💬 NLP

Revisiting Prompt Sensitivity in Large Language Models for Text Classification: The Role of Prompt Underspecification

This paper argues that a significant portion of observed prompt sensitivity in large language models stems from prompt underspecification rather than inherent model instability, demonstrating that specific instructions reduce performance variance and logit uncertainty with only marginal effects on internal representations.

Original authors: Branislav Pecher, Michal Spiegel, Robert Belanec, Jan Cegin

Published 2026-02-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Branislav Pecher, Michal Spiegel, Robert Belanec, Jan Cegin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to get a very smart, but slightly literal-minded robot to sort a pile of mail into "Junk" and "Important."

If you hand the robot a note that just says, "Here is a letter: [Letter Content]," and ask it to sort it, the robot might get confused. It might start writing a poem about the letter, or it might just stare at it blankly. If you change the note slightly to say, "Here is a letter: [Letter Content]. Please sort it," the robot might suddenly work perfectly. If you change it again to "Sort this: [Letter Content]," it might fail again.

This paper argues that the robot's confusion isn't because it's "bad" at sorting mail. It's because you didn't give it clear enough instructions.

Here is a breakdown of the paper's findings using simple analogies:

1. The Problem: "Vague Notes" vs. "Clear Instructions"

The researchers found that many studies testing AI models use "underspecified" prompts. Think of these as vague notes like "What is this?" or just pasting text without context.

The Result: When the instructions are vague, the AI's performance is all over the place. One day it's a genius; the next, it's guessing randomly. The paper calls this "prompt sensitivity."
The Cause: The AI actually knows how to do the task deep down, but because the instructions were too vague, it didn't know how to show you the answer. It's like a student who knows the math but forgets to write down the final number because the teacher didn't ask for it clearly.

2. The Solution: "Instruction Prompts"

When the researchers switched to "instruction prompts" (clear, specific directions like "Read this text and tell me if it is positive or negative"), the AI became much more stable and accurate.

The Analogy: It's the difference between handing a chef a raw ingredient and saying "Make something" versus saying "Make a grilled cheese sandwich." The chef (the AI) has the skills either way, but the second instruction ensures the output is what you actually wanted.

3. The "Magic Trick": In-Context Learning

The paper tested several ways to fix the AI's confusion. The most effective method was In-Context Learning (ICL).

The Analogy: Instead of just giving the robot a rule, you show it two examples first: "Here is a happy movie review. Here is a sad one. Now, here is a new review: Is it happy or sad?"
The Finding: This simple trick worked almost as well as complex mathematical fixes (called "calibration") that require digging into the robot's brain. It solved the confusion without needing to change the robot's internal code.

4. Peeking Inside the Robot's Brain (Logits and Probes)

The researchers looked under the hood to see where the confusion happened.

Logits (The "Confidence Score"): When the AI is confused by vague instructions, its internal "confidence scores" for the right answer are tiny—almost zero. It's like the robot whispering an answer so quietly you can't hear it. When instructions are clear, the robot shouts the answer with high confidence.
Linear Probes (The "Brain Scan"): They scanned the AI's internal layers (its "thought process").
- The Surprise: The AI's brain actually understood the task even when the instructions were vague. The "knowledge" was there in the middle layers of the brain.
- The Glitch: The problem only happened at the very end, when the AI tried to translate that knowledge into a final answer. The vague instructions messed up the output, not the thinking.

5. What Didn't Work

The paper tested a few other ideas, like adding "empty" tokens (nonsense words) to the prompt to "calibrate" the AI.

The Result: This often made things worse, causing the AI to perform even worse than random guessing. It's like trying to fix a confused chef by shouting random words at them; it just adds more noise.

The Bottom Line

The paper concludes that a huge chunk of the "instability" we see in AI models isn't because the models are broken or unpredictable. It's because we are asking them questions in a confusing way.

If we want reliable AI results, we need to stop using vague, minimal prompts and start using clear, specific instructions (and maybe show the AI a few examples first). The AI is capable of the job; it just needs a better job description.

Technical Summary: Revisiting Prompt Sensitivity in Large Language Models for Text Classification

Problem Statement

Large Language Models (LLMs) are increasingly utilized for zero-shot and few-shot text classification, where task behavior is dictated entirely by natural language prompts. A significant body of literature has established that LLMs exhibit high sensitivity to prompt variations, where minor phrasing changes can drastically alter performance rankings and state-of-the-art claims. However, this paper argues that a substantial portion of the observed sensitivity is not an inherent flaw in the models' reasoning capabilities, but rather a consequence of prompt underspecification.

Many existing studies investigating prompt sensitivity rely on "minimal" or "underspecified" prompts (inspired by early GPT-3 formulations) that provide little to no task instruction or label constraints. The authors hypothesize that these prompts cause LLMs to fail at the output generation stage—producing unstructured text or ignoring expected labels—rather than failing to understand the task internally. This leads to low logit values for relevant tokens and classification decisions based on randomly distributed probabilities, thereby inflating the perceived sensitivity of the models.

Methodology

The study systematically compares underspecified (minimal) prompts against well-specified (instruction) prompts across three text classification tasks (SST-2, AG News, MMLU) and six LLM variants (LLaMA-2-13B, LLaMA-3.1-8B, LLaMA-3.2-3B, in both base and instruction-tuned forms).

The evaluation framework employs three distinct analytical approaches:

Empirical Performance Analysis: The authors evaluate mean accuracy and standard deviation across 10 prompt variants per setting. They compare a "Base" setting against several mitigation strategies:
- UNK Strategy: Appending <unk> tokens to the prompt.
- Calibration: Adjusting label probabilities based on neutral prompts to ensure uniform distribution.
- In-Context Learning (ICL): Providing 2 examples per class within the prompt.
- Instruction-Tuning: Using instruction-tuned model variants.
- Meta-Tags/Chat Templates: Utilizing tokenizer-specific delimiters.
- Evaluation is conducted via both Logit Evaluation (selecting the class with the highest probability from the final token) and Generation Evaluation (extracting labels from free-text generation).
Logit Analysis: The authors analyze the magnitude and distribution of logit values for class labels. They investigate the correlation between logit values and classification accuracy, specifically comparing prompts that yield high versus low logit values.
Linear Probe Analysis: To investigate internal representations, the authors train linear Ridge classifiers on the activations of every layer of the LLM. This mechanistic interpretability technique determines whether the model's internal state contains sufficient information to solve the task, even if the final output generation fails.

Key Results

1. The Impact of Prompt Underspecification

Underspecified prompts consistently exhibit higher performance variance and significantly lower logit values for relevant tokens compared to instruction prompts. In many cases, minimal prompts result in label logits in the range of $10^{-5}$ to $10^{-3}$ , placing the correct labels in the lower half of the vocabulary probability distribution. Consequently, classification decisions under minimal prompts are often driven by noise rather than semantic understanding.

2. Mitigation Strategies

Instruction Prompts: While instruction prompts generally improve mean performance and reduce variance, they do not eliminate sensitivity entirely on their own.
In-Context Learning (ICL): ICL provides the most consistent and significant mitigation of prompt sensitivity. When combined with instruction-tuned models, ICL yields the highest accuracy and lowest standard deviation. Notably, ICL addresses the issues of underspecification as effectively as calibration but without requiring access to model internals or neutral prompts.
Calibration: Calibration offers improvements similar to ICL but is slightly less consistent. Combining calibration with ICL yields only marginal additional benefits, suggesting both methods target the same underlying issue.
UNK Strategy: The strategy of appending <unk> tokens often degrades performance significantly, sometimes resulting in performance worse than random chance.
Model Variants: Switching to instruction-tuned variants of the models significantly improves performance for both minimal and instruction prompts, though this may be partially attributed to data leakage if the training data overlaps with the evaluation sets.

3. Logit and Linear Probe Insights

Logit Correlation: There is a strong positive correlation (up to $r=0.757$ ) between the magnitude of label logit values and classification accuracy. High logit values serve as a reliable indicator of prompt quality and the absence of underspecification.
Internal Representations: Linear probe analysis reveals that internal representations (activations in middle layers) remain robust and contain sufficient information to classify tasks accurately, even when the final output is incorrect due to underspecification. The degradation in performance occurs primarily in the final layers of the network. This suggests that the model "knows" the answer internally but fails to articulate it correctly when the prompt is underspecified.

Contributions and Significance

The paper makes the following primary contributions:

Reframing Prompt Sensitivity: It identifies prompt underspecification as a primary driver of observed LLM sensitivity, arguing that models often fail at output formatting rather than task comprehension.
Methodological Rigor: It advocates for the use of well-specified instruction prompts and in-context learning to ensure reliable assessments of model capabilities, moving away from minimal prompts that induce high variance.
Diagnostic Tools:
- It establishes logit analysis as a practical tool for diagnosing prompt quality, showing that low logit values are a strong predictor of poor performance and high sensitivity.
- It utilizes linear probing to demonstrate that sensitivity issues often emerge in the final layers of the model, while internal representations remain stable.
Practical Mitigation: It demonstrates that simple modifications, specifically the combination of instruction-tuned models and in-context learning, can effectively mitigate the majority of sensitivity issues without complex calibration procedures.

Conclusion

The authors conclude that a significant portion of LLM instability in text classification stems from the use of underspecified prompts and ambiguous experimental setups rather than a fundamental inability of the models to solve the tasks. They advocate for greater methodological rigor in future research, emphasizing the necessity of well-specified prompts to distinguish between genuine model limitations and artifacts of poor prompting. The findings suggest that with appropriate prompting strategies (specifically instruction formats and ICL), the observed sensitivity can be largely mitigated, leading to more reliable and reproducible evaluations of LLMs.