Robust Is Salient: How an Informed Adversary Moves the Optimal Signal onto Salience in Forced-Choice Tasks

Forum topic · 小凯 · 2026-10-05

Summary

A new arXiv paper (2610.00233) by Cris Huynh studies a forced-choice signalling game, abstracted from the board game Deception: Murder in Hong Kong, in which an informed adversary knows the target, observes the signal, and argues for the strongest wrong answer under a persuasion budget beta. The key finding: the signal that best protects the truth against such an adversary is also the signal that best describes it. Across 108 confirmatory items, the adversary-robust optimum aligns exactly with the salience pole identified in prior work; in a 200,000-item pool the two differ on only 2,748 items, precisely where the prior salience-to-Bayes coordinate is undefined. As the budget beta grows, the optimal signal shifts from posterior-maximizing to marginal-maximizing options, with beta = 0 recovering the original oracle model at listener temperature tau = 1. The effect is substantive: 18.2% of items admit a finite-budget optimum shift with exactly computable thresholds. Evaluations across seven language models show 30 to 77 of 108 items changing options, yet no measurement can distinguish movement toward the adversary's perceived optimum from movement toward salience, since the two coincide—a structural limitation rather than a null result.

Paper Overview

Field: NLP
Author: Cris Huynh
Published: 2026-10-05
arXiv: 2610.00233

Abstract

When an informed adversary shares the audience of a constrained signalling channel, the signal that best protects the truth is the signal that best describes it. On 108 confirmatory items, the adversary-robust optimum aligns exactly with the salience pole from prior work. Across a 200,000-item pool, the two differ on only 2,748 items — lying exactly where the prior salience-to-Bayes coordinate is undefined. Where defined, robustness is achieved by moving from Bayesian discrimination entirely to salience.

Key Findings

The game is a forced-choice task abstracted from *Deception: Murder in Hong Kong*. The adversary knows the target, observes the signal, and argues for the strongest wrong answer using a persuasion budget $\beta$ .
As $\beta$ grows, the optimal signal shifts from posterior-maximizing to marginal-maximizing options; at $\beta = 0$ the game reproduces the original oracle model with listener temperature $\tau = 1$ .
The effect is real: 18.2% of items admit an optimum shift under finite budgets, with critical budgets exactly computable.
The coincidence structurally constrains empirical evaluation: two adversary frameworks change 30 to 77 of 108 items' options across seven language models, relative to an exact zero-effect rate.
However, no measurement can tell whether the movement is toward the adversary's perceived optimum or toward salience — the two options are identical. This is a structural limitation, not a null result.
Diagnostic checks are cheap: before evaluating adversary perception, first verify whether the robust objective coincides with the heuristic objective on the evaluation items.

---

*Auto-collected on 2026-10-05.*