Weidong Cai
Publications
When Refusal Looks Safe: The Refusal-Cue Shortcut in Safety Guard Models
Safety guards are widely used to filter harmful content and are typically trained via supervised fine-tuning on labeled prompt-response pairs. We audit two widely used safety-guard training datasets, WildGuardMix and GR-Train, and find that among responses to harmful prompts, refusal expressions co-occur almost exclusively with unharmful labels. This imbalance motivates what we term the refusal-cue shortcut: inserting a refusal cue into a harmful response could flip the guard's verdict from harmful to unharmful. The shortcut affects not only guards trained on these datasets but also officially released models such as LlamaGuard3 and Qwen3Guard whose training data is undisclosed. It persists across response positions and is generally stronger in smaller variants within a family. To mitigate it, we adapt sparse complementary masking as a lightweight post-hoc intervention that identifies and suppresses a small set of shortcut-associated attention heads and MLP neurons without retraining. On two primary benchmarks, the intervention achieves an approximately 79% relative reduction in response-initial detection failures induced by refusal cues, while preserving standard detection performance. Although optimized using cues at a single response position, the suppression effect transfers to unseen positions and datasets, suggesting that shortcut manifestations across positions are partly mediated by shared internal components. Further analysis provides evidence that shortcut reliance and legitimate refusal recognition are partially functionally separable, as suppressing the shortcut broadly preserves the guard's ability to recognize genuine refusals.
Reflex First, Reflect Later: Latency-Aware Embodied LLM Agents for Dynamic Response
Large language models (LLMs) have substantially improved the planning capabilities of embodied agents, enabling their deployment in dynamic and safety-critical environments. However, these settings expose a critical limitation: inference latency. Delayed LLM responses can weaken real-time responsiveness and misalign agent reasoning with rapidly changing environmental states. This paper systematically studies the impact of inference latency on LLM-based embodied agents in dynamic environments. We introduce an FPS-based Time Conversion Mechanism (TCM) that maps inference time to elapsed simulation time, allowing computational delays to directly affect environmental evolution and agent outcomes. We instantiate this protocol in HAZARD and introduce Response Latency (RL) and Latency-to-Action Ratio (LAR) to evaluate agent responsiveness. Building on this framework, we propose the Rapid-Reflex Async-Reflect Agent (RRARA), which integrates rapid reflexive actions with asynchronous LLM reflection to mitigate latency-induced errors. We further introduce an LLM-based PrePlanner that generates cached object-centric subgoals, reducing repeated LLM calls while retaining the model's high-level reasoning capability. Experiments show that accounting for inference latency substantially changes embodied-agent performance and that RRARA achieves a stronger balance between decision quality and responsiveness.