ActMap Turns a Generation's Whole Activation Trajectory Into a 96 KiB Tensor You Can Score in Under a Millisecond

arXiv 2609.11498·medium signalXBlueskyLinkedInCopy link

Practical uncertainty quantification has to decide from one generation whether a specific answer is trustworthy, but current methods either sample repeatedly, read only output-token probabilities, or collapse the model's internals to a single hidden state. ActMap compresses the generation-time hidden-state trajectory across every layer and every generated token into a fixed 12x32x128 tensor of temporal-statistic channels, captured during the generation pass with no measurable overhead, fixed in shape across model depths and hidden sizes, and small enough (96 KiB) to retain as an audit artifact. A compact Vision Transformer reads a correctness probability from each map in a fraction of a millisecond, and capacity-matched MLPs do comparably, indicating the representation rather than the classifier carries the signal.