PapersSelected

ActMap: Single-Pass Uncertainty Quantification from Generation-Time Activation Maps

Why we picked it

ActMap provides a practical way to assess the reliability of a model's output in a single pass, which is crucial for building more trustworthy AI systems.

ActMap is a new method for uncertainty quantification in large language models.

Body

It compresses the generation-time hidden-state trajectory into a fixed 12x32x128 tensor called an activation map.

This map is captured during generation with no measurable overhead and can be probed by a lightweight classifier to estimate correctness probability in a fraction of a millisecond.

ActMap was trained and evaluated on short-answer QA, direct-answer math, and summarization factuality with three 7-8B instruction-tuned models, outperforming baseline methods like sampling and token-probability.

Original · arXiv cs.AI

Practical uncertainty quantification (UQ) for large language models must decide, from a single generation, whether a specific answer should be trusted. Existing methods either sample multiple generations, read only output-token probabilities, or reduce the model's internal computation to a single hidden state. We introduce ActMap, a white-box representation that compresses the generation-time hidden- state trajectory (every layer, every generated token) into a fixed $12 \times 32 \times 128$ tensor of temporal-statistic channels that preserves structure across transformer depth and pooled hidden coordinates. The map is captured during the generation pass with no measurable overhead, has a fixed shape across model depths and hidden sizes, and occupies 96 KiB: a compact artifact that can be retained for audit-relevant generations and probed directly, with occlusion analysis localizing the classifier's signal to mid-depth regions of the map. A lightweight classifier, instantiated as a compact Vision Transformer, reads an estimated correctness probability from each map in a fraction of a millisecond; capacity-matched MLPs perform comparably, indicating the representation itself carries the result. Trained and evaluated in-domain on short-answer QA, direct- answer math, and summarization factuality with three instruction-tuned 7-8B models, ActMap consistently outperforms sampling, token-probability, attention, and embedding baselines, and matches ACT-ViT, a detector trained on dense activation tensors $67 \times$ larger, at essentially the same mean AUROC with lower calibration error on ten of twelve pairs. The resulting score supports abstention, routing, and selective verification from a single generation, making it a practical primitive for scalable oversight of deployed models.

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Key entities and technical topics under continuous tracking include ActMap, uncertainty quantification.

Source attribution: curated from arXiv cs.AI (@Jacopo Dardini, Roberta Calegari).