Benchmarking five optimizers (SGD, Adam, Adagrad, Adadelta, CGD) with three memory strategies across four transformer architectures; gradient accumulation proved most robust, optimizer performance varies by architecture.

Sarthak Mahapatra, Zihan Zhou, Khatoon Khedri et al.

Proposes QCPIKAN, a hybrid quantum-classical physics-informed network, for solving fuzzy PDEs with improved accuracy and theoretical guarantees.

Xiang Rao, Yuxuan Shen

Proposes dp_knapsack_sliding_hirschberg combining sliding window and Hirschberg’s algorithm for memory-efficient exact 0/1 knapsack in PyTorch, enabling larger models.

Jędrzej Maczan

Proposes FCAC framework combining evidence freshness and randomized audits, achieving 84.4% automation with controlled review workload.

Jie Deng

Introduces Graph Circuit Learning (GCL), using GNNs for cross-model circuit localization, achieving median edge AUROC of 0.902.

Chester Tan, Moritz Lampert, Courtney Maynard et al.

Biologically-inspired modular SNN (D-SNN) achieves 98% accuracy on MNIST, with tenfold fewer parameters, and resists catastrophic forgetting.

Maksim Bazhenov, Serafim Grubas, Vakhtang Putkaradze

Introduces quantum-inspired complex phase states, reducing training steps by ~3x, validated on byte-level text datasets with superior efficiency.

Ahmed Nebli, Hadi Saadatdoorabi, Christopher Keibel et al.

Momba enhances multi-objective RL by integrating SimbaV2 architecture with feature normalization, weight normalization, and distributional critic, achieving higher sample efficiency and better solution sets.

Adam Štafa, Santeri Heiskanen, Petr Novotný et al.

Modular TTT models test-time training as a composable graph, improving flexibility and efficiency in sequence modeling.

Bohao Tang, Zhen Qin, Yuqi Pan et al.

Introduces block-hybrid attention to retrofit LLaDA 2.1-mini, achieving 1.7× inference speedup with minimal performance loss.

Jinha Kim, Younghun Roh, Jaeyeon Kim

Proposes Ranking-based Reward Construction (RRC) to leverage generative reward models via relative rankings, improving RL performance with significant gains on benchmarks.

Chenglong Wang, Ziming Zhu, Yifu Huo et al.

Proposes SG-TULA, a subgradient-based Langevin sampling method with explicit convergence bounds for non-convex, non-smooth, superlinear potentials.

Iosif Lytras, Nikolaos Makras, Sotirios Sabanis

LC-GRPO combines inference-aligned ODE with Langevin correction, reducing training-inference gap and improving reward optimization in flow-based RL.

Yingqing Guo, Hui Yuan, Zijian He et al.

Sharding partitions requirements into smaller groups, boosting accuracy and robustness in LLM oversight, outperforming holistic models.

Victor Akinwande, J. Zico Kolter, Aran Nayebi

Introduces LLM-guided symbolic regression (SR) framework for automated kinetic model discovery, reducing iteration count by up to 79.3%.

Roberto Aliaga Medina, Paulina Quintanilla, Antonio del Rio Chanona

Marginal matching does not ensure factorized sampling; model leakage persists.

Duong Bach, Hai Nguyen Hong, Cuong Do

Introduces variational bounds for the finite-temperature perceptron trained on Gaussian mixtures, deriving upper and lower bounds via interpolation and log-concavity, with fixed point equations for energy and error estimates.

Francesco Camilli, Pierluigi Contucci, Federica Gerace et al.

OCSD compares matched replay views to calibrate environment feedback, significantly improving multi-task reinforcement learning performance.

Yi Yang, Cong Qin, Xiaodan Liu et al.

BinaryPC employs data-aware binary principal components for training-free sparse attention, maintaining accuracy and boosting decoding throughput by 3.56×.

Daohai Yu, Zhanpeng Zeng, Keyu Chen et al.

Omega-S uses graph topology metrics to regularize LLM fine-tuning, improving memory retention by controlling weight connectivity diversity.

Alberto Acedo