💬 NLP

Do Large Language Models Plan Answer Positions? Position Bias in Multiple-Choice Question Generation

This paper reveals that large language models exhibit systematic position biases when generating multiple-choice questions due to implicit planning of answer positions within their hidden representations, and demonstrates that activation steering can effectively manipulate these internal states to control and correct such biases.

Original authors: Xuemei Tang, Xufeng Duan, Zhenguang G. Cai

Published 2026-05-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xuemei Tang, Xufeng Duan, Zhenguang G. Cai

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 hiring a robot chef to bake a batch of "Multiple Choice Question" cakes. You tell the robot, "Make 100 cakes, and make sure the 'Correct Answer' ingredient is hidden randomly in slot A, B, C, or D, just like a fair game."

You expect the robot to be perfectly random. But when you taste the cakes, you notice a pattern: Robot A always hides the correct answer in slot A. Robot B always hides it in slot B or C. Robot C loves slot A so much it forgets to use slot D entirely.

This paper, "Do Large Language Models Plan Answer Positions?", is a investigation into why these robot chefs (Large Language Models, or LLMs) have such strong, predictable habits when creating quizzes.

Here is the breakdown of their findings, using simple analogies:

1. The "Hidden Habit" (The Problem)

The researchers tested 10 different robot chefs (LLMs) and 5 robot chefs that can also see pictures (Vision-Language Models). They asked them to create quizzes in three different kitchens:

Knowledge Kitchen: General facts (like history or math).
Reading Kitchen: Questions based on news articles.
Visual Kitchen: Questions based on images.

The Finding: The robots were not random. They had "position bias."

Some models (like Llama) were obsessed with the first slot (A).
Others (like Gemini) loved the middle slots (B and C).
Almost everyone ignored the last slot (D).

It's as if every chef has a secret "favorite shelf" where they instinctively put the winning cookie, regardless of what the recipe says.

2. The "Crystal Ball" (The Investigation)

The big question was: Is the robot just guessing based on the letters A, B, C, D, or is it actually "planning" the answer before it even starts writing the options?

To find out, the researchers looked inside the robot's brain (its internal code) while it was writing the question part of the quiz, before it even wrote the answers.

The Finding: They found a "crystal ball" inside the robot's brain.
Even while the robot was just writing the question stem (e.g., "What is the capital of France?"), its internal code had already "decided" where the correct answer would go. It was like the robot was whispering to itself, "I'm going to put the answer in slot B," before it even wrote the word "B."

This suggests the bias isn't just a surface-level mistake; it's a deep-seated implicit plan formed early in the thinking process.

3. The "Remote Control" (The Solution)

Since they found this "plan" inside the robot's brain, the researchers tried to use a "remote control" to change the robot's mind. They used a technique called Activation Steering.

Think of the robot's brain as a river flowing toward a specific destination (e.g., Slot A). The researchers found a way to build a small dam or a paddle to nudge the river slightly toward a different destination (e.g., Slot B).

The Finding:

It worked! By nudging the internal code, they could successfully force the robot to pick a different slot more often.
Timing matters: They found that nudging the robot just before it finished the question (the "penultimate" moment) worked much better than nudging it at the very end. It's like steering a car is easier before you hit the final turn, not after you've already crashed into the wall.
The Trade-off: While they could change the slot, sometimes the robot got confused. It might pick the right slot but change the meaning of the question or make the answer wrong. It's like forcing the robot to bake a chocolate cake in a vanilla mold; the shape changes, but the flavor might get weird.

4. The "Label Trap" (A Surprising Twist)

The researchers also tested what happens if you tell the robot to write the answers without the letters A, B, C, D. You'd think this would make it more random.

The Finding: It actually made the bias worse. Without the letters to guide it, the robots became even more obsessed with putting the correct answer in the very first position they generated. It turns out the letters A, B, C, D actually help the robots be more balanced, acting like guardrails that keep them from falling too far to the left.

Summary

LLMs are not random: They have strong, predictable habits about where to put correct answers in quizzes.
They plan ahead: They decide the answer's position early in the process, before writing the options.
We can nudge them: We can use internal "steering" to change these habits, but it's a delicate operation that can sometimes break the logic of the question.
Labels help: Surprisingly, giving the options labels (A, B, C, D) helps reduce the bias, while removing them makes the robots even more stubborn about picking the first option.

The Bottom Line: If you use AI to generate tests, you can't just trust it to be fair. It has its own hidden preferences, and if you want a truly random quiz, you need to actively intervene or use specific prompts to force it to break its habits.

Technical Summary: Do Large Language Models Plan Answer Positions?

Problem Statement

Large Language Models (LLMs) and Vision-Language Models (VLMs) are increasingly utilized to generate Multiple-Choice Questions (MCQs) for educational assessment, reading comprehension, and visual question answering. Ideally, the correct answers in generated MCQs should be uniformly distributed across options (A, B, C, D) to ensure unbiased evaluation. However, empirical observations indicate that these models exhibit systematic position biases, disproportionately favoring specific answer positions (e.g., a "primacy bias" toward option A or a "center bias" toward B/C).

While prior research has established that LLMs suffer from selection bias during the inference stage (favoring specific option identifiers when answering), it remains unclear whether similar biases arise during the generation process itself. Specifically, it is unknown if LLMs exhibit consistent positional preferences when constructing MCQs and whether these preferences are driven by internal representations that implicitly "plan" the answer position before generating the options.

Methodology

The authors propose a framework consisting of three components to investigate position bias, its underlying mechanisms, and potential controllability:

Controlled MCQ Generation Tasks:
The study evaluates 10 LLMs and 5 VLMs across three distinct generation scenarios:
- Knowledge-based MCQ (K-MCQ): Generating questions across five disciplines (Math, Chemistry, History, Geography, Biology).
- Reading Comprehension MCQ (RC-MCQ): Generating questions based on news articles.
- Vision-Language MCQ (VL-MCQ): Generating questions based on abstract images derived from COCO dataset descriptions.
  The authors standardize the output format (options A–D) and explicitly specify the correct answer to ensure observed biases stem from the model rather than task artifacts.
Probing for Implicit Planning:
To determine if answer positions are encoded during generation, the authors train lightweight classifiers (MLP probes) on token-level hidden representations of the question stem. The goal is to predict the eventual correct answer position ( $y \in \{1, \dots, K\}$ ) from the stem's hidden states ( $H^{(l)}_x$ ) before the options are generated. This tests the hypothesis that the model performs implicit forward planning.
Activation Steering (Intervention):
To establish a causal link between internal representations and answer position, the authors apply activation steering at decoding time. They construct steering vectors using two methods:
- Mean-Difference: Calculating the vector difference between mean hidden states associated with a source position (e.g., A) and a target position (e.g., B).
- Classifier-Based: Using the weight vector of a logistic regression classifier trained to distinguish between source and target positions.
  These vectors are added to hidden states at specific layers and token positions (e.g., the penultimate token of the stem) to manipulate the generation trajectory.

Key Results

1. Systematic Position Bias

Across all three tasks and diverse model families (including Llama, Qwen, DeepSeek, and Gemini), the models exhibit structured position biases.

Model-Specific Patterns: Llama models consistently favor option "A" (primacy bias), while Qwen and DeepSeek variants often prefer "B", and Gemini models show a tendency toward "B" and "C".
Underrepresentation: Option "D" is systematically underrepresented across nearly all models.
Identifier Influence: Removing explicit option identifiers (A/B/C/D) does not eliminate bias; instead, it often exacerbates the preference for the first position, suggesting that while identifiers provide cues, the bias is also driven by intrinsic ordering preferences.

2. Evidence of Implicit Planning

Probing experiments reveal that hidden representations in the question stem encode predictive signals of the correct answer position.

Predictability: Classifiers trained on stem representations achieve macro-F1 scores significantly above random and majority baselines, indicating that answer position information is encoded early in the generation process.
Layer-wise Emergence: Predictive signals are distributed across many layers rather than confined to the final layers. In some models (e.g., Qwen3-30B-Thinking), strong signals emerge in intermediate layers (24–36).
Token Position: Predictive performance is generally higher for later tokens in the stem, particularly the final token, suggesting that positional information becomes more accessible as contextual integration deepens.

3. Controllability via Activation Steering

Intervention experiments demonstrate that answer positions can be causally manipulated.

Effectiveness: Steering hidden states from a source position (A) to a target position (B) successfully shifts the distribution of generated answers.
Optimal Injection: Interventions applied at the penultimate token of the question stem yield stronger and more stable effects than those at the final token (often a punctuation mark).
Layer Sensitivity: The Mean-Difference method shows peak effectiveness in intermediate layers (e.g., layers 11–15), achieving shifts of over 22% (A→B).
Trade-offs: While steering can shift the target position, it introduces a trade-off. Interventions can lead to "off-target" shifts (e.g., A→C/D) and, in some cases, semantic drift where the question semantics or option space are altered to accommodate the new position.

4. Mitigation Strategies

Prompting: Adding explicit instructions to "balance" the distribution of correct answers reduces bias, particularly in larger or reasoning-oriented models (e.g., DeepSeek-Reasoner, Gemini-3-Pro). However, smaller models struggle to follow these global distributional constraints, and bias is not fully eliminated.

Significance and Claims

The paper claims to provide a practical framework for studying implicit positional planning in LLMs. Its primary contributions are:

Empirical Evidence: Demonstrating that position bias in MCQ generation is systematic, model-specific, and persists even without explicit option identifiers.
Mechanistic Insight: Providing evidence that answer positions are not merely incidental but are likely part of an implicit planning strategy encoded in the model's hidden representations. The ability to predict the answer position from the question stem suggests the model "decides" the position before generating the options.
Causal Manipulation: Establishing that internal representations can be steered to influence answer positions, confirming a causal link between latent planning signals and output distributions.
Reliability Implications: Highlighting that reliable MCQ construction requires more than just prompt engineering; it necessitates understanding and potentially controlling the internal dynamics of generation to ensure unbiased evaluation tools.

The authors maintain a modest stance, noting that while their findings suggest "planning-like mechanisms," they do not constitute definitive proof of explicit symbolic planning. They also acknowledge limitations, such as the dependence of intervention effectiveness on injection positions and the potential for semantic incoherence when forcing positional changes.