193 articles tagged with #llm-safety. AI-curated summaries with sentiment analysis and key takeaways from 50+ sources.
Researchers introduce DataShield, a novel method for identifying safety-degrading samples in benign datasets used to fine-tune large language models. The approach efficiently detects data points that compromise LLM safety through compliance vector analysis, addressing a critical vulnerability in current model training practices.
Researchers propose Domain-Shift-Aware Conformal Prediction (DS-CP), a framework that improves reliability of large language model outputs by adapting conformal prediction methods to handle domain shift. The approach reweights calibration samples based on proximity to test prompts, delivering more reliable uncertainty quantification and reducing hallucinations in real-world deployments.
InFerActive is an interactive system that improves how AI safety evaluators assess large language models by visualizing sampling results as navigable trees rather than static spreadsheets. The tool uses breadth-first sampling to achieve equivalent harmful-response coverage with up to 5x fewer samples, significantly improving evaluation efficiency according to controlled user studies.
Researchers introduce a Privacy Policy Enforcement framework that detects subtle data leakage in RAG systems beyond standard PII filters, using dual one-class density estimators to identify contextual attribute clusters that collectively identify individuals. The T3+OCSVM detector achieves 93%+ AUROC while reducing false positives by 44-55% and maintaining millisecond latency, outperforming traditional supervised approaches.
Researchers introduce Rationalize, a framework enabling shared semantic reasoning between humans and AI models through complementary role pairs (Explorer-Guide, Investigator-Informant, Teacher-Student, Judge-Advocate). The framework aims to align AI systems not just at the output level but by making purposes, questions, assumptions, and evidence explicit during human-AI collaboration, addressing bidirectional alignment challenges.
Researchers introduce COMPASS, a safety alignment framework for LLM-powered search agents that prevents harmful outcomes from seemingly innocent multi-step queries. The method combines cognitive tree exploration and step-wise alignment to achieve robust safety while maintaining utility, requiring less training data than existing approaches.
Researchers introduce Temporal Logit Observability (TLO), a training-free diagnostic tool that reveals how LLM jailbreak attacks unfold over time by analyzing logit patterns during decoding, rather than just whether attacks succeed. The method identifies that attacks with identical success rates actually follow different failure pathways, enabling better safety evaluation and early-stopping defenses that reduce successful jailbreaks by over 50%.
Researchers evaluated 14 open-source safety guard models across 79,331 samples and found that smaller models like Qwen Guard (4B parameters) significantly outperform larger counterparts in detecting harmful content, achieving 83.97% recall compared to just 25% for some 20B parameter models. The study reveals that model size does not correlate with safety detection performance and that recall—minimizing missed harmful content—is the critical metric for production deployments.
Researchers introduce SCOPE, a lightweight LLM framework designed to monitor pilot readbacks of Air Traffic Control instructions, addressing a critical aviation safety gap where readback anomalies contribute to approximately 80% of aviation incidents. The system achieves 91% accuracy in detecting anomalies and 96.63% correction rates while requiring minimal computational overhead, offering a practical deployment pathway for automated safety monitoring in high-stakes operational environments.
Researchers introduce Opir, a family of efficient encoder-based safety classification models designed to detect toxic content, jailbreaks, and harmful prompts in LLM applications without requiring expensive large guardrail models. The models achieve competitive performance across 12 safety tasks against eight contemporary systems while maintaining significantly smaller deployment footprints, with edge variants containing fewer than 100M parameters.
Researchers introduce LaRA, a framework for detecting data contamination in reinforcement learning post-trained large language models by analyzing layer-wise representations. The method identifies contamination through geometric deviations across neural network layers, outperforming existing detection approaches that rely on output-level signals unreliable for RL-trained models.
Researchers propose Calibrated Entropy Score (CES), a novel method for detecting hallucinations in large language models using entropy distribution patterns from a single forward pass. The technique achieves performance comparable to computationally expensive multi-sample methods while requiring only black-box access to token logits, with formal mathematical guarantees for detection accuracy.
Researchers introduce EVADE-Bench, a multimodal benchmark for evaluating how well AI models detect deliberately obfuscated content in e-commerce, such as products using word splitting or euphemistic language to evade moderation policies. Testing 26 leading LLMs and VLMs reveals significant vulnerabilities in even state-of-the-art models, with findings suggesting that clearer rule design and multi-agent reasoning architectures can substantially improve detection accuracy.
Researchers have developed READER, a compact AI text detector with only 1.5B parameters that outperforms much larger language models and existing detection systems. READER combines classification with explainable reasoning, providing both AI/human verdicts and structured rationales for its decisions, addressing critical limitations in current detection methods that fail under distribution shifts.
Researchers developed a reflective storytelling agent that combines large language models with knowledge graphs and argumentation theory to generate personalized narratives for older adults. Testing with 55 participants showed the system successfully identified personally relevant purposes in two-thirds of narratives, with argument-based grounding and hallucination detection significantly improving perceived consistency and clarity.
FragileFlow introduces a theoretical framework and practical regularizer to detect and mitigate a hidden failure mode in large language models and vision-language models where predictions remain technically correct but confidence margins narrow dangerously. The research provides the first PAC-Bayes bounds for margin-aware error flow, addressing robustness gaps that standard accuracy metrics overlook.
Researchers introduce MELD, an advanced AI-generated text detector that uses multi-task learning to improve robustness against adversarial attacks, transfer across unseen models and domains, and maintain low false-positive rates. The detector outperforms most open-source competitors and matches leading commercial systems on public benchmarks.
Researchers have developed an AI Teaching & Learning Assistant, a Moodle plugin using Retrieval-Augmented Generation (RAG) to provide students with Socratic tutoring while enabling educators to supervise content generation. The system grounds LLM responses in teacher-provided materials to minimize hallucinations and misinformation, achieving high faithfulness scores (0.97) and strong user satisfaction (4.00/5.00 rating).
Researchers introduce a proxy-analyzer framework that detects hallucinations in large language models by analyzing internal activations of a small open-weight reader model rather than the generator itself. The system achieves competitive or superior performance compared to existing methods across multiple model architectures, with notably consistent results showing that model size has minimal impact on detection accuracy.
Researchers propose Multilingual Self-Distillation (MSD), a framework that transfers safety safeguards from high-resource languages like English to vulnerable low-resource languages in large language models. The method eliminates the need for expensive multilingual response data by leveraging an LLM's existing safety capabilities, demonstrating effective cross-lingual protection across diverse jailbreak benchmarks.
Researchers propose a novel black-box confidence estimation method for chain-of-thought reasoning that measures trajectory convergence rather than relying on expensive sampling. Testing across multiple benchmarks and AI models shows significant improvements over self-consistency baselines while requiring only 4 samples instead of 8, with potential applications for safer API-based AI deployment.
Researchers propose WARDEN, an information-theoretic adversarial training framework that improves Large Language Model robustness against prompt attacks by dynamically reweighting adversarial examples using f-divergence principles. The method achieves comparable computational efficiency to existing approaches while substantially reducing attack success rates, advancing the scalability of AI safety mechanisms.
Researchers identify a critical flaw in machine-generated text detection: token-level likelihood signals vary inconsistently across a detector model's hidden space, causing Simpson's paradox that undermines existing detectors. They propose a learned local calibration method that dramatically improves detection performance, with calibrated variants achieving AUROC improvements from 0.63 to 0.85 on GPT-5.4 text.
Researchers propose a framework for comparing language models on safety without labeled benchmark data, introducing SimpleAudit as a validation tool that uses controlled contrasts and variance analysis to establish model safety rankings. The study demonstrates that comparative safety scores are inherently context-dependent, requiring detailed reporting of methods rather than single rankings.
Researchers have identified a critical vulnerability in LLM safety alignment where fine-tuning on benign samples causes parameters to drift toward unsafe behaviors, erasing safety gains from millions of preference examples. The study proposes SQSD, a method to quantify and score individual training samples by their contribution to safety degradation, with demonstrated transferability across different model architectures and scales.
