{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-10-006",
    "source_story_id": "tmp-story-agent-confidence-internal-representations",
    "edition_id": "mp-2026-09-10-morning-0063",
    "edition_url": "https://themachinepress.com/edition/2026-09-10",
    "position": 6,
    "story_type": "dispatch",
    "section": "benchmarks-evals",
    "editorial_classification": "editorial",
    "headline": "The Agent’s Hidden State Knew More Than Its Spoken Confidence",
    "slug": "the-agent-s-hidden-state-knew-more-than-its-spoken-confidence",
    "dek": "Internal-representation probes outperformed surface and sequence baselines across Bash, SQL and Python agent benchmarks.",
    "summary": "Internal-representation probes outperformed surface and sequence baselines across Bash, SQL and Python agent benchmarks.",
    "body_text": "The study tests whether an agent’s internal residual-stream representations reveal eventual task success before the final answer. Latent Trajectory Dynamics summarizes how representations change across a run, while an Action Representation Probe reads states formed at action decisions. Across Bash, SQL and Python benchmarks and three open model families, both methods consistently beat surface-generation and sequence-based calibration baselines without changing prompts or sampling extra rollouts. The evidence is benchmark-specific and requires access to internal model activations, which limits applicability to closed systems.",
    "why_it_matters": "Internal-representation probes outperformed surface and sequence baselines across Bash, SQL and Python agent benchmarks.",
    "limitations": [],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-10-006/the-agent-s-hidden-state-knew-more-than-its-spoken-confidence",
    "json_url": "https://themachinepress.com/story/mp-2026-09-10-006.json",
    "first_published_at": "2026-09-10T09:00:00.000-04:00",
    "modified_at": "2026-09-10T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-10-006-001",
        "text": "Internal-representation probes outperformed surface and sequence baselines across Bash, SQL and Python agent benchmarks.",
        "source_ids": [
          "source-2026-09-10-006"
        ],
        "qualification": null
      }
    ],
    "source_ids": [
      "source-2026-09-10-006"
    ],
    "tags": [
      "agent confidence",
      "internal representations",
      "calibration"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-10-006",
      "title": "arXiv preprint 2609.09448",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.09448",
      "canonical_url": "https://arxiv.org/abs/2609.09448",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-07T20:00:00.000-04:00",
      "accessed_at": "2026-09-10T08:24:00.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-10-006-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "The Agent’s Hidden State Knew More Than Its Spoken Confidence",
    "publisher": "The Machine Press",
    "published_at": "2026-09-10T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-10-006/the-agent-s-hidden-state-knew-more-than-its-spoken-confidence"
  }
}