Papers

FuseReg: Regularizing Layer Fusion Mitigates the

Representation autoencoders must pick which pretrained encoder layers feed both the pixel decoder and the generator, a choice that trades fine detail for generation quality. FuseReg trains on random layer subsets, yielding a decoder that handles full, sparse, and single‑layer fusions and improves PSNR, while decoder swapping cuts unguided gFID by 27% and joint regularization lowers it another 29% on diffusion models.

Hongyang Du, Yunfei Xie2026-09-25

Read

arXiv3 min read

Learning to Stop without Learning to Stop

Reasoning models often produce overly long chains of thought, making inference expensive. A self‑supervised confidence fine‑tune reduces token usage by up to a quarter while keeping accuracy unchanged across multiple model families and reasoning domains.

Parsa Hosseini, Akasha Tigalappanavara2026-09-25

Read

arXiv3 min read

Gap-free Differentially Private PCA for Gaussian Data

The work addresses the challenge of performing principal component analysis under differential privacy when the data follow a Gaussian distribution, and it does so without assuming any spectral gap. It presents an algorithm that attains privacy guarantees while delivering error bounds that are independent of the eigengap.

Alina Ene, Huy L. Nguyen2026-09-25

Read

arXiv3 min read

First-Order Stationarity of Reverse Diffusions

The authors aim to build a first‑order theoretical framework that links optimization guarantees to diffusion‑based sampling. They prove exponential Fisher‑divergence contraction for SDE reverse flows and derive averaged first‑order stationarity bounds for discretized overdamped and underdamped samplers.

Zhifeng Chen, Chenyang Jiang2026-09-25

Read

arXiv3 min read

Statistical attribute alignment for black-box

The work tackles the challenge of steering the statistical distribution of a protected attribute in black‑box generative models toward a user‑specified target. Empirical tests on text‑to‑image and geocoded persona generators demonstrate that the proposed post‑processing markedly improves alignment over prompting alone.

Kevin Jiang, Morgane Austern2026-09-25

Read

arXiv3 min read

User Model Extraction via Belief Self-Distillation

LLMs silently infer user attributes and adapt their responses, but these internal beliefs are hard to inspect or control. The new Belief Self-Distillation framework extracts a compact user representation that can be decoded and written back, enabling stronger interventions on model behavior such as refusal decisions.

Ali Holmov, Yiran Huang2026-09-25

Read

arXiv3 min read

New LoRA Skills Should Read but Never Write

Combining independently trained LoRA adapters into a single model is hampered by weight‑space interference and costly retraining. READ eliminates that overhead, delivering over 20‑point gains on SuperGLUE and more than 7‑point improvements on a domain benchmark across several model families.

Zeyan Li, Panqi Yang2026-09-25

Read

arXiv3 min read

GraphWrit3R: End-to-End 3D Scene Graph Writing

GraphWrit3R tackles 3D scene graph generation by eliminating multi‑stage pipelines and the need for ground‑truth object boxes, directly converting raw point clouds, Gaussian splats, or both into a complete JSON graph of objects, attributes, and relationships. On the 3DSSG benchmark it reaches state‑of‑the‑art recall for object classes, predicates, and triplets, surpassing approaches that still depend on ground‑truth detections.

Luka Milivojevic, Nikola Popovic2026-09-25

Read

AAAI3 min read

OpenDriveVLA: Towards End-to-end Autonomous Driving

We present OpenDriveVLA, a Vision-Language Action (VLA) model designed for end-to-end autonomous driving, built upon open-source large language models. OpenDriveVLA generates spatially-grounded driving actions by leveraging multimodal inputs, including both 2D and 3D instance-aware visual representations, ego vehicle states, and language commands.

Xingcheng Zhou, Xuyuan Han2026-03-1425 cites

Read

AAAI3 min read

AdaptCLIP: Adapting CLIP for Universal Visual Anomaly

Universal visual anomaly detection aims to identify anomalies from novel or unseen vision domains without additional fine-tuning, which is critical in open scenarios. Recent studies have demonstrated that pre-trained vision-language models like CLIP exhibit strong generalization with just zero or a few normal images.

Bin-Bin Gao, Yue Zhou2026-03-1415 cites

Read