arXiv:2608. 08451v1 Announce Type: cross Abstract: Harmful chat dialogues are ever-shifting through type-shifting and lexical evasion, yet we find they share invariant principles, i.
arXiv:2604. 17301v2 Announce Type: replace-cross Abstract: Detecting harmful content in multi turn dialogue requires reasoning over the full conversational context rather than isolated utterances.
arXiv:2604. 09544v2 Announce Type: replace-cross Abstract: Large language models (LLMs) undergo alignment training to avoid harmful behaviors, yet the resulting safeguards remain brittle: jailbreaks routinely bypass them, and fine-tuning on narrow domains can induce ``emergent misalignment'' that generalizes broadly.
arXiv:2606. 02423v1 Announce Type: cross Abstract: Large language models (LLMs) can serve as helpful assistants, yet they can equally function as harm amplifiers that enable malicious users to achieve harmful outcomes beyond their capabilities through extended interactions.
arXiv:2601. 14340v3 Announce Type: replace-cross Abstract: Large Language Models (LLMs) are increasingly deployed as multi-turn assistants and customized through instruction tuning with project-specific training components.
arXiv:2608. 16622v1 Announce Type: cross Abstract: Multimodal harmful meme detection is typically formulated as image--text harmfulness classification.
arXiv:2606. 18852v1 Announce Type: cross Abstract: Classifying implicit hate speech remains a challenge, as intent is often masked through insinuation and context rather than explicit slurs.
Large language models increasingly stream long, reasoning-intensive responses in real time, making when to moderate as critical as whether to moderate. Existing guardrails fall into two unsatisfactory extremes: response-level methods delay intervention until the full output is generated, whereas token-level methods act on incomplete semantics, often producing unstable decisions and excessive guard invocations.
arXiv:2610.16312v1 Announce Type: new Abstract: Explicit reasoning in large language models (LLMs) is often associated with reliability, yet it can still pose risks when a model infers harmful goals...
The paper investigates how activation steering—adding a vector to a large language model’s residual-stream activations—alters discourse-level aspects of generated text, such as formality, evidence usage, and ambiguity. It demonstrates that these unintended changes can be largely mitigated by applying a second, complementary steering vector that targets the affected discourse dimension, restoring about 78% of the original discourse characteristics while maintaining the intended behavior in roughly 93% of outputs.
The paper introduces CAM-Steer, a Category‑Adaptive Multi‑category Safety Steering framework that estimates risk for each harm category by comparing hidden states to safe and unsafe prototypes. It then combines safety directions into a single steering vector and applies a rotation whose angle is set by the estimated risks, preserving the hidden‑state norm. Experiments on three LLM backbones and seven harm categories show that CAM‑Steer outperforms baselines in defense success rate, even when multiple harm categories co‑occur, with negligible inference overhead.
The study investigates how category‑specific harmfulness signals, isolated by removing the shared general harmfulness component, influence large language model (LLM) safety. Using activation steering across 11 risk categories in three instruction‑tuned LLMs, the authors find that the presence of harmfulness in these category residuals varies by category and that the pattern of inducing refusal is even more model‑dependent. Additionally, category residuals were shown to enhance the models’ downstream alignment with the shared general harmfulness representation, indicating that fine‑grained signals play a role beyond the general component.
The paper introduces a counterfactually anchored evidence attribution approach for multi‑turn large language model safety failures. It presents a new dataset of 1,762 conversations, including adversarial, benign twins, and high‑risk vocabulary variants, and trains a lightweight hierarchical model that accurately predicts safety violations and attributes them to specific user turns and token spans. The model achieves high detection performance (F1 = 0.988) and significantly reduces adversarial confidence when top‑attributed tokens are removed, while maintaining low false‑positive rates on benign conversations.