ActMap compresses a generation into a 96 KiB tensor

A new technique called ActMap compresses a model's full generation into a 96 KiB tensor, captured with no measurable slowdown during generation.

That matters because judging a single answer usually means sampling a model several times, or reading only the output-token probabilities and missing what happened inside. ActMap keeps the whole internal trajectory instead, across every layer and every generated token, then scores it with a classifier in under a millisecond.

The ActMap paper trains a compact Vision Transformer to read that tensor and return a correctness probability. Separately trained MLPs, matched for capacity, score about as well. That result puts the signal in the compressed representation itself, not in whichever classifier reads it.

96 KiB per generation is small enough to log for every answer a system produces, not a sampled subset of them. For anyone building correctness checks into a live generation system, the paper doesn't say whether a tensor this small still carries the signal on a different model or task than the ones it was tested on.