Activation maps enable fast uncertainty estimates for single answers
ActMap: Single-Pass Uncertainty Quantification from Generation-Time Activation Maps
Artificial Intelligence
Summary
Knowing when to trust a computer's answer is important, especially for language models that generate text. The authors created ActMap, a way to look inside the model’s internal steps during answer generation and produce a simple map that summarizes this process. This map can be quickly analyzed by a small classifier to judge how likely the answer is correct, without needing to sample multiple answers or do extra heavy computations. Their experiments show ActMap works better than other methods on tasks like short answers, math, and summarization.
What this means in practice
- •For ai service operators: Estimate answer reliability from a single generated response to support deciding when to show or verify answers automatically.
- •For customer support teams: Identify uncertain AI-generated replies in chatbots to trigger escalation or human review without slowing down user interactions.
Authors
Jacopo Dardini, Roberta Calegari
Abstract
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.