4 papers
When Cosine Similarity Fails to Reflect Linearly Accessible Structure in Dialogue Models
Yu Sun, Mengyin Lu, Cong Feng +2
Cosine similarity is widely used to analyze transformer representations, implicitly assuming that similarity reflects task-relevant structure. We study when this assumption fails i…
K/V-Cache Interventions Dissociate Representation Alignment from Persona Expression in Decoder-Only Language Models
Yu Sun, Mengyin Lu, Cong Feng +2
We study K/V-cache interventions -- transplanting a target-conditioned K/V trajectory into a source-persona generation -- as a structured surface for persona control in decoder-onl…
SPEAR: Code-Augmented Agentic Prompt Optimization
Mengyin Lu, Cong Feng, Huimin Han +6
Automatic prompt engineering (APE) rewrites prompts to improve downstream task performance, but existing APE loops treat the optimizer itself as a fixed pipeline. We port the code-…
A Set of Generalized Components to Achieve Effective Poison-only Clean-label Backdoor Attacks with Collaborative Sample Selection and Triggers
Zhixiao Wu, Yao Lu, Jie Wen +3
Poison-only Clean-label Backdoor Attacks aim to covertly inject attacker-desired behavior into DNNs by merely poisoning the dataset without changing the labels. To effectively impl…