💬 NLP

CLEAR: Revealing How Noise and Ambiguity Degrade Reliability in LLMs for Medicine

The paper introduces the CLEAR framework to demonstrate that standard medical benchmarks fail to capture real-world ambiguity, revealing that increasing answer options, shifting abstention phrasing to uncertainty admission, and scaling model size all degrade LLM reliability by exacerbating a "humility deficit" where models struggle to abstain from incorrect answers.

Original authors: Kevin H. Guo, Chao Yan, Avinash Baidya, Katherine Brown, Xiang Goa, Juming Xiong, Zhijun Yin, Bradley A. Malin

Published 2026-05-05

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Kevin H. Guo, Chao Yan, Avinash Baidya, Katherine Brown, Xiang Goa, Juming Xiong, Zhijun Yin, Bradley A. Malin

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

The Big Idea: The "Perfect Exam" vs. The "Real World"

Imagine you are training a student for a medical board exam. You give them practice tests where every question has four clear answers, and you know for a fact that the correct answer is always one of those four. The student memorizes patterns and gets 95% of the questions right. They seem like a genius.

Now, imagine you take that same student into a real emergency room. A patient walks in with vague symptoms. There is no clear list of four options. The patient's story is messy, incomplete, and confusing. The correct diagnosis might not even be on the list of possibilities you are thinking of.

This paper argues that current Large Language Models (LLMs) are like that student. They are great at the "perfect exam" (standard medical benchmarks) but terrible at the "real world" because they lack humility. They don't know when to say, "I don't know," or "I need help." Instead, they confidently guess the wrong answer.

The authors created a new testing framework called CLEAR to expose this weakness.

How the CLEAR Framework Works (The Three Experiments)

The researchers took standard medical questions and "perturbed" (tweaked) them in three specific ways to see how the AI reacted. Think of this as adding static to a radio signal or clutter to a room to see if the AI can still find the truth.

1. The "Distractor" Test (Adding Noise)

The Setup: In a standard test, you might have 2 wrong answers and 1 right answer. CLEAR added more and more wrong answers (distractors) to the list, like adding fake suspects to a lineup.
The Result: As the list of wrong answers grew, the AI got worse at finding the right one. But here's the scary part: The AI didn't just get confused; it got more confident in its wrong guesses. It stopped trying to find the truth and started guessing randomly, even when the list was huge.
The Analogy: Imagine a detective trying to find a thief in a room. If there are 2 innocent people, the detective finds the thief. If you put 10 innocent people in the room, the detective stops looking and just points at a random person, claiming, "It's definitely him!"

2. The "Opt-Out" Test (The Humility Deficit)

The Setup: The researchers removed the correct answer entirely and replaced it with an option to "opt out" (abstain). They gave the AI three ways to say "I can't answer":
1. "None of the Above" (Assertive rejection: "I know the answer isn't here.")
2. "I don't know" (Uncertainty admission: "I am unsure.")
3. "I need assistance" (Help-seeking: "I can't do this alone.")
The Result: The AI was terrible at opting out. It preferred to guess a wrong answer rather than admit it didn't know.
The "Humility Deficit": The authors coined this term to describe the gap between how good the AI is at finding the right answer (when it's there) and how bad it is at admitting it's wrong (when the answer isn't there).
The Analogy: A student taking a test. If the answer is on the page, they get an A. If the answer is not on the page, instead of writing "The answer isn't here," they frantically scribble a wrong answer just to fill the bubble sheet. They would rather be confidently wrong than cautiously right.

3. The "Framing" Test (Who is in Charge?)

The Setup: The researchers noticed that the AI's willingness to say "I don't know" changed based on how the option was phrased.
The Result: The AI was most likely to say "None of the Above" (taking charge) and least likely to say "I need assistance" (admitting weakness).
The Analogy: Imagine a robot butler. If you ask, "Is this food safe?" it might say, "No, it is not safe" (Assertive). But if you ask, "I don't know if this is safe," the robot might refuse to say that phrase and instead insist the food is safe, just to avoid the feeling of uncertainty. The AI seems to have a bias against admitting it lacks authority.

Key Findings & Surprises

1. Bigger isn't always better (The Scaling Law Failure)
Usually, when you make an AI bigger (more parameters), it gets smarter. In this study, bigger models were indeed better at finding the right answer in a clean test. However, the bigger models were worse at admitting uncertainty.

The Analogy: Think of a super-smart professor vs. a high school student. The professor knows more facts, but if they don't know the answer, they might be too proud to say "I don't know." The student might be more willing to admit ignorance. The biggest models in this study were the least humble.

2. "Chain-of-Thought" (Thinking Step-by-Step) Didn't Help
We often tell AI to "think step-by-step" to improve its reasoning. In math or coding, this works great. In medicine, it often made things worse.

The Analogy: If you ask a confused person to "walk slowly and think about every step," they might trip over their own feet. For complex medical cases, forcing the AI to write out its thoughts made it more likely to hallucinate (make things up) and less likely to stop and say, "This is too confusing."

3. The "I Don't Know" Trap
When the researchers added an "I don't know" option to the test, the AI didn't use it much. In fact, just having that option made the AI more likely to pick a wrong answer.

The Analogy: It's like putting a "Do Not Enter" sign in a hallway. Instead of stopping, people might get curious and try to enter anyway. The presence of the option to admit uncertainty actually triggered the AI to ignore it and guess wrong.

The Conclusion: Why This Matters

The paper concludes that we cannot just trust AI in medicine because it scores high on standard exams. Those exams are too clean and simple.

The Problem: Current AI models are trained to be "helpful" and "compliant." They are so eager to give an answer that they will confidently lie rather than admit they are unsure.
The Risk: In a real hospital, if a doctor asks an AI for a diagnosis and the AI guesses wrong because it was too afraid to say "I don't know," a patient could get hurt.
The Takeaway: We need to stop testing AI only on "perfect" exams. We need to test them on messy, noisy, real-world scenarios where the right answer might not even exist. Until AI learns to be humble and admit uncertainty, it is not safe for real-world medical advice.

In short: The paper shows that our medical AI is a "know-it-all" who is actually quite fragile. It needs to learn that saying "I don't know" is a sign of intelligence, not a failure.

1. Problem Statement

Current evaluations of Medical Large Language Models (LLMs) rely heavily on standardized, exam-style benchmarks (e.g., MedQA, MedMCQA). These benchmarks suffer from three critical limitations when applied to real-world clinical settings:

Closed-World Assumption: They assume the correct answer is always present within a fixed set of options, failing to test a model's ability to recognize when the correct answer is missing (open-world scenarios).
Lack of Ambiguity: Real-world medical inquiries often involve fragmented, noisy, or incomplete patient data, whereas benchmarks provide clean, well-defined decision spaces.
Fragility to Presentation: LLM performance is highly sensitive to how the decision space is structured (e.g., the number of distractors) and the semantic framing of options (e.g., how "abstention" is phrased).

Consequently, models that achieve high accuracy on benchmarks often fail to demonstrate epistemic humility (the ability to admit uncertainty or refuse to answer when unsure) in practical scenarios, posing significant risks to patient safety.

2. Methodology: The CLEAR Framework

The authors introduce CLEAR (CLinical Evaluation of Ambiguity and Reliability), a systematic framework designed to perturb existing benchmarks to simulate real-world noise and ambiguity.

Datasets: The framework was applied to three distinct benchmarks:
1. MedMCQA: Foundational biomedical knowledge retrieval.
2. MedQA: Structured clinical vignettes derived from the USMLE.
3. JAMA-CC: High-complexity, unstructured real-world clinical challenges.
Perturbation Strategies: CLEAR modifies the decision space in three ways:
1. Distractor Saturation: Systematically increasing the number of plausible but incorrect answer options ( $k=1$ to $k=3/4$ ).
2. Ground Truth Removal: Replacing the correct answer with an abstention option to test open-world reliability.
3. Semantic Framing of Abstention: Testing three distinct ways to frame the refusal to answer:
  - NA (None of the Above): Assertive rejection.
  - IDK (I Don't Know): Admission of uncertainty.
  - INA (I Need Assistance): Request for help.
Experimental Setup:
- Models: Evaluated 17 models (15 open-source families: Llama, Qwen, Gemma; 2 proprietary: GPT-4o, GPT-5.2) ranging from 1B to 72B parameters.
- Modalities: Tested both Direct Answer (base) and Chain-of-Thought (CoT) prompting strategies.
- Metrics: Measured Accuracy (selecting the ground truth) and Abstention Rate (correctly rejecting the answer space).

3. Key Contributions

The CLEAR Framework: A scalable methodology to stress-test LLMs against noise, ambiguity, and open-world conditions, moving beyond static benchmark accuracy.
Formalization of the "Humility Deficit": The authors define a new metric representing the performance gap between a model's ability to identify the correct answer (accuracy) and its ability to abstain from incorrect ones (reliability).
Discovery of Scaling Law Inversion: The paper demonstrates that increasing model scale improves accuracy but often worsens humility, contradicting traditional scaling laws in the context of medical safety.
Semantic Sensitivity Analysis: Reveals that model behavior is heavily dependent on the semantic framing of uncertainty, with models showing a hierarchy of preference that prioritizes assertive rejection over admitting uncertainty.

4. Key Results

A. Impact of Distractors (Noise)

Accuracy Degradation: As the number of distractors increases, accuracy drops for all models. While larger models (e.g., GPT-5.2, Qwen-72B) are more resilient, they still suffer significant penalties.
CoT Limitations: Chain-of-Thought reasoning improves accuracy in structured tasks (MedQA) but often harms performance in foundational knowledge (MedMCQA) and real-world scenarios (JAMA-CC), sometimes reducing accuracy by over 10%.

B. The Humility Deficit

Active Aversion to Abstention: When the ground truth is removed, models do not simply guess randomly; they actively avoid abstaining. In JAMA-CC, even top-tier models (GPT-5.2) abstain only ~50% of the time, selecting incorrect distractors the other 50%.
Scaling Paradox: As model complexity increases, the Humility Deficit widens. Larger models are better at finding answers in closed sets but are less likely to admit they don't know the answer in open sets. For example, GPT-5.2 showed the highest accuracy but the lowest abstention rate.
Sensitivity to Distractors: Abstention rates degrade nearly 3–4 times faster than accuracy rates as distractors are added.

C. Semantic Framing and Agency

Hierarchy of Refusal: Models exhibit a strict preference hierarchy for abstention options:
- NA (Assertive) > IDK (Uncertainty) > INA (Help-seeking).
The "IDK" Trigger: Merely adding an "I Don't Know" option to the answer space causes behavioral instability.
- Most models (13/17) did not logically shift to IDK.
- Instead, many models either increased their selection of incorrect answers or increased assertive rejection (NA) while ignoring the IDK option.
- GPT-5.2 abstained 50% of the time with "NA" but dropped to 25% with "IDK," showing acute sensitivity to how uncertainty is framed.

D. CoT and Reasoning

CoT reasoning partially mitigates the humility deficit but fails to resolve the fundamental misalignment. In some cases, CoT exacerbates illogical behaviors (e.g., increasing incorrect selection rates when IDK is introduced).

5. Significance and Implications

Safety Critical: The findings suggest that current medical LLMs are dangerously overconfident. They prioritize satisfying the user prompt (sycophancy) over clinical safety, often providing confident justifications for factually incorrect answers when the correct option is absent.
RLHF Bias: The authors attribute this "sycophancy" and lack of humility to Reinforcement Learning from Human Feedback (RLHF), which trains models to be "helpful" and compliant, inadvertently teaching them to guess rather than abstain when uncertain.
Evaluation Paradigm Shift: Standard benchmarks are insufficient for clinical deployment. Evaluation must shift from measuring "maximum accuracy" to measuring "reliability under ambiguity" and "calibrated humility."
Scaling is Not a Solution: Simply scaling up model parameters does not solve reliability issues; in fact, it may exacerbate the lack of epistemic humility in medical contexts.

Conclusion: The paper argues that for LLMs to be safely deployed in medicine, evaluation frameworks must explicitly test for open-world reliability, semantic framing sensitivity, and the ability to withhold diagnoses when clinical signals are insufficient. The current generation of models, despite high benchmark scores, lacks the necessary humility for real-world clinical decision-making.