This paper introduces robust diffusion operators and hidden-state re-propagation, significantly improving pre-propagation GNN performance on heterophilic graphs.

Zichao Yue, Zhiru Zhang

UTTSI scales inference by uncertainty, achieving 5.3% relative online CTR gain at about 2.8× average cost.

Moyu Zhang, Yun Chen, Yujun Jin et al.

Establishes global linear convergence of entropy-regularized softmax policy gradient in continuous-space MDPs via non-uniform Polyak-Łojasiewicz inequality, extending beyond tabular cases.

Ziyue Chen, David Šiška, Lukasz Szpruch

Proposed CFCP method improves local coverage in multi-class classification using cluster frequency.

Tomer Lavi, Bracha Shapira, Nadav Rappoport

Proposes a template-controlled DiD protocol to accurately measure alignment-induced activation shifts, removing formatting confounds.

Yuki Nakamura

ECHO integrates environment observation prediction with policy optimization, doubling task success rate from 2.70% to 5.17% on Qwen3-8B in TerminalBench-2.0.

Vaishnavi Shrivastava, Piero Kauffmann, Ahmed Awadallah et al.

DiLaDiff combines latent continuous diffusion with discrete masking, achieving high-quality, fast language generation.

Jean-Marie Lemercier, Tomas Geffner, Karsten Kreis et al.

Proposes latent policy gradients to predict out-of-distribution RL behavior with 89.73% accuracy.

Jason Ross Brown, Edward James Young

RankElastor combines parameterized full mixing and GLU P-FFNs, improving AUC by over 0.001 on Criteo and Avazu versus the strongest baseline.

Guoming Li, Shangyu Zhang, Junwei Pan et al.

GEMQ uses global linear programming and efficient router fine-tuning for mixed-precision quantization of MoE LLMs, reducing memory usage.

Jianing Deng, Song Wang, Dongwei Wang et al.

Vector Policy Optimization (VPO) trains diverse policies to improve test-time search, achieving over 20% gains on best@k metrics across multiple tasks.

Ryan Bahlous-Boldi, Isha Puri, Idan Shenfeld et al.

Introduces curiosity-driven 3D exploration using persistent 3D Gaussian Splatting world model and Transformer policy, achieving 74.94% 3D coverage on HM3D.

Lily Goli, Justin Kerr, Daniele Reda et al.

The Matching Principle unifies nuisance-robust learning by estimating deployment nuisance covariance and regularizing encoder Jacobian accordingly; validated on 7B-parameter Qwen2.5-7B.

Vishal Rajput

Proposes EMO-STA, a multi-task evolutionary framework leveraging shared program archives, improving efficiency and generalization in LLM-guided program discovery.

Halil Alperen Gozeten, Xuechen Zhang, Emrullah Ildiz et al.

Neural compiler converts symbolic physics programs into exact, differentiable PyTorch modules, enabling precise, composable hybrid models with minimal parameters.

Lucas Sheneman

Proposes a three-step evaluation protocol addressing biases of pointwise metrics in multimodal inverse problems, emphasizing distribution fidelity and calibration.

Mads H. Baattrup, Jörn Bach, Laurids Jeppe et al.

Data gating is critical for self-play RL stability; reward signals alone cannot prevent collapse.

Sophia Xiao Pu, Zhaotian Weng, Chengzhi Liu et al.

Introduces stable-worldmodel (swm), a unified platform supporting reproducible research with high-performance data loading, multiple world model baselines, and diverse environment benchmarks.

Lucas Maes, Quentin Le Lidec, Luiz Facury et al.

Proposes the multi-class unhinged loss as the unique convex symmetric loss, improving robustness to noisy labels.

Alexandre Lemire Paquin, Brahim Chaib-Draa, Philippe Giguère

Quant.npu improves mobile NPU inference efficiency via fully static quantization, reducing latency by 15.1%.

Jinghe Zhang, Daliang Xu, Chenghua Wang et al.