Layer-adaptive regularization based on Hessian top eigenvalues improves continual learning by reducing forgetting.

Brian B. Moser, Ahmed Anwar, Tobias Christian Nauen et al.

SubZero+ enhances zeroth-order optimization stability with large learning rates, outperforming baselines on SuperGLUE.

Ziming Yu, Shuyao Xiao, Xingyu Zhao et al.

AATC employs transform coding and reverse waterfilling for KV cache compression, achieving near-lossless accuracy at 5.8× compression.

Hannah Laus, Claudio Mayrink Verdun, Hao Wang et al.

CForce enhances early-stage prediction reliability via self-trajectory consistency, significantly improving high-parallelism decoding in dLLMs.

Yuji Ren, Chenkai Xu, Zhuocheng Gong et al.

Event-driven Transformer-based DRL framework for multi-depot dynamic VRP, trained with behavior cloning and PPO, improves stability and responsiveness.

Faezeh Ardali, Gerald M. Knapp

Vero benchmark assesses AI's ability to jointly generate verified multi-module repositories; strongest model solves 27 out of 43 instances.

Zhe Ye, Hantao Lou, Yuechun Sun et al.

Intervention-aware clinical world model encodes patient anatomy into a 3D latent state, evolving it through asynchronous post-intervention events for long-term recurrence prediction, achieving AUROC 0.756 on DECAAF-II.

Yunsung Chung, Yingshuo Liu, Abboud F. Hassan et al.

Using the SbCD model, complete crystal structures are generated via Markovian jump diffusion, excelling on MP20.

Van Khoa Nguyen, Alexandros Kalousis

LOPD introduces learnable latent privileged context, outperforming OPSD variants with less than 30% rollout budget.

Guibin Zhang, Jiayang Lyu, Ran Sun et al.

Introduces Action-Conditioned Predictive Consistency (ACPC) to diagnose robustness of JEPA world models against visual perturbations, with theoretical bounds and empirical validation.

Guo An, Zijing Wu, Honghua Dong et al.

Introduces WMRL, replacing environment execution with a world model, accelerating AutoResearch agent training 3-4×, surpassing standard RL.

Xiyuan Yang, Sheikh Sarwar, Jingru Cheng et al.

Proposes a formal reward design framework combining objectives, causal features, and preference queries, ensuring conflict-free linear reward parameters via geometric optimization.

Di Yang Shi, W. Bradley Knox

Proposes an efficient near-optimal adversarial m-set bandit algorithm with high-probability regret bound of O(√dT log(K/δ)), avoiding exponential complexity.

Francesco Bacchiocchi, Tommaso Cesari, Roberto Colomboni

ScreenShot employs a hierarchical transformer pretrained on 40 drug screening datasets, enabling few-shot prediction and active learning to significantly improve combination drug screening efficiency.

Antoine de Mathelin, Christopher Tosh, Wesley Tansey

MARCH extends recurrent memory via content routing and periodic state checkpoints, improving long-range dependency modeling.

Ming Zhang, Kaisen Yang, Shu Yu et al.

Introduces low-interaction rank functions for unified dual-encoder design, validating spectral decay and normalization effects.

Zijian Zhao, Sen Li

Causal optimization framework using Transformer and Bayesian neural network improves recommendation incrementality by 7.20%.

Changshuai Wei, John Bencina, Phuc Nguyen et al.

Macaron-V1 employs Mixture-of-LoRA architecture with recursive self-improvement, achieving continuous learning and outperforming benchmarks.

Mind Lab, :, Vin Bo et al.

This work demonstrates that random untrained transformers can achieve universal approximation via soft prompts, leveraging kernel methods for theoretical guarantees.

Alexander Hsu, Rongjie Lai

Proposed TrackEFG algorithm achieves switching regret $ ilde{O}((1/ρ+ρK)√HAT)$ with per-trial complexity $O(HB)$.

Stephen Pasteris, Rahul Savani, Theodore Turocy