{"slug": "actmap-single-pass-uncertainty-quantification-from-generation-time-activation", "title": "ActMap: Single-Pass Uncertainty Quantification from Generation-Time Activation Maps", "summary": "Researchers introduced ActMap, a white-box representation that compresses a large language model's generation-time hidden-state trajectory into a fixed 12 × 32 × 128 tensor occupying 96 KiB, enabling single-pass uncertainty quantification. A lightweight Vision Transformer classifier reads an estimated correctness probability from each map in under a millisecond, and ActMap matched ACT-ViT — a detector trained on dense activation tensors 67× larger — at essentially the same mean AUROC with lower calibration error on ten of twelve pairs. Tested in-domain on short-answer QA, direct-answer math, and summarization factuality with three instruction-tuned 7-8B models, ActMap outperformed sampling, token-probability, attention, and embedding baselines, supporting abstention, routing, and selective verification from a single generation.", "body_md": "arXiv:2609.11498v1 Announce Type: new \nAbstract: Practical uncertainty quantification (UQ) for large language models must decide,\n  from a single generation, whether a specific answer should be trusted. Existing\n  methods either sample multiple generations, read only output-token probabilities,\n  or reduce the model's internal computation to a single hidden state. We introduce\n  ActMap, a white-box representation that compresses the generation-time hidden-\n  state trajectory (every layer, every generated token) into a fixed $12 \\times 32\n  \\times 128$ tensor of temporal-statistic channels that preserves structure across\n  transformer depth and pooled hidden coordinates. The map is captured during the\n  generation pass with no measurable overhead, has a fixed shape across model\n  depths and hidden sizes, and occupies 96 KiB: a compact artifact that can be\n  retained for audit-relevant generations and probed directly, with occlusion\n  analysis localizing the classifier's signal to mid-depth regions of the map. A\n  lightweight classifier, instantiated as a compact Vision Transformer, reads an\n  estimated correctness probability from each map in a fraction of a millisecond;\n  capacity-matched MLPs perform comparably, indicating the representation itself\n  carries the result. Trained and evaluated in-domain on short-answer QA, direct-\n  answer math, and summarization factuality with three instruction-tuned 7-8B\n  models, ActMap consistently outperforms sampling, token-probability, attention,\n  and embedding baselines, and matches ACT-ViT, a detector trained on dense\n  activation tensors $67 \\times$ larger, at essentially the same mean AUROC with\n  lower calibration error on ten of twelve pairs. The resulting score supports\n  abstention, routing, and selective verification from a single generation, making\n  it a practical primitive for scalable oversight of deployed models.", "url": "https://wpnews.pro/news/actmap-single-pass-uncertainty-quantification-from-generation-time-activation", "canonical_source": "https://www.machinebrief.com/news/actmap-single-pass-uncertainty-quantification-from-generatio-37d5", "published_at": "2026-09-12 04:00:00+00:00", "updated_at": "2026-09-12 04:28:09.455121+00:00", "lang": "en", "topics": ["large-language-models", "ai-research", "ai-safety", "machine-learning"], "entities": ["ActMap", "ACT-ViT", "Vision Transformer", "arXiv"], "alternates": {"html": "https://wpnews.pro/news/actmap-single-pass-uncertainty-quantification-from-generation-time-activation", "markdown": "https://wpnews.pro/news/actmap-single-pass-uncertainty-quantification-from-generation-time-activation.md", "text": "https://wpnews.pro/news/actmap-single-pass-uncertainty-quantification-from-generation-time-activation.txt", "jsonld": "https://wpnews.pro/news/actmap-single-pass-uncertainty-quantification-from-generation-time-activation.jsonld"}}