💬 NLP

Clustered Self-Assessment: A Simple yet Effective Method for Uncertainty Quantification in Large Language Models

The paper proposes "Clustered Self-Assessment," a simple and efficient method that improves uncertainty quantification in large language models by clustering sampled generations into semantic options for self-evaluation, consistently outperforming existing baselines even with minimal additional sampling.

Original authors: Qi Cao, Takeshi Kojima, Andrew Gambardella, Helinyi Peng, Yutaka Matsuo, Yusuke Iwasawa

Published 2026-06-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Qi Cao, Takeshi Kojima, Andrew Gambardella, Helinyi Peng, Yutaka Matsuo, Yusuke Iwasawa

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 asking a very smart, well-read friend (a Large Language Model, or LLM) a question. They answer you instantly with perfect grammar and total confidence. But here's the catch: sometimes, they are confidently wrong. They might tell you the Eiffel Tower is in Berlin, sounding just as sure as if they told you it's in Paris.

The problem is, you have no way of knowing if they are telling the truth or just "hallucinating" a plausible-sounding lie.

This paper introduces a simple trick called Clustered Self-Assessment to help you figure out how much you should trust your AI friend's answer.

The Problem: The "Confident Liar"

Current AI models are great at talking, but they are terrible at admitting when they don't know something. If you ask them, "How sure are you?" they will often say, "100% sure!" even when they are wrong. Other methods try to guess their confidence by looking at how many different answers the AI gives if you ask it the same question ten times, but these methods are often hard to understand and don't use the AI's own "brain" effectively.

The Solution: The "Multiple-Choice Test"

The authors propose a clever two-step game to make the AI check its own work.

Step 1: The "Brainstorming" Session
First, you ask the AI the same question a few times (just a couple of extra times is enough).

Analogy: Imagine asking your friend, "Where is the Eiffel Tower?" five times in a row. They might say "Paris," "Paris, France," "The capital of France," "Berlin," and "Rome."

Step 2: The "Grouping" Game
Next, you take those answers and group the ones that mean the same thing.

Analogy: You realize that "Paris," "Paris, France," and "The capital of France" are all saying the same thing. You put them in one pile. "Berlin" and "Rome" are different, so they get their own piles. Now you have three distinct groups: The Paris Group, The Berlin Group, and The Rome Group.

Step 3: The "Quiz Show"
Now, you turn this back on the AI. You present the original question to it, but this time, you give it a multiple-choice quiz where the options are those groups you just made.

The Question: "Where is the Eiffel Tower?"
The Options:
- A) The Paris Group
- B) The Berlin Group
- C) The Rome Group
- D) None of the above

Step 4: The "Confidence Score"
Finally, you ask the AI to pick an answer. The magic happens here: you look at the mathematical probability the AI assigns to the option it picks.

If the AI picks "A" (Paris) and gives it a 95% probability, it's saying, "I am very confident this is the right group."
If it picks "A" but only gives it a 40% probability (and is split between A, B, and C), it's saying, "I'm not really sure which group is right."

Why This is a Big Deal

The paper claims this method is a winner for three main reasons:

It's Fast and Cheap: You don't need to ask the AI 16 times to get a good result. Just two extra samples (asking the question twice more) are enough to get results that beat other methods that ask 16 times. It's like getting a great answer with a quick coffee break instead of a long lunch.
It's Easy to Understand: Instead of giving you a confusing math score like "Entropy: 0.84," it gives you a simple percentage (e.g., "85% confidence"). This is something a regular human can actually use to decide if they should trust the answer.
It Works Better: When the researchers tested this on different AI models and different types of questions (from trivia to summarizing news), their method consistently guessed the AI's confidence more accurately than all the other popular methods.

The Catch (Limitations)

The authors are honest about the flaws:

The "Black Box" Issue: To do this, you need to see the AI's internal math (its "logits"). If you are using a closed AI (like a paid API where you can't see the inner workings), you can't use this method.
The "Referee" Cost: To group the answers in Step 2, the method uses a separate, smaller AI model to act as a referee. This adds a tiny bit of extra work.
No "Calibration" Step: The method uses the raw numbers the AI gives. While they work well, the authors admit that adding a final "tuning" step could make them even more perfect.

Summary

In short, this paper suggests that instead of guessing how sure an AI is, we should ask the AI to take a multiple-choice test based on its own previous answers. By seeing how strongly the AI picks one option over the others, we get a clear, simple, and highly accurate "confidence meter" that tells us when to trust the AI and when to double-check its work.

Technical Summary: Clustered Self-Assessment for Uncertainty Quantification in LLMs

Problem Statement

Large Language Models (LLMs) exhibit remarkable performance across diverse tasks but frequently generate responses that appear fluent and plausible while being factually incorrect. This "hallucination" problem is particularly critical in high-stakes domains such as healthcare, law, and scientific research. A significant barrier to mitigating these risks is the lack of explicit, reliable uncertainty estimates. Existing methods for quantifying uncertainty often rely on indirect signals, such as entropy across sampled generations or semantic divergence metrics. These approaches have two primary limitations:

Interpretability: Indirect scores (e.g., entropy-based metrics) are difficult for end-users to interpret, limiting practical utility.
Underutilization of Internal Knowledge: Many methods measure output discrepancies without leveraging the LLM's internal understanding of its own knowledge. Furthermore, attempts to have LLMs verbalize uncertainty directly often fail due to the well-documented issue of model overconfidence.

Methodology: Clustered Self-Assessment

The authors propose Clustered Self-Assessment, a two-stage method designed to elicit and quantify uncertainty by prompting the LLM to evaluate its own generated answers. The process transforms the uncertainty estimation problem into a structured multiple-choice question (MCQ) task.

Stage 1: Answer Clustering

For a given input question, the LLM generates multiple responses (samples). To reduce semantic redundancy while preserving diversity, these samples are grouped into semantically distinct clusters:

Semantic Relationship Estimation: An external Natural Language Inference (NLI) model (specifically deberta-large-mnli) is used to estimate pairwise semantic relationships between sampled answers. For each pair $(a_i, a_j)$ , the NLI model predicts labels such as entailment, neutral, or contradiction.
Clustering Logic: The method employs a rule-based filter on bidirectional NLI labels. Two answers are grouped into the same cluster if their bidirectional relationship contains no contradiction and includes at least one entailment in either direction.
Representative Selection: Each cluster is represented by a single answer (typically the first assigned to the cluster). This process ensures that the final set of options represents semantically distinct candidate answers rather than minor variations of the same response.

Stage 2: Self-Assessment via MCQ

The clustered answers are used to construct a structured MCQ:

Question Formulation: The original question is presented alongside the representative answers from each cluster as distinct choices (e.g., Option A, Option B, Option C).
Safety Option: An additional option, "None of the above," is included to account for scenarios where all generated clusters are incorrect or the model is highly uncertain.
Confidence Scoring: The constructed MCQ is fed back into the original LLM. The model's token probability assigned to the specific choice corresponding to its primary (greedy-decoded) answer serves as the confidence score.
- Formally, if $c_{i^*}$ is the token label for the correct choice, the confidence score $S$ is $P(c_{i^*})$ .

Key Contributions

Eliciting Self-Assessment: The method leverages the internal uncertainty representations of LLMs by structuring sampled generations into an MCQ format. This prompts the model to explicitly evaluate the correctness of its own answers rather than relying on indirect statistical proxies.
Superior Performance and Efficiency: The approach consistently outperforms existing baselines (including Semantic Entropy, EigenScore, and SAR) across multiple models and datasets. Notably, it achieves competitive performance with as few as two additional samples, whereas some baselines require up to sixteen, significantly reducing computational overhead.
Human-Interpretability: By using the token probability of a specific choice as the confidence score, the method provides a metric that is intuitive for users to interpret, directly mapping to the likelihood of the answer being correct.

Experimental Results

The method was evaluated across seven state-of-the-art open-source models (Qwen-2.5, Qwen-3, and Gemma-3 series) and three datasets: TriviaQA (TQA), Natural Questions (NQ), and XSum (summarization).

Accuracy (AUROC): The method achieved the highest Area Under the Receiver Operating Characteristic Curve (AUROC) across all models and datasets. For instance, on TQA with Qwen2.5-32B, it achieved an AUROC of 0.940, outperforming the next best baseline (Deg) which scored 0.848.
Sample Efficiency: Performance remained robust even with minimal sampling. With only two additional samples, the method outperformed baselines that used eight or sixteen samples.
Calibration: Using the Brier score, the method demonstrated superior calibration compared to probability-based baselines (P(True) and raw Probability), indicating that the confidence scores align well with empirical correctness.
Probe Training: When used as a training signal for a probe to predict uncertainty from LLM hidden states, the method's scores enabled the probe to outperform baselines in out-of-distribution settings without requiring correctness labels during inference.

Significance and Limitations

The paper posits that Clustered Self-Assessment offers a practical, effective, and computationally efficient solution for uncertainty quantification. By converting the problem into a self-evaluation task, it bridges the gap between the model's internal knowledge and user-interpretable confidence metrics.

The authors acknowledge specific limitations:

Logit Access: The method requires access to output logits, which may not be available for closed-source models.
External Dependency: It relies on an external NLI model for clustering, which introduces computational overhead and potential semantic misalignment. The authors suggest future work could leverage the LLM's own internal representations for clustering.
Calibration: The method uses raw softmax probabilities without post-hoc calibration, though it already demonstrates strong performance.

In conclusion, the paper presents a simple yet robust framework that enhances the reliability of LLM outputs by enabling models to assess their own uncertainty through structured self-reflection.