Search papers, labs, and topics across Lattice.
The paper introduces "Evaluation without Generation," a framework for assessing harmful model specialization by analyzing internal representations rather than generated outputs, addressing limitations in auditing open-weight generative models, particularly in domains like CSAM. They propose Gaussian probing, which characterizes how LoRA adaptors perturb a model's internal representations by measuring responses to Gaussian latent ensembles. Experiments demonstrate that Gaussian probing effectively distinguishes benign from harmful specialization in high-risk domains, even against adversarial manipulations like weight rescaling, offering a scalable alternative to output-based evaluation.
You can now detect harmful specializations in generative models, like those trained on CSAM, without ever generating a single risky output.
Auditing the fine-tunes of open-weight generative models for harmful specialization has become a new governance challenge for model hosting platforms. The standard toolkit, generative evaluation via curated prompts or red-teaming, does not scale to platform-level auditing and breaks down entirely for domains like CSAM where generation is legally constrained. This motivates the Evaluation without Generation problem: assessing model capabilities without producing outputs. We argue that in such settings, capability must be inferred from the model's state, either its parameters or internal representations, rather than its outputs. We introduce Gaussian probing, a method that characterizes how LoRA adaptors perturb a model's internal representations by measuring responses to Gaussian latent ensembles. Unlike raw-weight baselines, Gaussian probing reliably distinguishes benign from harmful specialization without sampling outputs. We demonstrate effectiveness in high-risk domains, including detecting models specialized for child sexual abuse material (CSAM), where output-based evaluation is legally and ethically constrained. Our results show that Gaussian probing provides a scalable non-generative alternative for evaluating high-risk generative systems and remains robust to weight rescaling, a representative adversarial manipulation.
