This study analyzes optimizer-dependent training dynamics via Hessian eigenvector displacement and localization, revealing SGD stabilizes directions while Adam induces eigenvector reorganization.

Marcelina Marjankowska, Valerio Modugno, Paolo Barucca

Proposes modality-specific sparse autoencoders and post-hoc alignment to address cross-modal feature heterogeneity, boosting retrieval accuracy to 85.3% on MS-COCO.

Chungpa Lee, Jihoon Kwon, Kyle Min et al.

Dual-Flow RL employs conditional flow matching to jointly model return distribution and multimodal policy, enhancing exploration and value estimation in continuous control.

Qijun Li, Zheng Fu, Qi Song et al.

Introduces TILR, a low-rank subspace method that identifies stable reasoning directions, improving model consistency by ~10% and reducing trajectory variance by 50%.

Arun Vignesh Malarkkan, Manan Roy Choudhury, Utkarsh Byahut et al.

PEHT integrates LoRA-enhanced Transformer with urban mobility and congestion data, achieving state-of-the-art network traffic prediction with over 90% parameter reduction.

Abdolazim Rezaei, Mehdi Sookhak, Mahboobeh Haghparast

COCOLogic-V2 evaluates visual inductive reasoning with truly hard negatives, revealing models' struggles on boundary samples.

David Steinmann, Antonia Wüst, Kristian Kersting et al.

LLawCo framework uses failure reflection to extract behavioral laws, improving multi-agent cooperation success rate by 4.5%.

Qinhong Zhou, Chuang Gan, Anoop Cherian

NormGuard applies reward-preserving norm constraints to suppress velocity norm inflation in RL fine-tuning, improving image quality.

Tianlin Pan, Lianyu Pang, Cheng Da et al.

PerturbCellRL uses four biological verifiers to post-train scDFM, improving single-cell consistency after 1,600 RL steps while preserving population-level competitiveness.

Dongxia Wu, Mingyu Li, Yuhui Zhang et al.

RAC reduces policy bias by 47.9x via delay-aware reward correction for RLHF.

Arnav Raj

RiVER leverages score-based optimization with instance ranking and winner-heavy rewards, improving LLMs on algorithmic tasks by 8.9%-9.4%.

Yingyu Lin, Qiyue Gao, Nikki Lijing Kuang et al.

Error-conditioned neural solver (ENS) uses PDE residuals as direct inputs, achieving 10× accuracy improvement and avoiding costly optimizers.

Haina Jiang, Liam Wang, Peng-Chen Chen et al.

Introduces two activation regularizers for Top-k sparse autoencoders, enhancing interpretability and robustness without sacrificing reconstruction quality.

Nathanaël Jacquier, Maria Vakalopoulou, Mahdi S. Hosseini

This paper establishes the algorithmic equivalence between Blackwell approachability and Gradient Equilibrium (GEQ), unifying online decision frameworks.

Brian W. Lee, Nika Haghtalab, Michael I. Jordan et al.

Introduces decision-alignment framework for uncertainty quantification (UQ), critiques traditional metrics, and proposes prior-weighted utility (PWU) metrics, validated through experiments showing superior decision utility alignment.

Annika Schneider, Tommy Rochussen, Joshua Stiller et al.

Proposes FL-based quantization methods reducing communication and computation costs, enhancing robustness and privacy.

Farwa Ikram, Dipanwita Thakur, Antonella Guzzo et al.

Introduces a preference-conditioned Bellman operator using Chebyshev scalarization for deterministic Pareto-optimal policies in MOMDPs, ensuring full Pareto front coverage.

Aniruddha Joshi, Niklas Lauffer, Sanjit Seshia

Introduces Red Queen Gödel Machine (RQGM), a co-evolving framework with learned evaluators, achieving 1.35-1.86× efficiency gains in code, science, and proof tasks under non-stationary utilities.

Alex Iacob, Andrej Jovanović, William F. Shen et al.

This study reveals the asymmetric dynamics of rule survival in pretraining models, with support frequency predicting rule fate and demonstrating irreversibility via editing interventions.

Juliana Li, Diya Sreedhar

ASAP integrates diverse optimizers with agent-system co-design, optimizing wall-clock time for ML hyperparameter tuning.

Taicheng Guo, Haomin Zhuang, Kehan Guo et al.