Entity · model

Qwen2.5-3B

modelactiveprovisionalqwen2-5-3b-ff43ac4b·1 events·first seen 3d ago

Aliases: Qwen2.5-3B

Co-occurring entities

SafeAgentBench Phi-3.5 Qwen2.5-7B EARBench SmolLM2 SafeText Qwen2.5-14B PhysicalSafetyBench-1K Qwen-2.5-VL-3B Prism

More like this (12)

Qwen2.5-7B Qwen2.5-8B Qwen2.5-0.5B Qwen2.5-14B Qwen3.5-0.8B Qwen2.5-1.5B Qwen-2.5-VL-3B Qwen3.5-2B-Base Qwen2.5 Qwen3.5-Plus Qwen 2.5-7B Qwen3.5-122B

Recent events (1)

6arXiv · cs.AI·3d ago·source ↗

PRISM probe separates physical danger from text-level safety in LLM hidden states for embodied agents

Researchers introduce PRISM, a single-layer logistic probe over LLM hidden states that distinguishes physically grounded danger (e.g., unsafe robot actions) from ordinary text-level content danger, showing these are separable signals in representations across multiple model families. PRISM achieves 86.2–87.7% accuracy on SafeAgentBench with substantially lower false-positive rates than same-scale LLM judges, which over-block safe tasks at 24.7–39.0% FPR. The authors also release PhysicalSafetyBench-1K (PSB-1K), a 1,000-pair contrastive benchmark for evaluating physical-risk detection without relying on explicit harm keywords. The work is relevant to safety in embodied AI and agentic systems where linguistic safety filters are insufficient.

Evaluation and Benchmarking AI Safety Research SafeAgentBench Phi-3.5 Qwen2.5-7B +9 more