Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2609.16192v1 Announce Type: cross Abstract: The current trend of digitalisation has revolutionised the organisation of work and the way it is measured and performed across the globe, with AI be...
arXiv:2609.16255v1 Announce Type: cross Abstract: We present an efficient method to distill reasoning capabilities into compact video-language models (VLMs) for video question answering (VideoQA). Ou...
arXiv:2408.15538v4 Announce Type: replace Abstract: While modern Autonomous Vehicle (AV) systems can develop reliable driving policies under regular traffic conditions, they frequently struggle with...
arXiv:2602.06652v2 Announce Type: replace Abstract: The robustness of Vision Language Models (VLMs) is commonly assessed through output-level invariance, implicitly assuming that stable predictions r...
arXiv:2512.01782v4 Announce Type: replace-cross Abstract: Randomized Smoothing (RS) is a prominent technique for certifying the robustness of neural networks against adversarial perturbations. With R...
arXiv:2603.21676v2 Announce Type: replace-cross Abstract: Standard Transformers have a fixed computational depth, limiting their ability to generalize to tasks that require variable-depth reasoning....
arXiv:2609.15997v1 Announce Type: cross Abstract: Improving safety at intersections requires identifying crash mechanisms and recommending appropriate countermeasures. However, this process tradition...
Meta-LinEXP3 is an online-within-online algorithm designed for adversarial linear contextual bandits with random action sets. It builds a task-level prior from completed tasks to guide an inner LinEXP3 learner, achieving an σO(√n) per‑task regret when context distributions are known and an σO(n^{2/3}) regret with a past‑only regularized moment estimator when they are unknown. The paper also links prior accuracy to transfer regret, showing that better priors yield sublinear, transfer‑dependent regret across tasks, and demonstrates the method on structured hyperspectral tensor sampling.
arXiv:2609.16689v1 Announce Type: new Abstract: Large-scale vision-language models (VLM) such as CLIP enable strong open-vocabulary reasoning, yet deploying these capabilities on resource-constrained...
arXiv:2609.17269v1 Announce Type: new Abstract: Multimodal large language models generate natural-language responses from visual inputs, yet may mention objects absent from an image. In medication as...
The paper introduces Differential Privacy Representation Geometry for Medical Imaging (DP‑RGMI), a framework that interprets differential privacy as a structured transformation of representation space. DP‑RGMI decomposes performance loss into encoder geometry—measured by representation displacement and spectral effective dimension—and task‑head utilization, quantified by the gap between linear‑probe and end‑to‑end utility. Across 594,000 chest X‑ray images from four datasets, the study finds that differential privacy consistently creates a utilization gap even when linear separability remains, while displacement and spectral dimension vary non‑monotonically with initialization and dataset, indicating that privacy alters representation anisotropy rather than uniformly collapsing features.
The paper evaluates how robust three text‑to‑audio models—MusicGen‑small, MusicGen‑large, and Stable Audio 2.5—are to small changes in prompts that could affect adaptive game soundtracks. Using metrics such as log‑Mel distance, MFCC/chroma‑DTW, and CLAP similarity, the study finds that Stable Audio 2.5 consistently yields the lowest acoustic distances and highest CLAP similarity when prompts are structurally rephrased, while MusicGen‑large performs best under lexical substitutions and intensity shifts. The authors also observe that Stable Audio 2.5 shows the greatest variation in prompt‑to‑audio alignment across different random seeds, highlighting the need for multi‑seed robustness testing in game audio applications.
The paper investigates how large language models (LLMs) generate distractor answers for multiple‑choice questions (MCQs) by modeling student misconceptions. It introduces a learning‑science‑based taxonomy of reasoning strategies and applies it to LLM‑generated reasoning traces in math and science MCQs. The study finds that in math, LLMs often follow a misconception‑based process that can be diagnostically useful, whereas in science they rely more on semantic similarity, with frequent failures when the model cannot produce a correct solution or discards plausible distractors. Providing the correct solution in the prompt improves alignment with human distractors by 6.4%. "whyItMatters":"The findings show that anchoring distractor generation to the correct solution enhances LLM alignment with human‑authored distractors, underscoring the importance of correct‑answer cues in educational AI."
The paper introduces ViD, a vision‑dominant gender bias mitigation framework for large vision‑language models. ViD uses causal analysis of attention patterns and dual mechanisms—backdoor adjustment and refined token selection—to suppress bias while preserving reasoning and generation quality. Experiments show a 14.7% reduction in gender bias on FACET and significant improvements on MS COCO image captioning, all without extra training overhead.
The paper introduces Affect-Prototype Guided Fusion (APCF), a framework for open‑vocabulary multimodal emotion recognition that handles incomplete and unsynchronized modal data. APCF builds an affect‑prototype library to model how different emotions contribute across modalities, enabling dynamic fusion of available features. The fused representations are then decoded by an LLM to generate open‑vocabulary emotion labels, achieving superior performance on OV‑MERD+ and MER‑FG datasets compared to existing methods.
arXiv:2609.15998v1 Announce Type: new Abstract: Every large language model (LLM) has behavioral traits and moral preferences that comprise its character. Whether by design or as an emergent property...
arXiv:2609.16739v1 Announce Type: new Abstract: Background: Large language models (LLMs) have achieved physician-comparable performance on multiple-choice medical knowledge examinations, but their ca...
arXiv:2609.16006v1 Announce Type: cross Abstract: Large language models (LLMs) increasingly serve users whose expectations are shaped by their cultural context, yet most cultural evaluations test wha...
arXiv:2601.09050v2 Announce Type: replace Abstract: Tonal low-resource languages are widely spoken but remain underserved by modern speech technologies. A central challenge is learning speech represe...
The paper examines how the tight integration of multimodal understanding and image generation in Unified Multimodal Models (UMMs) can create new safety vulnerabilities. It introduces RICE, an attack framework that exploits bidirectional interactions between generation and understanding to propagate unsafe signals across modalities. Experiments demonstrate high attack success rates in both Generation‑to‑Understanding and Understanding‑to‑Generation pathways, revealing previously overlooked safety weaknesses in UMMs.