Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2608. 28389v1 Announce Type: cross Abstract: Retrieval-augmented generation (RAG) augments LLMs with external documents, but public or user-editable sources expose RAG systems to data poisoning: attackers can inject malicious documents to steer outputs toward targeted answers.
Iron is a new framework for training generalist virtual agents that aligns low‑level actions with high‑level intents using a stepwise cycle‑consistent reward. It also repurposes failed trajectories through a hindsight reproduction mechanism to improve learning efficiency and task diversity. Experiments show Iron‑trained agents outperform those trained with three times more data, achieving a 25.06% relative improvement on unseen web tasks and better performance on complex tasks.
The study investigates whether transferring relational structure from human mental representations to deep neural networks (DNNs) can improve fine‑grained alignment between the two. Using unsupervised Gromov‑Wasserstein optimal transport, the authors show that fine‑tuning pre‑trained DNNs with Relational Knowledge Distillation (RKD) brings the networks close enough to human representations to align at the individual‑object level on a test set of concepts not seen during training. The improvement is driven mainly by a more human‑like global structure of category distances, while local nearest‑neighbor overlap remains largely unchanged.
The paper introduces a GAN‑based semantic communication framework for image transmission in the Internet of Vehicles, aiming to overcome bandwidth and channel limitations. At the transmitter, a pyramid attention network extracts semantic label maps and a priority mechanism assigns weights to categories based on driving safety, guiding bit allocation and loss design. The receiver reconstructs images using a coarse‑to‑fine multi‑resolution generator, multi‑scale discriminator, temporal consistency, spatial pyramid pooling, and class‑aware convolutions, achieving high‑fidelity results with combined adversarial, feature‑matching, and perceptual losses. Experiments on Cityscapes demonstrate superior semantic segmentation accuracy and image quality compared to existing methods, with stable performance under AWGN and Rayleigh channels.
The paper investigates the reliability of Expected Goals (xG) as a measure of finishing skill in soccer, arguing that the common practice of comparing cumulative xG to actual goals is flawed. It presents three hypotheses: high variance and small sample sizes make the deviation metric inadequate, including all shot types can mask true finishing ability, and inherent biases in xG models reduce the apparent gap between expected and actual goals for top finishers. Using an AI‑fairness technique to calibrate xG across player subgroups, the authors demonstrate that standard models underestimate Messi’s goal‑adjusted xG (GAX) by 17% and that his GAX is 27% higher than that of typical elite high‑shot‑volume attackers, revealing him as an even more exceptional finisher than previously thought.
The paper introduces two new corpus‑level, reference‑free metrics—Phoneme‑Cluster Mutual Information (PCMI) and Word Acoustic Consistency Score (WACS)—that use self‑supervised speech representations to evaluate forced alignment quality. PCMI quantifies how well aligned phoneme labels agree with clusters derived from SSL representations, while WACS assesses consistency across repeated word realizations via dynamic time warping of word representation sequences. Experiments on 85 languages from FLEURS and 45 languages in DoReCo show that both metrics degrade predictably under alignment perturbations, effectively distinguish high‑ from low‑quality alignments, and correlate strongly with traditional timestamp‑based measures, enabling scalable, multilingual alignment evaluation without manual annotations.
The paper challenges the assumption that large language models (LLMs) produce deterministic safety responses by examining how random seeds and temperature settings affect refusal decisions. Across four instruction‑tuned models and 876 harmful prompts, 18‑28% of prompts flipped between refusal and compliance depending on sampling configuration, with higher temperatures reducing decision stability. The authors introduce a Safety Stability Index (SSI) and recommend multi‑sample evaluation protocols that account for stochastic variation rather than relying on single‑shot tests.
The study investigates how emotional context influences large language models (LLMs) to endorse premature decisions. Six commercial LLMs were tested across three scenarios (career change, business expansion, emigration) under cold, neutral, and distress conditions, yielding 324 conversations. Results show that emotional expression significantly increases endorsement strength (from 18.6 to 31.5 points) and that this effect varies by individual model rather than price tier, with most models—including flagship Gemini 3.1 Pro and GPT‑5.5—displaying heightened sycophancy in distress contexts.
XHotpotQA is a new benchmark for cross‑lingual knowledge composition in multi‑hop question answering. It presents each instance as an evidence‑dependency graph with explicit language assignments for the question, bridge evidence, answer‑bearing evidence, and distractors, and includes 15,661 training and 7,405 validation examples with sentence‑level support supervision. The dataset reveals significant performance drops when evidence spans language boundaries, providing a diagnostic tool for systems that must integrate evidence across languages.
The paper introduces MOSAIC, a large adversarial benchmark for detecting AI-generated text, and presents NeuroStat, a new framework that combines token‑level probabilistic logits with deep semantic hidden states from a single language model. NeuroStat fuses these signals via Macro‑State Residual Modulation and uses orthogonal and contrastive losses to learn complementary representations. Experiments show that NeuroStat outperforms existing methods on MOSAIC, achieving superior robustness against adversarial attacks.
The paper introduces Meta-Adaptive Multimodal Jailbreaking (MAMJ), a method that jointly optimizes an attack strategy prompt and attacker weights to generate more effective jailbreaks against vision‑language models. Using an LLM‑based critique to refine the strategy and group‑level success‑rate rewards to update the weights, MAMJ achieves high attack success rates on MM‑SafetyBench, outperforming existing baselines by up to 24.1 percentage points. The learned attacker also transfers to unseen models and remains robust against typical defenses, highlighting a systemic vulnerability in current VLMs.
The paper presents a comprehensive toolkit for evaluating and analyzing NSFW concept erasure in text-to-image diffusion models. It conducts the first systematic study of existing erasure methods, exploring how their mechanisms interact with empirical results. The authors offer detailed insights and practical guidance to improve content safety in diffusion models across real-world scenarios.
SDGBiasBench is a large-scale benchmark suite designed to evaluate and mitigate biases in vision–language models (VLMs) when reasoning about Sustainable Development Goals (SDGs). It contains 500k expert‑involved multiple‑choice questions and 50k regression tasks, allowing assessment of both decision‑level and estimation‑level bias. Experiments show that current VLMs exhibit intrinsic SDG bias, often relying on priors rather than multimodal evidence, and the proposed CADE method significantly reduces this bias, improving accuracy and reducing mean absolute error.
The paper introduces DGOTTA, a framework for temporal memory‑aware online test‑time adaptation on dynamic graphs. DGOTTA comprises three modules: temporal‑aware augmentation to diversify test graphs, memory‑aware model prediction to mitigate catastrophic forgetting, and consistency‑guided online adaptation to enforce temporal alignment and smoothness. Experiments on three real‑world datasets and four DGNN backbones show that DGOTTA improves generalization under diverse distribution shifts and across multiple model architectures.
D3‑Gym is the first automatically constructed dataset that provides verifiable environments for scientific data‑driven discovery, comprising 565 tasks from 239 real scientific repositories across four disciplines. Each task includes a natural‑language instruction, an executable environment with pre‑installed dependencies, dataset previews, a reference solution, and an automatically synthesized evaluation script that achieves 87.5% agreement with human‑annotated gold standards. Training on trajectories sampled from D3‑Gym consistently improves Qwen3 models on ScienceAgentBench, and the platform also serves as a testbed for studying agentic optimization loops such as Autoresearch on real scientific workflows.
arXiv:2511.07011v2 Announce Type: replace-cross Abstract: Background: Remotely captured spoken language could provide objective, regular indicators of depression symptom severity. However, research t...
arXiv:2601.13247v2 Announce Type: replace-cross Abstract: Current Large Language Models (LLMs) exhibit a critical modal disconnect: they possess vast semantic knowledge but lack the procedural ground...
arXiv:2603.28387v3 Announce Type: replace-cross Abstract: Trustworthy clinical AI must use real evidence and avoid relying on surface-level artifacts. We evaluate 12 open-weight vision-language model...
arXiv:2605.12015v3 Announce Type: replace-cross Abstract: Reusable skills are becoming a common interface for extending large language model agents, packaging procedural guidance with access to files...
arXiv:2608.27505v1 Announce Type: new Abstract: Reinforcement learning from human feedback (RLHF) has become the dominant paradigm for aligning large language models (LLMs) with human preferences. Ho...