ThinkTrap
An optimization-based black-box adversarial attack strategy that optimizes latent variables in a continuous latent space using gradient estimation or evolutionary strategies so that the target model produces high-risk internal reasoning.
Timeline (last 90 days)
This study evaluated the dLLM mapping method under two attack settings: sampling-based RareTrap and optimization-based ThinkTrap.
In the optimization-based ThinkTrap, using dLLM mapping reduced character-level perplexity from 20–28 to 4.3–5.4.
RareTrap is sampling-based: it randomly samples latent variables in a continuous latent space and decodes them into prompts, searching for samples that can trigger abnormal behaviors such as overthinking; ThinkTrap is optimization-based: it uses gradient estimation or evolutionary strategies to optimize latent variables so that the model produces high-risk internal reasoning.
All Facts (3)
This study evaluated the dLLM mapping method under two attack settings: sampling-based RareTrap and optimization-based ThinkTrap.
50%UnverifiedIn the optimization-based ThinkTrap, using dLLM mapping reduced character-level perplexity from 20–28 to 4.3–5.4.
50%UnverifiedRareTrap is sampling-based: it randomly samples latent variables in a continuous latent space and decodes them into prompts, searching for samples that can trigger abnormal behaviors such as overthinking; ThinkTrap is optimization-based: it uses gradient estimation or evolutionary strategies to optimize latent variables so that the model produces high-risk internal reasoning.
50%