💻 computer science

Credal Concept Bottleneck Models for Epistemic-Aleatoric Uncertainty Decomposition

The paper introduces CREDENCE, a Credal Ensemble Concept Bottleneck Model framework that decomposes concept-level uncertainty into epistemic and aleatoric components using probability intervals and ensemble disagreement, thereby enabling more actionable decision-making strategies such as automated routing, targeted data collection, and human review.

Original authors: Tanmoy Mukherjee, Thomas Bailleux, Pierre Marquis, Zied Bouraoui

Published 2026-04-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tanmoy Mukherjee, Thomas Bailleux, Pierre Marquis, Zied Bouraoui

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 a restaurant critic trying to decide if a review is positive or negative. Sometimes, the review is crystal clear: "The food was amazing!" You know exactly what to do. But other times, the review is tricky: "The waiter was efficient, I suppose."

Is that a compliment? A complaint? Or just someone being unsure?

Current AI models often give you a single number to answer this, like "60% chance this is positive." But that single number hides a big problem. It doesn't tell you why the AI is unsure. Is the AI confused because it hasn't seen enough examples like this in its training? Or is the sentence just so vague that even a human expert would argue about it?

This paper introduces a new system called CREDENCE (Credal Ensemble Concept Estimation) that solves this by acting like a team of detectives instead of a single detective.

The Problem: The "Confused" vs. The "Ambiguous"

The authors say current AI mixes up two very different types of uncertainty:

Epistemic Uncertainty (The "I haven't learned this yet" problem): The AI is confused because it lacks experience. If you showed it more examples, it could learn the answer.
Aleatoric Uncertainty (The "This is just vague" problem): The input itself is ambiguous. Even if the AI studied forever, humans would still disagree on the answer because the sentence is inherently unclear.

Mixing these up is dangerous. If the AI is confused, you should train it more. If the sentence is vague, you should ask a human to decide. But if you just see a "low confidence" score, you don't know which action to take.

The Solution: A Team of Detectives (CREDENCE)

CREDENCE changes how the AI thinks. Instead of asking one AI to guess the answer, it uses a team of five slightly different AI "heads" (think of them as five detectives with different magnifying glasses) to look at the same sentence.

Here is how it works, step-by-step:

1. The Team of Detectives (Ensemble Heads)
Imagine five detectives looking at the sentence "The waiter was efficient, I suppose."

Detective A (trained with a simple lens) thinks: "Maybe positive."
Detective B (trained with a detailed lens) thinks: "Maybe negative."
Detective C thinks: "It's a toss-up."

If all five detectives agree, the AI is confident. If they all disagree, the AI knows it is confused (Epistemic Uncertainty). This disagreement tells the system: "Hey, we need more training data to learn how to handle sentences like this."

2. The "Ambiguity Meter" (Aleatoric Head)
Separately, the system has a special sensor trained to spot when humans themselves can't agree. It looks at the sentence and asks: "Is this sentence so vague that even a group of people would argue about it?"
If the answer is yes, the system flags it as Aleatoric Uncertainty. This tells the system: "Don't waste time training the AI on this; just send it to a human reviewer."

3. The "Interval" Answer
Instead of giving a single number (like 60%), CREDENCE gives a range (like "between 40% and 80%").

A narrow range means the detectives agree (low uncertainty).
A wide range means they are fighting (high uncertainty).

What Can You Do With This?

Because CREDENCE separates "confusion" from "vague input," it can give you a specific instruction for what to do next. The paper tests this on tasks like reading restaurant reviews and spotting toxic comments.

They found four distinct situations, like a traffic light system for AI decisions:

🟢 TRUST (Low Confusion, Low Vague): The detectives agree, and the sentence is clear. Action: Let the AI handle it automatically.
🟡 DATA (High Confusion, Low Vague): The detectives are fighting, but the sentence is clear. This means the AI just hasn't seen enough examples. Action: Collect more training data to teach the AI.
🟠 REVIEW (Low Confusion, High Vague): The detectives agree, but the sentence is so weird or vague that humans would argue. Action: Send it to a human reviewer. The AI isn't "wrong"; the situation just requires human judgment.
🔴 ABSTAIN (High Confusion, High Vague): The detectives are fighting, and the sentence is a mess. Action: Don't guess at all. Decline the task or escalate to an expert.

The Results

The paper tested this on several datasets (like restaurant reviews and emotion detection). They found:

When the AI made a mistake, it was usually because the "detectives" were disagreeing (High Confusion).
When the AI flagged a sentence as "vague," it was almost always because real humans actually disagreed on how to label it (High Vague).
Most importantly, if you tried to fix the AI by correcting the "vague" sentences, you got much better results than trying to fix the "confused" ones. This suggests that the things humans find ambiguous are actually the most important parts of the decision-making process.

In a Nutshell

CREDENCE is like giving an AI a dashboard with two separate warning lights instead of one big "Warning" light. One light tells you, "We need to study more," and the other tells you, "This is too tricky for a machine, call a human." By separating these two problems, the system knows exactly how to react, making it safer and more useful for real-world decisions.

1. Problem Statement

In high-stakes Natural Language Processing (NLP) applications (e.g., healthcare, legal, finance), standard models output a single confidence score that conflates two distinct types of uncertainty:

Epistemic Uncertainty: Reducible uncertainty caused by model ignorance or underspecification (e.g., the model hasn't seen enough similar training examples).
Aleatoric Uncertainty: Irreducible uncertainty caused by inherent ambiguity or noise in the input data (e.g., human annotators genuinely disagree on the label).

Limitations of Current Approaches:

Concept Bottleneck Models (CBMs): While CBMs improve interpretability by predicting intermediate human-understandable concepts, they typically output point probabilities (single values). This collapses epistemic and aleatoric uncertainty, making it impossible to determine why a prediction is uncertain.
Actionability Gap: Without decomposition, practitioners cannot distinguish between cases where the model needs more training data (high epistemic) versus cases where human review is required due to ambiguous input (high aleatoric).

2. Methodology: CREDENCE

The authors propose CREDENCE (Credal Ensemble Concept Estimation), a framework that decomposes uncertainty at the concept level by construction.

Core Architectural Innovations

Credal Predictions (Intervals): Instead of a single probability $p_k$ , each concept $k$ is predicted as a probability interval $[p_k, \bar{p}_k]$ . This interval represents a Credal Set, capturing the range of plausible probabilities.
Structural Separation of Uncertainty: The architecture ensures epistemic and aleatoric signals come from distinct sources to prevent them from collapsing into a single score.
- Epistemic Uncertainty ( $U_{epi}$ ): Derived from ensemble disagreement. The model uses $H$ diverse lightweight heads (LoRA adapters) with varying ranks and dropout rates. The variance in their predictions quantifies model confusion.
- Aleatoric Uncertainty ( $U_{ale}$ ): Derived from a dedicated head trained specifically to predict annotator disagreement (ambiguity). This head learns to output a value proportional to the variance in human labels for a given concept.

Pipeline Stages

Encoder: A frozen pre-trained model (e.g., RoBERTa, LLMs) encodes the input text $x$ into a representation $h$ .
Ensemble Heads: $H$ diverse LoRA heads predict concept probabilities $\hat{p}_h^{(k)}$ .
Credal Aggregation:
- Interval bounds are formed: $C_k = [\min_h \hat{p}_h^{(k)}, \max_h \hat{p}_h^{(k)}]$ .
- Epistemic uncertainty is calculated as the variance across heads: $U_{epi}^{(k)} = \text{Var}_h[\hat{p}_h^{(k)}]$ .
Aleatoric Head: A separate head predicts $\sigma_\theta^{(k)}$ , trained to match the observed annotator disagreement rate (ground truth) during training.
Classification & Decision:
- Final labels are predicted using the mean of concept probabilities.
- Decision Rules: The system uses the decomposed signals to route samples:
  - Trust: Low Epi, Low Ale $\rightarrow$ Automate.
  - Data Collection: High Epi, Low Ale $\rightarrow$ Collect more training data.
  - Human Review: Low Epi, High Ale $\rightarrow$ Route to humans (input is ambiguous).
  - Abstain: High Epi, High Ale $\rightarrow$ Decline prediction.

3. Key Contributions

Credal CBM Formulation: A novel CBM variant that represents concept predictions as probability intervals rather than point estimates, enabling principled uncertainty quantification.
Structural Decomposition: A method that separates epistemic (model confusion) and aleatoric (data ambiguity) signals using distinct architectural components (ensemble variance vs. supervised ambiguity head), preventing signal collapse.
Actionable Prescriptive Guidance: Demonstrates that decomposing uncertainty allows for specific, optimized interventions (e.g., prioritizing data collection for epistemic errors vs. human review for aleatoric cases).

4. Experimental Results

The framework was evaluated on multiple NLP tasks (Sentiment, Emotion, Toxicity) using datasets like CEBaB, GoEmotions, HateXplain, and SST-2.

Decomposition Validity (RQ1 & RQ2):
- Epistemic Correlation: CREDENCE shows a strong positive correlation between epistemic uncertainty and prediction errors ( $\rho_{epi} \approx 0.287$ on CEBaB). When the model is wrong, ensemble heads disagree.
- Aleatoric Correlation: Aleatoric uncertainty strongly correlates with annotator disagreement ( $\rho_{ale} \approx 0.785$ on CEBaB), far outperforming baselines (e.g., CBM+Ensemble $\rho_{ale} \approx 0.356$ ).
- Separation: On SST-2 (which lacks annotator disagreement labels), aleatoric uncertainty correctly shows no correlation with errors, proving it measures ambiguity, not model failure.
Actionability (RQ3):
- Concept Intervention: Replacing concepts with high aleatoric uncertainty with ground-truth labels yields significantly higher accuracy gains ( $\Delta Acc \approx 18.9\%$ ) compared to replacing high epistemic concepts ( $\Delta Acc \approx 4.3\%$ ). This suggests that ambiguous concepts drive task outcomes more than model confusion.
- Routing Efficiency: The four-quadrant decision framework identifies "Human Review" cases (Low Epi, High Ale) where the model is actually correct but humans disagree, achieving the highest potential accuracy gain ( $85.7\%$ ) if routed correctly.
Efficiency: CREDENCE achieves these results with minimal overhead (using LoRA adapters) compared to full Deep Ensembles, maintaining competitive accuracy while providing rich uncertainty signals.

5. Significance and Impact

Bridging Theory and Practice: The paper moves beyond theoretical uncertainty quantification to provide prescriptive actions. It tells practitioners what to do based on the type of uncertainty (e.g., "collect data" vs. "send to human").
Handling Subjectivity: In NLP tasks where "ground truth" is often a composite of disputed human judgments (subjectivity), distinguishing between model ignorance and inherent ambiguity is critical. CREDENCE provides the first CBM framework to explicitly model this distinction.
Human-AI Collaboration: By identifying cases where human judgment is essential (high aleatoric) versus cases where the model simply needs more training (high epistemic), the framework optimizes the allocation of human resources and data collection budgets.
Robustness: The method generalizes across different model scales (from 66M parameter encoders to 8B parameter LLMs) and diverse NLP tasks.

In summary, CREDENCE transforms uncertainty from a single, opaque confidence score into a structured, actionable diagnostic tool, enabling more reliable and efficient deployment of AI systems in high-stakes environments.