Consistency Without Alignment: Item-Sensitive Language Models Indistinguishable From Random

2609.00576 ↗ · 2026-09-01 · cs.AI, cs.CL · h-max 0 · 0 citations

Cris Huynh

13 pages, 3 figures

The finding

Headline claim

A language model being reliably item-sensitive (its choice depends on the input) is not evidence it is doing the task: in a forced-choice signalling task with closed-form references, 8 of 21 model-by-rule cells are statistically indistinguishable from random and 5 score worse than random despite universal item-sensitivity.

The test

Seven language models across two model families and four orders of magnitude of parameter count, plus a post-training ablation, scored on one pre-registered Deception-abstracted signalling tile (100 hypotheses, one tile carrying the primary analysis) under three token-normalisation scoring rules, against exactly computable fit-maximising, posterior-maximising (Bayesian), and uniform-random references built from human-derived similarity data (SPoSE, THINGSplus, SWOW).

Effect size & scope

Item-sensitivity is positive in 21 of 21 cells; model coordinates range from indistinguishable from random to 91 percent of the way to a salience reference; item-sensitivity and distance from random correlate at only r = 0.30; a literal cosine-similarity baseline with no pragmatics outperforms most models; adding an RSA-style pragmatic layer moves choosers toward random; the labelled multiple-choice format carries no content signal (permutation consistency at or below chance in all 32 model-by-tile-by-label cells).

Limitations

The authors concede: no human data collected yet (instrument designed and piloted only), the primary analysis rests on a single tile, the oracle is a literal-listener account, the vocabulary carries a measured natural-kind sampling bias from SWOW cue construction, and all model-side measurements are post hoc against the fixed instrument. Unstated: three of four tiles carry no item-sensitivity, two conditions are structurally ineligible for the pragmatic layer, the setting is one narrow abstracted party game far from any economic or real-world evaluation, and results may not transfer to standard benchmarks.

Counterpoint

This is a single-author, single-tile, single-task lab study with no human comparison data in hand; the finding that models fail a deliberately ill-fitting closed-menu game may say little about whether item-sensitivity metrics mislead on the benchmarks actually used in the field, where no closed-form oracle exists precisely because real tasks are open-ended; the dramatic framing (consistency without alignment) generalises from a stylised instrument whose main virtue is analytic tractability.

Economy implication

Little or nothing yet for AI economics: the paper is a methodological caution about evaluation methodology, and its only economic relevance is indirect, that headline consistency statistics used as evidence of agentic competence may overstate readiness without independent baselines; no cost, revenue, or labor quantity is measured.

Who this matters for

Technical Capability: Anyone reading permutation-consistency or self-consistency numbers as evidence of task competence should demand an independent reference for the measured quantity, since this paper shows the statistic alone cannot distinguish responding to the intended feature from responding to any input feature.

Societal & Cultural: Eval consumers and journalists treating benchmark consistency as proof of understanding get a concrete demonstration that a model can look responsive and still be statistically indistinguishable from a random chooser.

Abstract

Item-sensitivity, defined as whether a model's choice depends on the specific input rather than on its own output prior, is widely reported as evidence of task competence. We show this evidence is necessary but not sufficient using a forced-choice signalling task abstracted from the board game Deception: Murder in Hong Kong. In this environment, the reference points against which a coordinate should be judged (a fit-maximising strategy, a posterior-maximising strategy, and uniform random selection) are all computable in closed form. Across seven language models, two model families, a post-training ablation, and three independent scoring rules, every one of 21 model-by-rule cells is reliably item-sensitive. Yet 8 of those 21 cells are not statistically distinguishable from a chooser that ignores the item and selects at random, and 5 score worse than random at describing the target. Item-sensitivity and distance from random correlate at only r = 0.30. We call this consistency without alignment and argue it generalises to any evaluation that relies on item-sensitivity, permutation consistency, or self-consistency without an independent reference for the measured quantity. We further find that a literal-similarity baseline with no pragmatics outperforms most tested language models, that adding a pragmatic layer over two baseline similarity sources moves choosers toward random rather than toward the Bayesian reference, and that a standard labelled multiple-choice format carries no measurable content signal here. All results represent the model side of a pre-registered instrument; a matched human condition is designed and piloted but not yet collected.