KronQ employs Kronecker-factored Hessian approximation with gradient covariance, achieving state-of-the-art 2-bit quantization, e.g., perplexity 7.93 on LLaMA-3-70B.

Donghyun Lee, Yuhang Li, Ruokai Yin et al.

Proposes a circuit-based mechanistic interpretability framework using SAEs and transcoders to disentangle polysemantic features in Transformer models.

Pranav Sawant, Jakub Krejčí

Identifies reward hacking in reference-free LLM judges via hidden-anchor audit; “commit-answer-first” strategy effectively reduces false positives.

Chenyu Zhou

Proposes TuneNNGen, leveraging source models and LLMs to boost CIFAR-10 accuracy from 23.98% to 50.49%.

Kabir Dev Paul Baghel, Radu Timofte, Dmitry Ignatov

Using Classifier Discrimination Score (CDS) to address class overlap in single-cell perturbation data, significantly improving identification accuracy.

Youssef Marrakchi, Davide D'Ascenzo, Sebastiano Cultrera di Montesano

MANCE leverages manifold constraints for nonlinear concept erasure, significantly reducing information leakage while preserving other concepts.

Matan Avitan, Yoav Goldberg, Yanai Elazar

Prefix-GRPO improves small models in long-horizon environments by reusing teacher trajectory prefixes and online continuation.

Yihan Wang, Zhong Guan, Haoran Sun et al.

Proposed a partial bidirectional BDLM Mamba-attention hybrid, boosting long-context inference throughput with parameters from 87M to 350M.

Pranshu Chaturvedi, Parth Shroff, Tarun Suresh et al.

HO-HE post-generation curation balances fidelity and diversity, reaching 95% on CIFAR-10 with 300K synthetic images.

Disheng Liu, Tuo Liang, Chaoda Song et al.

Proposes distribution-wise reward with subset-replace strategy for RL fine-tuning, reducing FID from 8.30 to 5.77.

Ruihang Li, Mengde Xu, Shuyang Gu et al.

Zeus is a tuning-free foundation model for time series, using multi-scale Transformer and multi-objective masking to excel across tasks.

Yisong Fu, Zezhi Shao, Chengqing Yu et al.

Study finds SDPO accelerates in-domain learning under specific conditions but struggles in cross-domain scenarios.

Meng Wang, Haohan Zhao, Wenzhuo Liu et al.

Active-GRPO combines active imitate-reinforce and dynamic referencing, boosting SR×Sim from 0.0959 to 0.1773 in molecular optimization.

Xuefeng Liu, Mingxuan Cao, Qinan Huang et al.

Introduces signed-permutation coordinate transport for RMSNorm transformers, recovering 91.1% of coordinates.

John Sweeney

AETDICE unifies AET framework for offline nonlinear multi-objective RL, bridging SER and ESR paradigms with a novel decomposition approach.

Woosung Kim, Youngjun Suh, Jinho Lee et al.

HSAP integrates hierarchical sequence-aware parallelism with JIT-compiled SAP for ultra-long sequences, outperforming SOTA methods.

Songxin Zhang, Zejian Xie, Zhuoyang Song et al.

Arko-T is a 4B-parameter transformer that maps natural language directly into executable, editable parametric CAD programs, outperforming seven frontier LLMs on 12 metrics.

Liang Wang, Zhaoyang Xi, Zekai Xiang et al.

FlowAWR performs advantage-weighted velocity rectification without SDEs or CFG, reaching PickScore 24.12 in 1.2k steps on SD3.5-Medium.

Zheming Fu, Ruizhe He, Wei Shang et al.

Develops KL divergence bounds for acceptance criteria in speculative decoding, applicable to greedy, relaxed, and tree-based decoding, enhancing practical inference reliability.

Aaryam Sharma

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