLMs introduces Path Log‑Likelihood (Path LL) as a trajectory‑level objective that correlates strongly with downstream accuracy. The authors design POKE, a K‑step optimistic likelihood estimator, and embed it in a Sequential Monte Carlo (SMC) search (POKE‑SMC). Experiments on LLaDA‑8B‑Instruct and LLaDA‑1.5‑8B across six benchmarks (GSM8K, MATH500, HumanEval, MBPP, Countdown, Sudoku) achieve 3%–5% average accuracy improvement with comparable inference cost, and larger gains under aggressive parallel decoding.

Selecting Samples on Graphs models a training dataset as a weighted graph, converting pruning to a maximum‑weight‑clique problem. A normalized greedy algorithm based on marginal sample benefit provides provable near‑optimal guarantees under mild constraints. Empirical results on ImageNet‑1k with ResNet‑50 show over 40% reduction in training time while preserving model accuracy, outperforming traditional pruning methods.