probe

An Nvidia employee has been detained by Taiwanese authorities in connection with a probe into the smuggling of chips to China.

Introduces PROPEL, a solver-amortized framework that trains a lightweight activation probe to predict solver pass rates, enabling efficient training of task generators for RL without costly solver rollouts. The method improves generation at the learnable frontier across math, code, and software-engineering tasks.

Introduces PRIG, a gradient attribution method that localizes prompt ambiguity in large language models by training a linear probe to distinguish clear from ambiguous prompts and attributing the probe score to token representations in the residual stream, achieving strong performance on synthetic and human-written benchmarks.

The paper argues that unlearning in LLMs should be goal-dependent, proposing a cosine-based meta-learned variant of RMU for dangerous knowledge and a multi-layer objective with probe directions for toxicity, achieving strong results across four 7-8B models.