Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2511.10661v2 Announce Type: replace Abstract: It is increasingly important to evaluate the characteristics of systems based on large language models (LLMs). Evaluations in this context often re...
arXiv:2604.02543v2 Announce Type: replace Abstract: As vision-language models (VLMs) are increasingly deployed in clinical decision support, more than accuracy is required: knowing when to trust thei...
arXiv:2605.25765v2 Announce Type: replace-cross Abstract: Existing closed-form methods for concept unlearning in text-to-image diffusion models typically derive editing directions from fixed text emb...
The study introduces an automated system for estimating the Verbal Fluency Index (VFI) in individuals with Motor Neuron Disease (MND) by combining ASR (WhisperX) and VAD (Silero) with precise timestamping. Using a unique MND dataset, the approach outperformed traditional acoustic and self‑supervised embedding methods, achieving high predictive accuracy (R² up to 0.9 for P‑words and 0.8 for S‑words). Clinically inspired features were consistently superior, demonstrating the feasibility of automated VFI estimation for monitoring cognitive impairment in MND.
arXiv:2609.38260v1 Announce Type: cross Abstract: Values such as honesty, autonomy, and confidentiality are often regarded as general principles underpinning AI alignment. However, what it means to a...
The paper investigates how conversational safety in language models degrades over extended, adversarial interactions. By testing three instruction‑tuned models with persistent adversarial users across up to 101 turns, the study finds that safe‑response rates drop sharply from 85–100% at the first turn to 15–44% by the end. This demonstrates that strong single‑turn safety does not guarantee continued safety in long conversations.
arXiv:2609.39027v1 Announce Type: new Abstract: AI reviewers can assign different judgments to manuscripts that report the same science in different wording, potentially rewarding rhetorical optimiza...
The paper investigates on‑policy self‑distillation (OPSD) as a method to enhance reasoning in language models, focusing on mathematical reasoning across models from 0.6B to 8B parameters. Through controlled experiments and token‑level analysis, the authors find that OPSD’s effectiveness depends on alignment between the teacher’s reasoning mode and the full teacher prefix, rather than on privileged semantics alone. They observe that OPSD only improves reasoning in limited compatibility regimes, while often causing length growth, degradation, or behavioral collapse, and that the teacher’s signal is unstable and not predictive of downstream performance.
arXiv:2609.39846v1 Announce Type: cross Abstract: We investigate the problem of role-capability leakage (RCL), in which a role-prompted reasoning model generates convincing in-role text while continu...
arXiv:2609.39853v1 Announce Type: new Abstract: Role-playing prompting has become a popular yet simple technique for improving LLM reasoning and output quality. However, whether it consistently boost...
arXiv:2609.40124v1 Announce Type: new Abstract: Bias has long been studied in social psychology and cognitive science, where decades of research have produced a body of validated interventions that r...
The paper introduces a Voice Retention / Representation Ratio metric to assess bias in large language model (LLM) summaries of employee feedback. Using a bilingual corpus of 2,586 responses from a global professional services firm, the study finds that criticism is reported more reliably than praise, and that LLM summaries tend to filter by popularity rather than sentiment—criticism often survives while single-mention concerns, short or German-only content are frequently omitted. The authors argue that prevalence, not sentiment, drives the bias, and provide a metric, field evidence, and a disaggregated voice‑retention card for future audits.
arXiv:2609.39050v1 Announce Type: cross Abstract: As multi-agent systems enter high-stakes domains, the possibility that agents may circumvent safety boundaries is a growing concern. Prior work has e...
arXiv:2609.39549v1 Announce Type: cross Abstract: As Large Language Model (LLM) agents are increasingly deployed in complex environments, multi-turn interaction attacks have become a significant secu...
arXiv:2609.39688v1 Announce Type: cross Abstract: Multimodal encoders such as CLIP underlie many downstream systems, but their web-scale training data embed harmful associations that safety alignment...
arXiv:2609.39920v1 Announce Type: cross Abstract: Large vision-language models (LVLMs) exhibit strong multimodal in-context learning (ICL) capabilities, yet this ability degrades substantially as mod...
arXiv:2601.09717v2 Announce Type: replace-cross Abstract: Online medical consultations contain sensitive health information whose privacy implications depend not only on the entities mentioned but al...
arXiv:2605.28306v2 Announce Type: replace-cross Abstract: Mixture-of-Experts (MoE) models enable efficient LLM scaling, yet adapting them to non-English downstream tasks remains challenging. Standard...
The paper evaluates safety monitors by measuring recall only on prompts that the target model actually answers, rather than on all harmful prompts. Across several guard systems, recall at a 1% false‑positive rate drops sharply when focusing on answered prompts, with monitors catching refused requests 1.1–6.4 times more often than answered ones. Rewriting prompts to be less explicit dramatically increases compliance and reveals that many harmful requests slip past monitors, especially when phrasing is softened. Fine‑tuning guards on these rewritten prompts improves recall from 0.24 to 0.89 on answered requests and generalizes to unseen benchmarks.
arXiv:2609.34284v2 Announce Type: replace Abstract: Personalized LLMs must decide, for each stored preference, whether the current context calls for applying or suppressing it, which we call its appl...