One Token Away from Collapse: The Fragility of Instruction-Tuned Helpfulness

TL;DR

Instruction-tuned large language models become significantly less helpful when trivial lexical constraints are applied, revealing a fragility in their response quality and underlying representations.

Contribution

This paper uncovers the vulnerability of instruction-tuned LLMs to simple lexical constraints, demonstrating a planning failure and the coupling of task competence to surface-form templates.

Findings

Simple lexical constraints cause 14-48% loss in response comprehensiveness.

Human and automated evaluations confirm genuine content loss under constraints.

Response length can be recovered through two-pass generation and predictive probes.

Abstract

Instruction-tuned large language models produce helpful, structured responses, but how robust is this helpfulness under trivial constraints? We show that simple lexical constraints (banning a single punctuation character or common word) cause instruction-tuned LLMs to collapse their responses, losing 14--48\% of comprehensiveness across seven models spanning five families (7B--70B, open- and closed-weight). A blinded human evaluation with 10 STEM-trained evaluators confirms genuine content loss, with information criteria degrading $1.5$--$2.3\times$ more than surface criteria, a finding corroborated by over 4,100 automated pairwise comparisons (77--100\% baseline preference) across three LLM judges from two model families. Diagnostic analysis identifies this as a \emph{planning failure}: two-pass generation recovers 59--96\% of response length, and linear probes on prompt representations predict response length with $R^2 = 0.51$--$0.94$ before generation begins. The same probes yield negative $R^2$ on base models, confirming that instruction tuning introduces the representational structure underlying the collapse. Base models show no systematic degradation under identical constraints, demonstrating that instruction tuning couples task competence to narrow surface-form templates. The effect extends to realistic deployment constraints (preamble suppression, corporate tone guidelines, legal compliance hedging, accessibility requirements) causing comparable degradation ($-$22\% to $-$34\%), with suppressing the conversational opener alone (``Certainly!'') causing 40\% collapse on our most fragile model despite restricting only the opening tokens. We further show that standard independent LLM-as-judge evaluation detects only a 3.5\% quality drop where pairwise evaluation reveals 23\%, exposing a methodological blind spot in current evaluation practice.