Papers by Guangsheng Cheng
Mitigating Safety Context Amnesia in Multimodal Reasoning Models via Intent-Guided Safety Reasoning (2026.acl-long)
Copied to clipboard
| Challenge: | Recent advances in Multimodal Large Reasoning Models (MLRMs) have enabled explicit chain-of-thought inference across vision and language, improving performance on complex cognitive tasks. |
| Approach: | They propose an inference-time defense that uses a percept decoupler to extract objective visual evidence into a structured intent output and a cognitive arbiter to enforce explicit safety constraints prior to generation. |
| Outcome: | The proposed defense improves defense success rates by over 62% compared to baselines while preserving task utility. |