VLM safety steering differs by layer depth, study finds
On Qwen3-VL-4B, early vision-layer steering moved unsafe rejection between 6% and 96% without retraining.
Stories per day · 30 days
3 RSS
Relatedactivation-steering1evaluation-awareness1indic-languages1internvl1language-models1llm-safety1model-evals1multilingual1professional-ethics1qwen1safety-evals1vision-language-models1
15 stories, newest first · one card per story
Stories tagged #refusal
TenetBench tests language models against professional ethics codes
Eleven of fourteen model and reasoning-mode combinations produced conflicting content on most requests, usually without warning.
Evaluation Awareness Beyond English: Do LLMs Act Differently When Tested in Indic Languages?
LLMs refuse more often in Indic languages when told they are under a safety test, including some benign requests.
We’re putting too much faith in AI’s ability to say no
MIT Technology Review argues AI safety leans too heavily on probabilistic, company-defined refusals that can fail or censor.
SLDR: Defending Against Malicious Fine-tuning via Selective Layers Recovery and Dynamic Routing
SLDR restores LLM refusal after malicious fine-tuning by recovering sensitive layers and routing only harmful queries.
Frequency Is Not Sensitivity Identifying Safety-Sensitive Experts in Sparse MoE LLM
Router-gradient expert selection suppresses MoE LLM refusals more effectively than activation frequency, dropping OLMoE refusals 73.5% without breaking outputs.
Safer Content or Firmer Refusals? A Hybrid Perturbation Defense for Alignment under Harmful Fine-tuning
VaccineBooster lowers flagged harmful output after poisoned fine-tuning but can weaken explicit refusals.
Tool Mediation Alters Refusal Mechanisms in Large Language Models
Tool access raises LLM refusal thresholds and makes those refusals easier to weaken than chat refusals.
User Model Extraction via Belief Self-Distillation
Belief Self-Distillation reads and rewrites an LLM's user beliefs, changing refusals with the request fixed.
AEGIS: Audio Endogenous Guarding via Internal Signals Against Large Audio-Language Model Jailbreaks
AEGIS uses mid-layer risk signals to cut audio-model jailbreak unsafe rates from 17.9% to 0.4%.
When Should a Failing Robot Ask? Initiating Corrective Human-Robot Dialogue from Audited Sensor Evidence
Open vision-language models miss robot-failure causes and slash refusal when the refusal option is listed first.
How Fragile Is Safety Alignment at Frontier Scale? A Single-Direction Attack on a 320B MoE
Researchers show directional ablation breaks refusal in GLM-5.3-Flash, a 320B-parameter MoE, cutting refusal by 41–89 points across seven benchmarks.
Safety for Whom? Refusing the Right Subset of a Topic, Not the Whole Topic
Multiverse Computing's Hugging Face post argues language models should refuse only the relevant subset of a topic instead of over-refusing whole subjects.
Safety for Whom? Boundary-Aware Self-Distillation for Controlled LLM Safety Refusal
A self-distillation safety framework tunes narrow-boundary refusals in Qwen3-8B, raising target-domain refusal to 84.75% while cutting over-refusal from 15.20% to 5.20%.
Recognition-Refusal Misalignment in LLMs: Why Models Answer Structurally Unanswerable Questions
A linear hidden-state direction encodes question impossibility in 1.7B-70B LLMs, but misalignment with the safety-refusal pathway explains why models answer unanswerable questions.