Live intelligence

Latest AI updates

6 updates

Research Editor selected

SafeStart: Mitigating the Overlooked Safety Cost of Active Expert Reduction in MoE LLM Serving

This paper reveals that expert-reduction methods for MoE LLM serving, which speed up inference by running fewer experts per token, can weaken refusal of harmful requests. The authors introduce SafeStart, which runs all experts for the prompt and first eight generated tokens, then reduces experts afterward, preserving most speedup while restoring safety. In vLLM on an RTX A6000, SafeStart keeps at least 85% of the offline decode speedup at batch sizes 8 and 32 and reduces harmful compliance increases to near baseline levels across three models.

Why it matters: Provides a concrete, recent finding that expert reduction in MoE serving has safety costs and offers a practical mitigation, valuable for AI serving systems and safety research.

arxiv1d agoScore 85

Mixture-of-ExpertsLLM safetyInference optimization

Research Editor selected

Addition by Subtraction: Mitigating Fragile Safety Alignment in LLMs

This paper introduces Addition by Subtraction, a finetuning method that removes safety-critical neurons and uses preference optimization to redistribute safety behavior across the model. The resulting models are significantly more robust to pruning-based attacks, refusal-direction ablation, and various black-box jailbreaks, with attack success rates reduced by up to 90% across four models. The approach also improves truthfulness while largely preserving general capabilities.

Why it matters: Provides a concrete, novel method to strengthen LLM safety alignment against sparse-component vulnerabilities, with broad empirical validation.

arxiv1d agoScore 85

LLM safetyadversarial robustnessmodel pruning

Research Editor selected

False Reassurance in LLM Severe-Weather Guidance: A Replayable Evaluation Against Real NWS Alerts

This paper introduces 'false reassurance'—an LLM assistant telling users conditions are calm while a severe-weather alert is active—and measures it by replaying 117 frozen cases built from real NWS alerts through a Claude-based pipeline. The production pipeline achieved 0% false reassurance, while blind prompts caused 37-45% false reassurance, and real alerts significantly reduced it. The study also highlights failures like ingesting NWS test products as real alerts and token budget issues, and releases dataset, labels, and code.

Why it matters: Provides concrete, reproducible evaluation of a critical safety failure mode in LLM weather assistants, with actionable insights for system design.

arxiv2d agoScore 82

LLM safetyweather guidancealert integration

Research Editor selected

Predictive Safety Guardrails Do Not Always Make LLM Agents Safer

A systematic study of world-model-driven predictive safety guardrails for LLM agents finds that stronger prediction involvement does not guarantee safety gains and can reduce task completion while increasing inference cost. Predictions become unreliable over longer horizons, and safety gains can persist even under corrupted predictions, revealing a fundamental prediction-to-action gap. The findings challenge the assumption that predictive foresight alone reliably improves agent safety.

Why it matters: Provides critical empirical evidence that current predictive guardrails may not work as intended, informing future safety research.

arxiv3d agoScore 92

LLM safetypredictive guardrailsworld models

Research Editor selected

When Safety Checks Become Privacy Risks: A Privacy-Preserving Paradigm for LLM Agent Safety Auditing

The paper proposes a privacy-preserving paradigm for safety auditing of LLM agents, separating audit computation from plaintext access and execution authority. It introduces SealCall, a framework combining learned audit representations with CKKS-based encrypted similarity computation, achieving a mean F1 score of 0.840 on 1,252 public binary candidates. Encrypted scoring completes in 0.176 seconds per operation, preserving plaintext audit decisions in all measured cases. This addresses the risk of exposing sensitive user data during external safety checks.

Why it matters: Offers a concrete, evaluated solution to a critical privacy issue in LLM agent auditing, with measurable performance metrics.

arxiv4d agoScore 85

LLM safetyprivacyhomomorphic encryption

Research Editor selected

MSafeTemp: Template-Based Multilingual Safety Alignment Across 60 Languages

Researchers propose a scalable template-based augmentation pipeline that extracts safe templates from harmful prompts, translates them, and fills slots with culturally authentic harmful and benign entities. The work releases MSafeTemp, a validated corpus spanning 60 languages for multilingual safety alignment and evaluation. Experiments show English-only evaluation can misjudge multilingual safety: attacks that fail in English may succeed when translated, and culturally localized variants expose additional failures. Template-based augmentation also improves safety performance beyond English.

Why it matters: It provides a concrete multilingual safety dataset and pipeline that exposes gaps in English-only safety evaluation and training.

arxivOct 30, 2026Score 80

LLM safetymultilingual NLPsafety alignment