{
  "schema_version": "1.0",
  "id": "s13:https://arxiv.org/abs/2607.12985v1",
  "slug": "2607-12985v1-1yo2sej",
  "url": "https://feed7.dev/p/2607-12985v1-1yo2sej",
  "title": "Resist and Update: Counterfactual Report Coordinates for Incentive-Compatible LLMs",
  "why_included": "A training-free activation clamp separated resistance to user pressure from responsiveness to evidence in a controlled benchmark, but its deployable single-pass version lost substantial resistance.",
  "summary": "The method identifies separate activation coordinates for answers, confidence, and caveats, then clamps reports against an incentive-neutralized counterfactual. Its two-pass form reached **1.00 resist and 1.00 update** on a Bayesian-witness benchmark.",
  "practical_implication": "For agents making consequential judgments, test two behaviors separately: refusing unsupported pressure and changing when real evidence arrives. Resist-only tuning can suppress legitimate updates, while explicitly training both objectives performed better here.",
  "agent_context": "The method identifies separate activation coordinates for answers, confidence, and caveats, then clamps reports against an incentive-neutralized counterfactual. Its two-pass form reached **1.00 resist and 1.00 update** on a Bayesian-witness benchmark.\n\nFor agents making consequential judgments, test two behaviors separately: refusing unsupported pressure and changing when real evidence arrives. Resist-only tuning can suppress legitimate updates, while explicitly training both objectives performed better here.\n\nThe two-pass result is a causal certificate under a constructible reference, not a deployment recipe. The single-pass compilation fell to **0.73 resist and 0.97 update**, despite reproduction across **three model families** and transfer to SycophancyEval.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2607.12985v1",
    "published_at": "2026-07-14T17:28:25.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "benchmark",
  "domains": [
    "research"
  ],
  "topics": [
    "agent-evals",
    "agent-reliability"
  ],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "uncertainty": [
    "The two-pass result is a causal certificate under a constructible reference, not a deployment recipe. The single-pass compilation fell to **0.73 resist and 0.97 update**, despite reproduction across **three model families** and transfer to SycophancyEval."
  ],
  "lifecycle": "Current",
  "published_at": "2026-07-14T17:28:25.000Z",
  "modified_at": "2026-07-14T17:28:25.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2607-12985v1-1yo2sej",
    "json": "https://feed7.dev/p/2607-12985v1-1yo2sej.json",
    "markdown": "https://feed7.dev/p/2607-12985v1-1yo2sej.md"
  }
}