{
  "schema_version": "1.1",
  "id": "s13:https://arxiv.org/abs/2611.23218v1",
  "slug": "2611-23218v1-1ae46kz",
  "url": "https://feed7.dev/p/2611-23218v1-1ae46kz",
  "title": "What Does a Bias Head Measure? Non-Identifiability in Head-Level Bias Attribution",
  "why_included": "An audit finds head-level bias maps on two language models cannot be attributed to bias from observational data, and ablating selected heads did not reduce unfairness. The method may track prompt structure instead.",
  "summary": "The audit tested Qwen-2.5-7B-Instruct and Llama-3.1-8B-Instruct. Prompt identity explained **64%** of the selected statistic’s variance; response length explained **1%** after an intervention broke its link to labels. Ablating tested head sets did not lower unfairness.",
  "practical_implication": "If you use internal head maps to diagnose or mitigate model behavior, treat the score as a proxy until an intervention isolates what it measures. Validate the exact downstream behavior you care about, including refusal behavior.",
  "agent_context": "The audit tested Qwen-2.5-7B-Instruct and Llama-3.1-8B-Instruct. Prompt identity explained **64%** of the selected statistic’s variance; response length explained **1%** after an intervention broke its link to labels. Ablating tested head sets did not lower unfairness.\n\nIf you use internal head maps to diagnose or mitigate model behavior, treat the score as a proxy until an intervention isolates what it measures. Validate the exact downstream behavior you care about, including refusal behavior.\n\nWithin a fixed prompt, labels, response length, and protected-group tokens were highly collinear, limiting observational attribution. The paper reports refusal behavior was damaged by the complementary head set, but these results cover only two models and the tested pipeline.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2611.23218v1",
    "published_at": "2026-11-17T18:57:08.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "benchmark",
  "domains": [
    "research",
    "security"
  ],
  "topics": [
    "agent-evals",
    "benchmark-integrity"
  ],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "uncertainty": [
    "Within a fixed prompt, labels, response length, and protected-group tokens were highly collinear, limiting observational attribution. The paper reports refusal behavior was damaged by the complementary head set, but these results cover only two models and the tested pipeline."
  ],
  "connected_context": {
    "meaning": "The study warns that a head-level bias score may mostly track prompt identity rather than the intended construct, and that ablating selected heads did not reduce unfairness in the tested pipeline. This narrows what internal attribution can establish and makes downstream behavior checks essential, including refusal behavior that the complementary head set damaged.",
    "corpus_size": 1116,
    "generated_at": "2026-11-18T11:09:16.013Z",
    "connections": [
      {
        "title": "From Deceptive Outputs to Deceptive Mechanisms: A Causal Framework for Language-Model Deception Research",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2609.04166v1",
        "feed7_url": "https://feed7.dev/p/2609-04166v1-04gqa5r",
        "reason": "Both require causal evidence before attributing behavior to an internal mechanism: this audit finds a head-level statistic is confounded by prompt identity, while the deception framework separates observed output from the mechanism that produced it."
      },
      {
        "title": "BLOOM-WILT: Logit Tilting for Behaviour Elicitation in Automated LLM Auditing",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2608.31105v1",
        "feed7_url": "https://feed7.dev/p/2608-31105v1-0vdo3q3",
        "reason": "Both show that evaluation or intervention choices can distort behavioral conclusions: head ablations failed to reduce the target unfairness and affected refusal behavior, while BLOOM-WILT reports elicitation can change observed safety rankings."
      },
      {
        "title": "Comparative Analysis of Agent Evaluation Frameworks: Stability, Detection, and Discovery in Enterprise Analytics Agents",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2611.14493v1",
        "feed7_url": "https://feed7.dev/p/2611-14493v1-1ghdgem",
        "reason": "The Signal calls for validating downstream behavior after internal head interventions; the framework comparison reinforces that multiple evaluation methods can expose distinct failures, so one metric or check may miss consequential effects."
      }
    ]
  },
  "lifecycle": "Current",
  "published_at": "2026-11-17T18:57:08.000Z",
  "modified_at": "2026-11-17T18:57:08.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2611-23218v1-1ae46kz",
    "json": "https://feed7.dev/p/2611-23218v1-1ae46kz.json",
    "markdown": "https://feed7.dev/p/2611-23218v1-1ae46kz.md"
  }
}