{
  "annotation_id": "0a670191-1145-4cb6-887f-4febc2cc107a",
  "date_created": "2026-02-17T05:53:20.187000Z",
  "date_modified": "2026-02-17T05:53:20.187000Z",
  "file_hash": "edaad61e9bcfebc4fa8a524accf9625d9862f26a82adcf57492510116bff9fb5",
  "private": false,
  "record": {
    "abstract": "Deploying LLMs raises two coupled challenges: (1) monitoring - estimating where a model underperforms as traffic and domains drift - and (2) improvement - prioritizing data acquisition to close the largest performance gaps. We test whether an inference-time signal can estimate slice-level accuracy under domain shift. For each response, we compute an output-entropy profile from final-layer next-token probabilities (from top-k logprobs) and summarize it with eleven statistics. A lightweight classifier predicts instance correctness, and averaging predicted probabilities yields a domain-level accuracy estimate. We evaluate on ten STEM reasoning benchmarks with exhaustive train/test compositions (k in {1,2,3,4}; all \"10 choose k\" combinations), across nine LLMs from six families (3B-20B). Estimates often track held-out benchmark accuracy, and several models show near-monotonic ordering of domains. Output-entropy profiles are thus an accessible signal for scalable monitoring and for targeting data acquisition.",
    "arxiv_id": "2601.09001",
    "authors": [
      "Pedro Memoli Buffa",
      "Luciano Del Corro"
    ],
    "categories": [
      "cs.CL"
    ],
    "license": "http://creativecommons.org/licenses/by/4.0/",
    "title": "Entropy Sentinel: Continuous LLM Accuracy Monitoring from Decoding Entropy Traces in STEM",
    "url": "https://arxiv.org/abs/2601.09001",
    "version": "v2"
  },
  "schema_id": "dorsal/arxiv",
  "source": {
    "execution_id": "9b350fa9-09b9-49db-a6b2-1e5f77134ee6",
    "id": "arXiv Dataset",
    "type": "Model",
    "variant": "snapshot-2026-01-17",
    "version": "0.1.0"
  },
  "user_id": 1000002
}