Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2611.06247v1 Announce Type: new Abstract: Cross-modal remote sensing image retrieval (RS-CMIR) requires representations that are simultaneously modality-invariant and semantically discriminativ...
arXiv:2611.06302v1 Announce Type: new Abstract: Fine-grained action understanding requires vision-language models to ground what an agent does, how it acts, and why it acts in spatio-temporal evidenc...
arXiv:2611.06433v1 Announce Type: new Abstract: Multimodal large language models (MLLMs) remain prone to visual hallucinations when correct responses depend on subtle visual details. One line of post...
arXiv:2611.06635v1 Announce Type: new Abstract: Unmanned aerial vehicle (UAV)-based collaborative perception enables multiple agents to share complementary perceptual information, overcoming limitati...
arXiv:2611.06796v1 Announce Type: new Abstract: Few-step distillation offers efficient generation with diffusion and flow models, but preserving quality with only a few sampling steps remains challen...
arXiv:2611.06890v1 Announce Type: new Abstract: For many computer vision tasks, neural networks can be trained to produce a low average error, yet they still fail on seemingly simple scenes. Reducing...
arXiv:2606.18439v2 Announce Type: replace Abstract: Visual Geometry Grounded Transformer (VGGT) recovers dense 3D scene structure from multi-view images in one forward pass, but quadratic cross-frame...
arXiv:2608.12088v2 Announce Type: replace Abstract: Generative models can produce images nearly indistinguishable from real data, yet rigorous and interpretable evaluation remains challenging. Conven...
arXiv:2611.03114v2 Announce Type: replace Abstract: 3D Gaussian Splatting (3DGS) has become a vital tool in novel view synthesis. 3D Gaussian functions with attributes like color, opacity, and scale...
EmoReward is a self‑evolving reward calibration framework for emotional video generation using reinforcement learning. It distills reliability from reward traces during training to adaptively weight multiple reward primitives—such as visual quality, motion, and text alignment—based on emotional evidence. The method employs grounding cues, heterogeneous signal primitives, reflective reward gating, and a RewardMemory module to calibrate fusion weights across training runs, demonstrating improvements on HunyuanVideo 1.5 and transferability to SD 3.5 and FLUX.
The paper investigates end‑to‑end FP4 reinforcement learning for Mixture‑of‑Experts large language models and finds that existing recovery techniques for dense models fail to improve accuracy in the MoE setting. The authors attribute this to the discrete routing decisions in MoE, which make training‑inference policy alignment highly sensitive to asymmetry in recovery methods. They introduce MoEResQ, a symmetric quantization‑recovery recipe that restores alignment and significantly narrows the accuracy gap to full precision.
The study empirically investigates how length‑normalized Direct Preference Optimization (DPO) with a Structured Fine‑Tuning (SFT) retention term (RPO objective) affects downstream capabilities of a Qwen3.5‑4B model fine‑tuned with sparse LoRA on structured financial‑filing tasks. Results show that the alignment tax is graded and task‑selective: extraction performance is largely preserved, JSON extraction degrades modestly, and router‑based tasks suffer severe loss, with most degradation attributed to suppression of rejected completions. The analysis uses 8 seeds, artifact‑aware scoring, and TOST equivalence testing to quantify these effects across different reward margins and precision settings.
CARTBench is a new benchmark for evaluating large language model shopping assistants on complex, ambiguous requests that involve timelines. It compiles 1,603 timestamped dialogues from 17 retail domains and pairs them with product records, then uses case‑specific rubrics based on Ideal Response Principles to score responses across Core, Excellence, and Safety criteria. In a full‑suite evaluation of ten commercial APIs and several open‑weight models, only three models scored above 90% on Core and none surpassed 45% on Excellence, revealing a gap between required behavior and high‑quality purchase advice.
LITMUS is a benchmark for testing the safety of large language model agents in real operating system environments, focusing on behavior jailbreaks that can cause irreversible OS-level harm. It includes 819 high-risk test cases organized into a harmful seed subset and six attack-extended subsets covering jailbreak speaking, skill injection, and entity wrapping, and employs a semantic-physical dual verification mechanism with OS-level state rollback. Evaluation of frontier agents shows that even strong models can execute dangerous operations, exhibit execution hallucination, and are vulnerable to skill injection and entity wrapping attacks.
The paper investigates how different stakeholders—end‑users, AI builders, and policymakers—define and experience transparency in generative AI. It finds that while all groups agree on core goals of accuracy and privacy, end‑users perceive larger gaps between desired and available transparency features, with only 28% feeling well served compared to 53% of builders and 49% of policymakers. The study also shows that transparency is context‑dependent, shifting priorities based on task and role, and calls for pluralistic transparency infrastructures rather than a one‑size‑fits‑all approach.
MA‑SafeDiffuser is a multi‑agent extension of SafeDiffuser that incorporates diffusion‑based trajectory planners with finite‑time diffusion invariance guarantees for joint‑agent safety specifications. It formalizes joint safe sets as intersections of per‑agent and pairwise barrier sets, and derives both centralized and decentralized constrained denoising procedures that maintain invariance under mild assumptions. The approach also tackles local‑trap and deadlock issues using time‑varying specifications and liveness CBFs, and introduces a lightweight benchmarking suite including a multi‑agent Maze2D domain, showing reduced violation counts compared to unconstrained diffusion baselines while preserving planning quality.
The paper investigates confirmation bias in Agent-as-a-Judge systems, which evaluate open-ended agent tasks by gathering evidence from the agent’s trajectory. It demonstrates that a single false pre‑screening note can overturn up to 40% of correct verdicts across multiple judges and datasets, even when the judge collects decisive evidence. The authors classify the resulting errors into over‑trust, hallucination, and under‑investigation, noting that over‑trust dominates in larger Claude judges, and show that lightweight mitigations reduce but do not eliminate the bias.
The paper investigates how vision‑language models (VLMs) perceive uncertainty in visualizations and make decisions based on visual data. Two studies were conducted: one on agents’ ability to quantify uncertainty in charts, and another on their decision‑making when presented with images. Results show GPT‑4o best perceives uncertainty in pie and bar charts, while agents align most closely with rational decision theory when uncertainty is shown via circle charts, icon arrays, or text.
The paper presents a self‑supervised pretraining approach for facial expression recognition (FER) using optomyographic (OMG) smart glasses like OCOsense. By pretraining a Vision Transformer with a masked‑autoencoder on unlabelled recordings and fine‑tuning for a five‑class FER task, the authors demonstrate that the pretrained model outperforms a from‑scratch model even when only 5% of the labels are used. The study shows that SSL can substantially reduce labeling costs for smart eyewear FER and suggests benefits for future personalization and continual adaptation.
UniCompress is a deployment-oriented framework that transforms multi-task vision‑language expertise into a single compact, adapter‑free model suitable for edge intelligence. Starting from a BF16 Qwen3‑VL‑4B teacher, it builds an asymmetric INT2 student using GPTQ initialization and teacher‑guided quantization recovery, jointly compressing both the vision encoder and language backbone. Evaluated on four embodied tasks, UniCompress cuts weight storage from 8.9 GB to 2.1 GB (≈76% reduction) while improving the INT2 macro task score by 34.3 percentage points over GPTQ‑INT2 and staying within 0.9 points of the BF16 teacher.