💬 NLP

ABCD: All Biases Come Disguised

This paper introduces a bias-reduced evaluation protocol for multiple-choice question benchmarks that replaces standard option labels with uniform, unordered labels and utilizes sentence similarity for answer matching, thereby significantly improving evaluation robustness against answer permutations while maintaining model performance.

Original authors: Mateusz Nowak, Xavier Cadet, Peter Chin

Published 2026-02-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mateusz Nowak, Xavier Cadet, Peter Chin

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 taking a multiple-choice test, but instead of reading the questions and thinking about the answers, you just start guessing based on the format of the test.

Maybe you notice that the teacher always puts the correct answer in the "C" spot. Or maybe you see that in the practice examples at the top of the page, the answer is always "A," so you decide to just pick "A" for everything. You aren't actually solving the problem; you're just exploiting a pattern in the test itself.

This is exactly what happens with Large Language Models (LLMs)—the super-smart AI brains behind chatbots. They are incredibly good at finding these "cheat codes" in how we ask them questions, often giving the right answer for the wrong reasons.

This paper, titled "ABCD: All Biases Come Disguised," is like a group of detectives (researchers from Dartmouth College) who decided to expose these cheating habits and build a better, fairer test.

Here is the breakdown of their discovery and solution, using some everyday analogies:

1. The Problem: The "Magic 8-Ball" Effect

The researchers found that AI models are like students who have memorized the layout of the test paper rather than the subject matter.

The Position Bias: If the correct answer is always in the 3rd slot, the AI picks the 3rd slot every time, even if the question is nonsense.
The Label Bias: If the options are labeled A, B, C, D, the AI might prefer "A" because it looks nice, or "D" because it's at the end.
The "Few-Shot" Bias: This is the sneakiest one. If you give the AI a few examples at the start of the chat (like a practice round) where the answer is always "Blue," the AI will assume the next answer must also be "Blue," even if the question has nothing to do with color.

The Analogy: Imagine a waiter who knows you always order the "Special of the Day." If you walk in and ask for a recommendation, they don't ask what you like; they just say, "The Special," because that's what you usually pick. They aren't thinking; they are just following a pattern.

2. The Detective Tool: "NonsenseQA"

To prove that the AI was cheating, the researchers created a fake test called NonsenseQA.

What it is: A test made entirely of random words. The questions make no sense, and the answers are random words like "Purple" or "Toaster."
The Goal: Since the questions are nonsense, the AI should get them wrong 75% of the time (random guessing).
The Shocking Result: Under standard testing methods, some AIs got 95%+ of these nonsense questions right!
Why? They weren't reading the words. They were just looking at the few examples at the top of the page, seeing that the answer was usually the 3rd option, and picking the 3rd option every time. They were "gaming the system."

3. The Solution: The "M&D" Protocol

The researchers proposed a new way to test AI called Matched-and-Dashed (M&D). Think of this as changing the rules of the game so the cheat codes don't work.

Here is how they fixed the test:

Remove the Labels (The "Dashed" part): Instead of labeling options A, B, C, D, they just used dashes: -, -, -, -. This stops the AI from guessing "A" because it likes the letter A.
Force Full Sentences (The "Matched" part): Instead of letting the AI just say "A" or "C," they forced the AI to write out the whole answer in a sentence.
- Old way: AI says "C."
- New way: AI says "The answer is the red apple."
The Semantic Match: Since the AI wrote a sentence, the computer uses a "meaning detector" (a semantic similarity model) to check if the sentence the AI wrote matches the meaning of the correct answer.

The Analogy:

Old Test: You point to a picture and say "A." The teacher just checks if you pointed to the right letter.
New Test: You have to describe the picture out loud. The teacher listens to your description and checks if it matches the picture. You can't just point to the right spot; you have to actually know what the picture is.

4. The Results: A Fairer World

When they ran this new test on 13 different AI models:

The Cheaters Stopped: On the NonsenseQA test, the AIs that were previously getting 95% right dropped down to about 25% (which is just random guessing). This proved they were actually relying on the test format, not their intelligence.
The Real Thinkers Stayed: On real tests (like science or logic questions), the AIs didn't get much worse. They just became more stable.
Less Variance: Before, if you shuffled the order of the answers, the AI's score would jump up and down wildly. With the new method, the score stayed steady. It's like a basketball player who used to only score when the hoop was on the left side; now, they can score no matter where the hoop is.

Why This Matters

The authors argue that we need to stop praising AI for being "smart" when they are just good at pattern matching.

If a doctor uses an AI to help diagnose a patient, and that AI is just guessing based on the order of symptoms listed, that could be dangerous. This new method helps us separate the models that are actually reasoning from the ones that are just guessing the pattern.

In a nutshell: The paper teaches us that if an AI can answer a question perfectly just by looking at the formatting, it's not smart—it's just a trickster. By changing the rules of the test, we force the AI to show its true intelligence.

1. Problem Statement

Large Language Models (LLMs) are routinely evaluated using Multiple-Choice Question (MCQ) benchmarks (e.g., MMLU, GPQA, CommonsenseQA). However, standard evaluation protocols are susceptible to superficial artifacts that models exploit rather than relying on genuine reasoning. The authors identify three primary sources of bias:

Label Bias: Models exploit the semantic or ordering cues of option labels (e.g., "A", "B", "C", "D").
Position Bias: Models learn to select answers based on their position in the list (e.g., always choosing the third option).
Few-Shot Prompt Distribution Bias: Models analyze the distribution of correct answers in the few-shot examples (e.g., if 3 out of 4 examples have "A" as the answer, the model biases its prediction toward "A" for the test question).

Existing research has addressed position and label biases, but few-shot distribution bias remains largely unexplored. Furthermore, standard evaluation methods (Select-and-Letter) often reward shortcut exploitation, leading to inflated accuracy scores that do not reflect true reasoning capabilities.

2. Methodology

A. Diagnostic Tool: NonsenseQA

To quantify these biases, the authors introduce NonsenseQA, a synthetic dataset consisting of 1,000 questions generated from random words with randomly assigned correct answers.

Purpose: Since the content is semantically meaningless, any accuracy significantly above the chance level (25%) indicates the model is exploiting evaluation artifacts (labels, position, or few-shot patterns) rather than reasoning.
Findings: Under standard protocols, some models achieved >95% accuracy on NonsenseQA by explicitly or implicitly exploiting few-shot answer distributions.

B. Proposed Protocol: Matched-and-Dashed (M&D)

The authors propose a bias-reduced evaluation protocol designed to eliminate exploitable artifacts while preserving the MCQ structure:

Uniform, Unordered Labels: Instead of distinct letters (A, B, C, D), all options are labeled with a uniform symbol (e.g., "-"). This removes label-based ordering cues.
Full-Text Generation: Instead of asking the model to output a single letter, the model is instructed to generate the full text of the chosen answer.
Semantic Matching: The model's generated text is matched against the candidate options using a sentence similarity model (specifically Qwen3-Embedding-0.6B with cosine similarity). This allows for chain-of-thought reasoning and synonym variations.
Extraction: Regular expressions are used to extract the final answer sentence from the model's output.

C. Evaluation Metrics

SCORE: A metric measuring consistency across permutations.
Variance Ratio ( $\sigma^2_R$ ): Defined as $\sigma^2_{M\&D} / \sigma^2_{S\&L}$ . A value $< 1$ indicates the M&D protocol reduces performance variance (improving robustness) compared to the standard Select-and-Letter (S&L) protocol.
Answer-Moving Attacks: The authors permute the position of the ground-truth answer in both the test question and the few-shot prompt to test model sensitivity.

3. Key Contributions

Identification of Few-Shot Bias: The paper characterizes a previously underexplored bias where models exploit the distribution of correct answers in few-shot prompts, even on semantically meaningless inputs.
NonsenseQA Benchmark: A diagnostic tool to isolate and measure model reliance on shortcuts (labels, position, and few-shot patterns) without requiring ground-truth knowledge.
Bias-Reduced Evaluation Protocol (M&D): A practical, single-pass evaluation method that requires no model fine-tuning, no access to internal logits, and only a minimal increase in computational cost (~3%).
Comprehensive Empirical Analysis: Evaluation across 13 LLMs (ranging from 8B to 32B parameters) and 5 diverse benchmarks (MMLU-Pro, GPQA, CommonsenseQA, ARC, INCLUDE).

4. Results

A. Performance on NonsenseQA

Standard Protocol (S&L): Models exhibited high accuracy (often >95%) on random questions by exploiting few-shot patterns.
M&D Protocol: Median accuracy dropped sharply to near-chance levels (~25%) for most models, effectively suppressing bias exploitation. Models that previously relied on shortcuts were forced to behave randomly on nonsense data.

B. Robustness on Real Benchmarks

Reduced Variance: Across 13 models and 5 benchmarks, the M&D protocol reduced the variance of model accuracy under answer-moving attacks by a factor of 3x (Geometric mean variance ratio of 0.33).
Minimal Performance Drop: The reduction in bias came with only a minimal decrease in mean accuracy on standard benchmarks. For example, on ARC and CommonsenseQA, the drop was negligible, while robustness improved significantly.
Cross-Lingual Generalization: The protocol improved robustness in multilingual settings (INCLUDE dataset), proving that label and prediction-mode artifacts are not language-specific.

C. Model Categorization

The study revealed three distinct model behaviors regarding bias:

Explicit Bias Models: Directly reference few-shot answers in their reasoning (e.g., GPT-OSS).
Implicit Bias Models: Do not explicitly reference examples but consistently achieve >50% accuracy on nonsense data by inferring distribution patterns (e.g., Qwen3-8B).
Bias-Resistant Models: Only marginally exploit biases, performing near chance levels even under standard protocols (e.g., Gemma-3-27b-it).

D. Cross-Benchmark Agreement

The M&D protocol improved the correlation between different benchmarks (e.g., GPQA and ARC), suggesting that removing evaluation artifacts reveals the true underlying reasoning capabilities of models. Conversely, it highlighted that high performance on knowledge-heavy benchmarks (MMLU-Pro) does not always correlate with reasoning ability when biases are removed.

5. Significance and Impact

More Reliable Evaluation: The M&D protocol provides a more faithful assessment of an LLM's reasoning and knowledge retrieval capabilities by filtering out "shortcut" behaviors.
Practicality: Unlike previous debiasing methods that require fine-tuning or access to model internals (logits), this approach is a simple prompt engineering and post-processing change, making it immediately applicable to existing benchmarks.
Safety and Trust: By exposing models that rely on few-shot distribution biases, the protocol helps practitioners identify models that may fail in high-stakes, real-world scenarios where such patterns do not exist.
Future Research: The paper sets a new standard for MCQ evaluation, encouraging the community to move away from simple letter-selection metrics toward semantic matching and full-text generation to ensure robustness against evaluation artifacts.

In conclusion, the paper demonstrates that current MCQ benchmarks are significantly "gamed" by LLMs exploiting superficial cues. The proposed Matched-and-Dashed (M&D) protocol effectively neutralizes these biases, offering a more rigorous and stable framework for evaluating the true capabilities of large language models.