Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2608.22346v1 Announce Type: cross Abstract: This data article describes a multimodal collection of scanned examination answers paired with expert-designed Outcome-Based Education (OBE) grading...
arXiv:2608.21415v1 Announce Type: cross Abstract: Large Vision-Language Models (LVLMs) have achieved remarkable performance across a wide range of tasks; however, they often inherit social biases fro...
arXiv:2608.21606v1 Announce Type: new Abstract: Machine unlearning aims to remove the influence of targeted training data from a model while preserving its remaining capabilities, but evaluating whet...
arXiv:2608.21880v1 Announce Type: new Abstract: Bangla large language model (LLM) safety is difficult to evaluate with English-centric or standard-script benchmarks because Bangla users routinely wri...
AI University (AI‑U) is a flexible framework that uses a fine‑tuned large language model (LLM) combined with retrieval‑augmented generation (RAG) and a reasoning synthesis model to produce style‑aligned responses from lecture videos, notes, and textbooks. In a graduate‑level finite‑element‑method (FEM) course, the authors created a pipeline to generate course‑grounded training data, fine‑tuned an open‑source LLM with Low‑Rank Adaptation (LoRA), and applied RAG‑based synthesis. Evaluation through cosine similarity, LLM‑based assessment, expert review, and user studies showed that the expert model outperformed the base model in alignment with course materials, with 86 % of test cases scoring higher and human users preferring the expert model roughly twice as often. whyItMatters":"The study demonstrates a practical method for building course‑specific learning assistants that improve alignment with instructional content, offering a template that can be extended across STEM fields."
The paper proposes an information‑flow perspective on explainability, arguing that exposing reasons for observed effects is a positive flow of information that must be specified and verified. It introduces an epistemic temporal logic with counterfactual causes to formalize the requirement that agents gain knowledge about why an effect occurred, and presents an algorithm for checking finite‑state models against these specifications. A prototype implementation is evaluated on benchmarks, demonstrating the ability to distinguish explainable from unexplainable systems and to incorporate privacy constraints.
The paper introduces a new type of adversarial attack on automated fact‑checking systems that uses large language models to rephrase claims with persuasive techniques. By applying 15 persuasion methods across five categories, the authors evaluate how these rewrites affect claim verification and evidence retrieval on the FEVER and FEVEROUS benchmarks. Results show that persuasive rewrites significantly degrade both verification accuracy and evidence retrieval performance, underscoring the vulnerability of current fact‑checking systems to such attacks.
The paper investigates how Spoken Language Models (SLMs) process speech compared to text, noting that current SLMs show weak alignment between speech and text representations despite strong downstream performance. The authors propose a framework that separates length mismatch from semantic alignment to better match speech and text representations. Experiments on multiple benchmarks demonstrate that this approach yields competitive results against strong baselines, highlighting the need to explicitly address structural differences between speech and text in SLM training.
The study examines how multilingual large language models (LLMs) produce outputs that differ across sociocultural contexts, highlighting that identity labels and source-language cues can mislead assessments of cultural grounding. Using a human‑validated, multi‑agent audit on 89,253 outputs from 12 LLMs in English, French, and Chinese across 18 occupations and three task conditions, the authors find that bias representation varies systematically by language and task. Removing direct identity cues reduces identity‑label prediction in English and Chinese but not in French, and the source language’s cultural context consistently receives the highest relevance score, though this signal weakens after translation or name masking. "whyItMatters":"The findings show that surface cues can obscure true cross‑cultural patterns, underscoring the need for careful audit designs to avoid misleading conclusions about bias in multilingual LLMs."
SketchFlow is a new generative framework for creating high‑quality vector sketches from text prompts. It uses a Gaussian Mixture Model prior in the CLIP latent space and an Optimal Transport Conditional Flow Matching model to map this prior to sketch features, which are then decoded by a Hybrid Diffusion Decoder combining 1D U‑Net and Transformer architectures. The approach achieves superior visual quality and human‑like drawing styles, and supports zero‑shot synthesis for unseen concepts and smooth semantic interpolation.
arXiv:2608.22367v1 Announce Type: new Abstract: Diffusion multimodal large language models (dMLLMs) frequently produce long-form outputs marred by semantic drift and repetition, with quality generall...
arXiv:2608. 22915v1 Announce Type: cross Abstract: Inference-time pipelines often sample multiple outputs, filter them with a learned safety model, and return the proxy-feasible output with the highest learned reward.
arXiv:2606.24844v3 Announce Type: replace Abstract: One-step diffusion editors are fast because they avoid inversion and iterative optimization, but a single transport update must be aggressive enoug...
arXiv:2608.23034v1 Announce Type: cross Abstract: Controlling restricted knowledge in large language models is essential for model alignment and safe deployment. Test-time unlearning avoids costly re...
arXiv:2608.22176v1 Announce Type: new Abstract: Large Language Models (LLMs) are increasingly applied to clinical prediction tasks such as in-hospital mortality and readmission from electronic health...
arXiv:2608.22444v1 Announce Type: new Abstract: The unit of AI safety evaluation is still the individual model, yet language-model agents are increasingly deployed in interacting populations that rea...
arXiv:2608.22048v1 Announce Type: new Abstract: Large language models are increasingly deployed on local hardware for privacy, cost, and accessibility reasons. Yet many evaluations emphasize accuracy...
The paper studies how midtraining with synthetic documents can align AI models to the value of animal compassion. It introduces ANIMA, a 26‑question benchmark covering 13 ethical dimensions, and shows that training on 3,000 documents yields 77% accuracy versus 40% for instruction‑tuning. The improvement disappears after 5,000 samples, indicating that document‑based interventions may need explicit preservation strategies.
The paper introduces RASET, a router‑agnostic safety‑critical expert tuning framework for Mixture‑of‑Experts (MoE) large language models. RASET identifies a small subset of experts that are responsible for safety enforcement and applies parameter‑efficient tuning only to those experts, preserving the model’s intrinsic routing behavior. Experiments on five open‑weight MoE backbones show that RASET achieves a high safety‑bypass yield, outperforming existing baselines by a significant margin.
GRACE is a step‑level benchmark for evaluating the faithfulness of chain‑of‑thought reasoning over context. It provides human annotations for each step in CoT traces from 10 models across 4 datasets, labeling faithfulness, error category, and natural‑language explanations. The benchmark introduces a data‑driven taxonomy that splits errors into GRACE‑Inference (deductive) and GRACE‑Grounding (factual) tracks, each with four categories, and demonstrates that incorporating step‑level faithfulness signals can improve downstream accuracy and reasoning reliability.