💻 computer science

Auditing Multimodal LLM Raters: Central Tendency Bias in Clinical Ordinal Scoring

This paper reveals that while frontier multimodal LLMs can competitively score clinical images on ordinal scales, they exhibit a systematic central tendency bias that compresses predictions toward the middle of the scale, leading to critical errors at the high and low extremes that necessitate calibration before deployment in high-stakes clinical screening.

Original authors: Jiaqing Zhang, Sandeep Elluri, Bhanu Cherukuvada, Yonah Joffe, Jessica Sena, Miguel Contreras, Scott Siegel, Subhash Nerella, Catherine Price, Parisa Rashidi

Published 2026-05-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jiaqing Zhang, Sandeep Elluri, Bhanu Cherukuvada, Yonah Joffe, Jessica Sena, Miguel Contreras, Scott Siegel, Subhash Nerella, Catherine Price, Parisa Rashidi

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 have a new, incredibly smart robot assistant. You want to use it to grade students' drawings of clocks—a common test used by doctors to check for memory or thinking problems. The test is simple: draw a clock showing 11:10. Doctors give these drawings a score from 0 (a complete mess) to 5 (perfect).

The researchers in this paper asked: Can this robot assistant grade these drawings as well as a human doctor or a specialized computer program?

Here is what they found, explained simply:

1. The Robot is "Too Safe"

The robot (which is a type of advanced AI called a Multimodal Large Language Model) is very good at looking at a picture and understanding what it sees. However, when it comes to giving a score, it has a weird habit: it refuses to be extreme.

Think of the scoring scale like a volume knob on a radio, going from 0 (muted) to 5 (max volume).

The Human/Computer Expert: If a drawing is terrible, they confidently turn the knob to 0. If it's perfect, they turn it to 5.
The Robot: Even when the drawing is terrible, the robot hesitates to give it a 0. It thinks, "Maybe it's not that bad," and gives it a 1. When the drawing is perfect, it hesitates to give a 5, thinking, "Maybe it's not quite perfect," and gives it a 4.

The robot keeps pulling the score toward the middle (2 or 3). It's like a teacher who is afraid to give anyone an "A" or an "F," so they just give everyone a "B" or a "C" to be safe.

2. The "Average" Score Hides the Problem

If you just look at the robot's overall average score, it looks pretty good. It's often "close enough" to the human score (within one point).

The researchers compared this to a weather forecast. If a forecast says "It will be 70 degrees" every day, and the actual temperature is sometimes 60 and sometimes 80, the average error is small. But if you need to know if it's going to rain (a specific extreme), that forecast is useless.

Similarly, the robot's "safe" middle scores look good on paper, but they fail the most important cases:

The Danger: If a patient has a very bad drawing (score 0 or 1), the robot might say it's a 2. This could mean a sick person is missed and doesn't get the help they need.
The Confusion: If a patient has a perfect drawing (score 5), the robot might say it's a 4. This could cause unnecessary worry or extra tests for a healthy person.

3. Trying to Fix the Robot Didn't Work

The researchers tried to "teach" the robot to be more confident by showing it examples of bad and good drawings before it started grading (this is called "few-shot prompting").

Result: It got slightly better, but it still refused to give the extreme scores. It was still too cautious.

They also tried removing all the medical words from the instructions, asking the robot to just grade the "art" instead of the "medical test."

Result: This actually made the robot worse. It turns out the medical context helped the robot understand the task, but even with that help, it still wouldn't give the extreme scores.

4. The Specialized Computer Wins

The researchers also tested a specialized computer program (a Deep Learning model) that was trained specifically on thousands of these clock drawings.

Result: This specialized computer didn't have the "safety" problem. It confidently gave 0s to bad drawings and 5s to good ones. It was much more accurate at the extremes where it matters most.

The Bottom Line

The paper concludes that while these fancy AI robots are impressive and can get "close" to the right answer, they have a built-in bias toward the middle. They are like a nervous student who is afraid to raise their hand to say "I know the answer" or "I don't know the answer," so they just guess the middle ground.

Because of this, the researchers say we shouldn't use these robots alone to make final decisions on patient health screenings. They are too likely to miss the most critical cases (the very bad or the very good) because they are too afraid to be extreme.

Technical Summary: Auditing Multimodal LLM Raters for Clinical Ordinal Scoring

Problem Statement
Multimodal Large Language Models (LLMs) are increasingly deployed as automated evaluators ("LLM-as-a-judge") in clinical settings. However, their scoring behavior on ordinal clinical scales remains poorly understood. Standard evaluation metrics, such as Mean Absolute Error (MAE) and within-tolerance agreement, often rely on aggregate statistics that can conceal systematic failure modes. Specifically, a rater might achieve strong aggregate agreement while failing at the scale extremes. In clinical contexts like cognitive impairment screening, errors at these extremes (e.g., misclassifying a severely impaired patient as normal, or vice versa) carry disproportionate consequences for downstream screening decisions. The paper investigates whether LLMs exhibit systematic biases, specifically a "central tendency effect" (compression of scores toward the middle of the scale), when applied to clinical ordinal scoring tasks, and whether this bias can be mitigated through prompt engineering.

Methodology
The study benchmarks three frontier multimodal LLM families (GPT-5/GPT-5.4, Gemini-2.5-Pro, and Claude-4-Sonnet) against supervised deep learning models (CNNs and Vision Transformers) on the Clock Drawing Test (CDT).

Datasets: The primary evaluation uses 597 images from the National Health and Aging Trends Study (NHATS), partitioned into development and test sets. External validation is performed on an independent Thai CDT cohort (386 images).
Task: The task involves predicting an integer score $y \in \{0, \dots, 5\}$ based on the Shulman rubric, where 0 represents an unrecognizable drawing and 5 represents an accurate depiction.
Models:
- Supervised Baselines: A ResNet-101 CNN and two Vision Transformer (ViT) variants (Ordinal classification and Continuous regression). These are fully fine-tuned on NHATS data.
- LLM Raters: Zero-shot and few-shot multimodal LLMs. These models receive the image and a natural-language rubric describing the six score levels but undergo no gradient-based training on the target dataset.
Evaluation Protocol:
- Aggregate Metrics: MAE, RMSE, and within-1 accuracy (tolerance-based agreement).
- Per-Score Analysis: Examination of predicted-score distributions, directional error profiles, and confusion matrices to detect systematic compression toward the center.
- Prompt Ablations: To distinguish between prompt-engineering artifacts and intrinsic model behavior, the authors tested:
  1. Few-shot prompting: Providing 30 exemplar images (5 per score level) to anchor the model.
  2. De-clinicalized prompting: Removing medical terminology (e.g., "neuropsychology") to test if safety mechanisms trigger conservative scoring.

Key Results

Aggregate Performance: Fully fine-tuned ViT-Ordinal models achieved the best calibration (MAE 0.52, within-1 accuracy 91%). Among zero-shot LLMs, GPT-5 was the strongest performer (MAE 0.67, within-1 accuracy 92%), showing competitive tolerance-based agreement despite higher absolute error than the best supervised models.
Central Tendency Effect: Per-score analysis revealed a pronounced systematic bias in all LLM families. Predictions were compressed toward the middle of the scale:
- Low-end over-prediction: True scores of 0 were frequently predicted as 1.
- High-end under-prediction: True scores of 5 were frequently predicted as 4.
- This resulted in a calibration slope significantly shallower than that of supervised models, with errors concentrated in the (0→1) and (5→4) cells of the confusion matrix.
Ablation Findings:
- Few-shot prompting improved aggregate metrics (reducing MAE to 0.56) and increased accuracy at the high end, but did not eliminate the central tendency effect. A significant portion of extreme scores remained misclassified toward the center.
- De-clinicalized prompting worsened performance across all metrics and failed to reduce the bias, indicating that clinical terminology actually aids scoring fidelity rather than triggering excessive conservatism.
External Validation: The central tendency pattern (endpoint compression) was replicated on the independent Thai CDT cohort, confirming the robustness of the finding across different populations.

Key Contributions

Systematic Evaluation Protocol: The authors propose a protocol combining per-score error decomposition and calibration-slope analysis to detect failure modes that aggregate metrics conceal.
Identification of Central Tendency Bias: The study demonstrates that multimodal LLMs exhibit a consistent central tendency effect in clinical ordinal scoring, systematically avoiding extreme ratings (0 and 5) in favor of intermediate scores.
Isolation of Bias Source: Through targeted ablations, the paper isolates this bias as an intrinsic behavioral tendency of current multimodal LLMs rather than a result of insufficient prompting or specific clinical framing.

Significance and Claims
The paper argues that while frontier LLMs show promise as scalable proxies for expert annotation, their current scoring behavior is not yet suitable for standalone deployment in high-stakes clinical screening workflows where accurate identification of scale endpoints is critical. The central tendency effect poses a specific risk: it may lead to the under-detection of severe cognitive impairment (by predicting a 1 instead of a 0) or the over-referral of healthy individuals (by predicting a 4 instead of a 5).

The authors extend the literature on LLM-as-a-judge bias from NLP evaluation to clinical assessment. They hypothesize that this behavior may stem from Reinforcement Learning from Human Feedback (RLHF) alignment, which internalizes a human-like aversion to extreme ratings, or as a distributional prior in the model. The paper concludes that deploying LLM-based raters requires calibration-aware evaluation and suggests that post-hoc calibration or hybrid workflows (using LLMs for prescreening and supervised models/humans for final scoring) are necessary before these tools can be safely integrated into clinical practice.