Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2508.04080v2 Announce Type: replace Abstract: Standard large language model prompting treats geospatial inference as independent, instance-wise prediction, ignoring the fundamental spatial depe...
arXiv:2605.28313v2 Announce Type: replace Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities in tasks related to reasoning and judgment. However, assessing the quality o...
arXiv:2609.00666v1 Announce Type: new Abstract: InfRared Small Target Detection (IRSTD) is a prominent and challenging task in computer vision. In recent years, text-guided methods have significantly...
arXiv:2609.00920v1 Announce Type: cross Abstract: Vision-and-Language Navigation (VLN) requires an agent to navigate through unseen 3D environments according to natural-language instructions. Explici...
arXiv:2507.20881v3 Announce Type: replace Abstract: Endoscopic depth estimation is a critical technology for improving the safety and precision of minimally invasive surgery. It has attracted conside...
GenONet introduces a Spatio-Temporal U-DeepONet architecture that serves as a generator in a GAN framework for high‑resolution precipitation nowcasting up to three hours ahead. By learning continuous‑time precipitation dynamics with a Deep Operator Network and enforcing physics through a moisture‑conservation loss, the model produces sharp, physically consistent forecasts that outperform baselines, especially for high‑intensity events and longer lead times. Ablation studies confirm the added value of the physics‑informed regularizer and the synergy of operator learning with adversarial training.
The paper introduces a new entity alignment foundation model that overcomes the limitations of existing models by addressing the "reasoning horizon gap". It employs a parallel encoding strategy that uses seed entity pairs as local anchors to guide message passing, thereby shortening inference paths and improving alignment across sparse, heterogeneous knowledge graphs. The model also incorporates a merged relation graph and a learnable interaction module, and experimental results demonstrate its strong generalizability to unseen knowledge graphs.
The paper introduces a new privacy vulnerability in diffusion language models (DLMs) called token‑level memorization asymmetry, derived from theoretical analysis of diffusion training dynamics. It proposes Q‑Skew, a quantile‑weighted skewness indicator, to perform membership inference on fine‑tuned DLMs, outperforming existing baselines across multiple datasets and models. Additionally, Q‑Skew can be used to extract personally identifiable information (PII), demonstrating a broader privacy attack surface.
The paper presents a semi‑supervised biomedical image segmentation method that uses a diffusion‑based teacher–student framework. The teacher is pretrained via unsupervised diffusion reconstruction and then co‑trained with a student, leveraging supervised labels and cross pseudo‑supervision on unlabeled data. A multi‑round extension generates multiple stochastic reconstructions to further refine pseudo‑labels, achieving competitive or superior results on several 2D and 3D biomedical datasets, especially when labels are scarce.
The paper introduces Solver-Gradient Guided Reinforcement Learning (SG‑RL), a method that augments standard RL with bounded gradients from a differentiable MPC solver to adapt cost‑function weights online. SG‑RL integrates solver‑gradient guidance into PPO through actor‑update scaling, policy loss, advantage estimation, and value‑function learning, achieving comparable or superior closed‑loop performance while requiring up to 70.6% fewer samples. Experiments on two autonomous racing platforms with intentional model mismatch demonstrate that SG‑RL outperforms both RL and gradient‑based policy learning baselines and generalizes zero‑shot to unseen environments.
The paper introduces a rank‑2 projection subspace to disentangle parametric knowledge (PK) and context knowledge (CK) in natural language explanations (NLEs) generated by large language models. Unlike prior work that used a rank‑1 subspace and focused on single‑step generation, this study performs a multi‑step analysis across longer NLE sequences, revealing richer interactions such as complementary or supportive knowledge. Experiments on four QA datasets and three open‑weight LLMs show that the rank‑2 subspace better captures diverse PK–CK dynamics, with hallucinated generations aligning strongly with PK and context‑faithful generations balancing PK and CK alignment.
The paper "Beyond Static Summarization: Proactive Memory Extraction for LLM Agents" identifies two shortcomings in current memory extraction for large language model agents: (1) extraction occurs ahead of time and mixes multiple types of information, leading to loss of useful details, and (2) extraction is typically one‑off, allowing errors and hallucinations to persist. To address these issues, the authors propose ProMem, a proactive framework that separates details, events, and relations, applies distinct extraction strategies for each, checks for completeness, and verifies facts at an atomic level. Experiments demonstrate that ProMem enhances memory completeness and question‑answering accuracy while maintaining a favorable balance between quality and token cost.
The paper introduces the concept of multi-view hallucination (MVH), where large vision-language models produce incorrect answers when processing images from multiple viewpoints. It presents MVH-Bench, a benchmark of 4.8k question-answer pairs that target cross-instance and cross-view hallucinations, and shows that MVH is common across recent models. The authors propose Reference Shift Contrastive Decoding (RSCD), a training-free decoding method that mitigates visual interference, achieving significant performance gains on MVH-Bench with LLaVA-OneVision and Qwen2.5-VL.
The paper presents the first systematic reliability evaluation of diffusion-based Large Vision‑Language Models (dLVLMs), comparing six diffusion models to autoregressive (AR) baselines across four dimensions. Key findings include a reversal of the yes‑bias seen in AR models for binary visual queries, competitive hallucination rates but lower linguistic quality, near‑zero accuracy for underrepresented racial groups with opposite‑polarity gender bias, and accuracy collapse in multiple‑choice tasks when the correct option is shorter than distractors due to a length prior emerging at the first denoising step. Additionally, tokens committed late in denoising with low confidence correlate with hallucinated content, indicating a unique mechanistic signal in diffusion generation.
This study introduces a multipart 3D‑2D rigid registration method that aligns CAD models of gas turbine blades with X‑ray projections, offering an alternative to traditional CT reconstruction for inspecting complex internal structures. The greedy registration algorithm first aligns the blade’s exterior, then its internal components by maximizing mutual information, while a separate greedy approach optimizes view angles to minimize dimensional measurement errors. The method achieves sub‑pixel registration accuracy, with errors below one‑fifth of the magnified detector‑pixel pitch, and remains robust to image noise and defects when appropriate view selection is applied.
The paper introduces a new adversarial attack on multimodal large language models (MLLMs) that targets the models’ intrinsic focus in the frequency domain. By exploiting the models’ sensitivity to phase information, the authors design a phase‑aware perturbation strategy that restricts changes to structure‑relevant phase regions, making attacks both effective and imperceptible. An auxiliary adversarial prompt module further misaligns multimodal attention toward targeted structural patterns, and experiments on several MLLM models and datasets confirm the method’s superior performance over existing attacks.
The paper introduces the Detection Surface, a geometric framework that maps the decision boundaries of heterogeneous safety filters in text‑to‑image models. Using this insight, the authors propose CRACK, a multi‑agent debate system that iteratively mutates prompts, diagnoses layer‑specific constraints, and refines attacks to bypass composite defenses. Experiments demonstrate that CRACK can achieve attack success rates up to 99.63% while using fewer queries and preserving semantic fidelity.
SEBA is a sample‑efficient framework for black‑box adversarial attacks on visual reinforcement learning agents. It combines a shadow Q model, a generative adversarial network for imperceptible perturbations, and a world model to simulate dynamics, reducing real‑world queries. Experiments on MuJoCo and Atari show SEBA significantly lowers cumulative rewards while preserving visual fidelity and requiring far fewer environment interactions than previous methods.
The paper introduces a new online fair division framework where a learner must allocate indivisible items to agents in real time, balancing fairness and efficiency. Traditional methods rely on many copies of each item to estimate utilities, but this is unrealistic for platforms with many users and few interactions. By treating utility as an unknown function of item-agent features and framing the problem as a contextual bandit, the authors propose algorithms that achieve sublinear regret and demonstrate their effectiveness experimentally.
The paper investigates privacy risks in Vision Transformer (ViT) split‑inference systems that use token reduction and token shuffling to lower computation and communication costs. It shows that even after token shuffling, transmitted token embeddings still contain enough positional information for a new attack, the Spatially Aligned Reconstruction Attack (SARA), which predicts token positions, restores spatial layout, fills missing embeddings with a masked autoencoder, and reconstructs the input image. While token reduction offers stronger protection, significant leakage remains when retained tokens preserve semantic and positional cues, and the authors propose a lightweight edge‑side defense that removes positional embeddings and adapts transformer blocks via knowledge distillation to reduce SARA’s effectiveness without harming downstream accuracy.