# Resist and Update: Counterfactual Report Coordinates for Incentive-Compatible LLMs

Source: [arXiv](https://arxiv.org/abs/2607.12985v1)  
Feed7 permalink: https://feed7.dev/p/2607-12985v1-1yo2sej  
Published: 2026-07-14T17:28:25.000Z  
Trust: Needs Review (needs_review)

## 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.

## Source 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-Ready 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.

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.

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.

## Context Map

- Layer: benchmark
- Domains: research
- Topics: agent-evals, agent-reliability

## 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.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.