Large vision-language models (LVLMs) exhibit strong multimodal in-context learning (ICL) capabilities, yet this ability degrades substantially as model size decreases. Knowledge distillation offers a natural way to bridge this gap, but existing methods primarily align output distributions or hidden representations dire...
The Selection--Realization Hypothesis is proposed, which views demonstrations as inducing a compact family of internal changes from which the query selects, while the model's computation constrains how the selected change can be implemented.
SVF is proposed, which learns a differentiable concept scoring function whose local gradient defines the steering direction at each activation, making interventions explicitly context-dependent and enables efficient long-form and multi-attribute control within a unified framework.
TRACES is a representation-based proactive auditor that learns prefix-level trajectory risk states from the hidden representations of an observer LLM, and it is suggested that these risk states can help train a safer agent, highlighting the broader potential of proactive auditing for long-horizon agent safety.