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

Apr 14, 2026•Erfan Baghaei Potraghloo, Seyedarmin Azizi, Souvik Kundu +1•View PDF

TL;DR Highlight

We discovered that LLM responses can shrink by up to 48% with a single instruction: "Don't use commas".

Who Should Read

Backend developers and prompt engineers who apply format/safety constraints to LLMs in production. Specifically, teams operating AI services that prohibit the use of specific words or symbols in system prompts.

Core Mechanics

Even a simple constraint prohibiting punctuation like commas or colons results in a 14-48% loss of response coverage in Llama-3.1-8B-Instruct, Qwen-2.5-7B-Instruct, Mistral-7B-Instruct, and GPT-4o-mini.
This isn't a capability limitation, but a 'planning failure'. Using a two-pass approach – generating freely first and then rewriting with constraints – can recover 59-96% of the original response length.
Instruction-tuned models already decide 'whether to write short or long' upon receiving the prompt. Response length can be predicted with R²=0.51~0.93 level accuracy using a linear probe in the intermediate layer hidden state.
Base models (original models without instruction tuning) are barely affected by the same constraints. Applying the same probe to base models results in negative R², confirming that instruction tuning creates the vulnerability.
GPT-4o-mini is no exception. When commas are prohibited, response length decreases by 54% (from 472 to 216 words) and coverage is lost by 31%. Even closed-weight models, which are robust to format constraints (e.g., JSON output), crumble under lexical constraints.
Standard independent evaluation (LLM-as-judge individual scoring) barely catches this loss. Actual 23% quality degradation is underestimated 6.7x by independent evaluation, which only detects 3.5%.

Evidence

In 1,920 pairwise comparisons, unconstrained baseline responses were preferred 77-100% of the time. Qwen-2.5-7B-Instruct showed a 48.1% coverage loss and a 99.7% win rate based on GPT-4o judging criteria.
Two-pass recovery rates: Llama 96%, Mistral 91%, Qwen 59%. The two-pass approach fully recovers length in the Llama no-comma condition.
Linear probe R²: Qwen-2.5-7B-Instruct 0.925, Llama-3.1-8B-Instruct 0.747, Mistral-7B-Instruct 0.514. In contrast, base models show Llama -4.04 and Qwen -0.59 across all layers.
Independent evaluation vs. pairwise comparison gap: For Llama-3.1-8B-Instruct with the no-comma constraint, independent evaluation showed -5.4% while pairwise showed -27.0%, a difference of approximately 5x.

How to Apply

If you prohibit specific words or symbols in your system prompt, be sure to add pairwise evaluation. A/B testing based solely on independent scores can underestimate quality degradation by a factor of 6.7.
If you need to block certain expressions due to safety filters or brand guidelines, consider a two-pass approach. Generate freely first, then prompt a second call to 'rewrite this content while adhering to [constraints]', which can significantly reduce quality loss.
When building an LLM quality evaluation pipeline, always include side-by-side comparison with unconstrained responses for configurations with constraints (length limits, prohibited words, format constraints, etc.).

Code Example

snippet

# Use a two-pass approach to reduce quality loss under constraints
import openai

client = openai.OpenAI()

def two_pass_generate(question: str, constraint: str, model: str = "gpt-4o-mini") -> str:
    # Step 1: Generate freely without constraints
    free_response = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": question}],
        temperature=0.7
    ).choices[0].message.content

    # Step 2: Rewrite the result from step 1 under constraints (explicitly mention maintaining coverage)
    rewrite_prompt = f"""Rewrite the following response while adhering to the constraints below.
Be sure to maintain the same level of detail and structure as the original. Do not shorten the content.

Constraint: {constraint}

Original response:
{free_response}

Rewritten response:"""

    constrained_response = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": rewrite_prompt}],
        temperature=0.7
    ).choices[0].message.content

    return constrained_response

# Example usage
question = "Explain gradient descent easily"
constraint = "Do not use commas."

result = two_pass_generate(question, constraint)
print(result)

Terminology

Instruction TuningThe process of further training an LLM by showing it 'question-answer' pairs to behave like a helpful assistant. A key step in making models conversational, like ChatGPT or Claude.

Lexical ConstraintA text-level constraint that prohibits the use of specific words or symbols. Examples include 'Don't use commas' or 'Don't use "the"'.

Linear ProbeA simple linear regression analysis to check how much information can be read from the model's internal state (hidden state). It acts like a stethoscope to see what the model 'knows'.

R²A metric indicating how well a prediction model explains the variance of actual data. Closer to 1 means more accurate prediction, and 0 or less means worse than average.

Pairwise EvaluationAn evaluation method that compares two responses side-by-side to determine which is better. It is much more sensitive to differences than independent evaluation (assigning scores to each individually).

LLM-as-JudgeAn evaluation method where a powerful LLM like GPT-4 scores the quality of another LLM's output. AI acting as a judge instead of a human.

Two-pass GenerationA two-step generation method that first generates freely and then rewrites the result with constraints. A way to circumvent the problem of quality degradation when applying constraints at once.

Hidden StateInternal vector values created by the LLM at each layer when processing tokens. Can be considered the model's 'intermediate thought process'.

Original Abstract (Expand)

Instruction-tuned large language models produce helpful, structured responses, but how robust is this helpfulness when trivially constrained? 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 in pairwise evaluation across three open-weight model families and one closed-weight model (GPT-4o-mini). The baseline response is preferred in 77--100% of 1,920 pairwise comparisons judged by GPT-4o-mini and GPT-4o. Notably, GPT-4o-mini suffers 31% comprehensiveness loss (99% baseline win rate), demonstrating that the fragility extends to commercially deployed closed-weight models, contrary to prior findings on format-level constraints. Through mechanistic analysis, we identify this as a planning failure: two-pass generation (free generation followed by constrained rewriting) recovers 59--96% of response length, and linear probes on prompt representations predict response length with $R^2 = 0.51$--$0.93$ before generation begins, with $R^2$ tracking collapse severity across models. The same probes yield negative $R^2$ on base models, confirming that instruction tuning creates the representational structure encoding the collapse decision. Crucially, base models show no systematic collapse under identical constraints, with effects that are small, noisy, and bidirectional, demonstrating that instruction tuning creates this fragility by coupling task competence to narrow surface-form templates. The effect replicates on MT-Bench across all eight task categories. We further show that standard independent LLM-as-judge evaluation detects only a 3.5% average quality drop where pairwise evaluation reveals 23%, exposing a methodological blind spot in how constrained generation is assessed.