🤖 AI

The Dual Role of Abstracting over the Irrelevant in Symbolic Explanations: Cognitive Effort vs. Understanding

This paper demonstrates that applying formal abstractions to symbolic AI explanations—specifically by clustering irrelevant details to improve understanding and removing them to reduce cognitive effort—significantly enhances human reasoning performance and interpretability within an Answer Set Programming framework.

Original authors: Zeynep G. Saribatur, Johannes Langer, Ute Schmid

Published 2026-04-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zeynep G. Saribatur, Johannes Langer, Ute Schmid

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 trying to teach a friend how to solve a complex puzzle. You could give them the entire instruction manual, page by page, including every tiny detail about the history of the puzzle pieces, the factory where they were made, and the specific shade of blue on the box. Or, you could give them a simplified cheat sheet that highlights only the moves that actually matter to win the game.

This paper is about finding the perfect balance between those two extremes when teaching computers (specifically, "Symbolic AI") to explain their decisions to humans.

Here is the breakdown of what the researchers did and what they found, using simple analogies:

The Problem: Too Much Noise

Modern AI often works like a "black box"—it gives an answer, but we don't know how it got there. Symbolic AI is different; it's like a transparent machine where you can see the gears turning. However, the researchers found that when these machines show you exactly how they turned the gears, it can be overwhelming.

The Analogy: Imagine trying to read a map to find a coffee shop. The "raw" map shows every single tree, every crack in the sidewalk, and the name of every street in the city. It's technically accurate, but it's so cluttered with "irrelevant details" that you can't find the coffee shop. You need a map that abstracts away the noise.

The Solution: Two Ways to Clean the Map

The researchers used a logic system called Answer Set Programming (ASP) to create explanations. They tested two ways to "clean up" these explanations by removing or grouping irrelevant details:

Removal (The "Delete" Button): This takes out details that are true for everything and therefore don't help distinguish one option from another.
- Analogy: If you are trying to identify a cat, the fact that "it has a body" is true for all cats. Removing that detail from your explanation doesn't hurt; it just saves space. You only keep the details that matter, like "it has stripes" (for a tiger) or "it has spots" (for a leopard).
Clustering (The "Grouping" Button): This takes several similar details and bundles them into one big concept.
- Analogy: Instead of listing "wet from rain," "wet from a puddle," and "wet from a hose" as three separate reasons, you just group them under one label: "Wet."

The Experiment: The "Flower, Mushroom, and Cactus" Test

To see if these "cleaned-up" explanations actually helped humans, the researchers ran an online experiment.

The Setup: Participants were shown rules for classifying different biological specimens (flowers, mushrooms, and cacti).
The Groups: Some people saw the "Raw" map (all the details). Others saw the "Removed" map (just the unique details). Others saw the "Clustered" map (grouped details). A final group saw a mix of both.
The Task: After learning the rules, they had to classify new, unseen specimens as fast and accurately as possible.

The Results: A Double Win

The study found that "cleaning" the explanations helped humans in two very specific, different ways:

1. Clustering helps you get the answer right.
When details were grouped together, people made fewer mistakes.

The Metaphor: It's like organizing a messy closet. If you group all your "summer clothes" together, you can find what you need faster and are less likely to grab a winter coat by mistake. The structure helped people understand the logic better.

2. Removal helps you get the answer faster.
When irrelevant details were deleted, people made decisions much quicker.

The Metaphor: It's like turning off the radio while driving in a complex city. You don't need the music to navigate; you just need the road signs. By removing the "noise," people didn't waste time reading things that didn't matter.

The Surprise:
Even though people were faster and more accurate with the cleaned-up explanations, they didn't feel more confident about their answers. They knew they were doing better, but their internal sense of "I'm sure about this" didn't change.

The Bottom Line

The paper concludes that when we explain how an AI thinks, we shouldn't just dump all the data on the user. We need to be smart about what we show them:

Use grouping (clustering) to help people understand the logic and avoid errors.
Use deletion (removal) to help people make decisions quickly by cutting out the fluff.

By doing this, we can build AI systems that aren't just technically correct, but are also friendly to the human brain.

1. Problem Statement

While Symbolic AI offers inherent transparency compared to subsymbolic systems (like deep learning), raw logical traces generated by solvers often impose high extraneous cognitive load on human users. These traces frequently contain redundant or distracting details that do not contribute to the core decision logic.

The Gap: Existing formal methods for abstraction (e.g., removing or clustering irrelevant atoms) lack empirical validation regarding their impact on human reasoning performance and cognitive effort.
The Goal: To determine if formally abstracting "irrelevant" details in Answer Set Programming (ASP) explanations improves human understanding (accuracy), reduces cognitive effort (response time), and increases confidence.

2. Methodology

Theoretical Framework: $\chi$ -Irrelevance in ASP

The authors utilize Answer Set Programming (ASP) as the formal foundation. They define a notion of $\chi$ -irrelevance to identify atoms (facts) that can be safely abstracted without altering the solution space for a specific set of problem instances ( $\chi$ ).

Mapping ( $m$ ): A surjective mapping is applied to the program $P$ to create an abstracted program $m(P)$ .
Two Abstraction Operations:
1. Removal: Mapping atoms to $\top$ (logical truth). This "forgets" details that lack discriminative power across the instance set (e.g., if all flowers have spiky leaves, the "spiky" attribute is removed as it doesn't distinguish classes).
2. Clustering: Mapping multiple distinct atoms to a single representative atom. This groups related logical entities into higher-level concepts (e.g., grouping habitatWater and habitatMud into habitatWaterOrMud).
Formal Guarantee: The abstraction preserves the correspondence of answer sets for any potential extensions of the program within the defined scenario set $\chi$ .

Empirical Study Design

Task: A concept-learning binary classification task. Participants learned rules to classify biological specimens (Flowers, Mushrooms, Cacti) based on attributes (e.g., "slow growth," "scent").
Domains: Three distinct biological domains to prevent proactive interference.
Experimental Groups (Between-subjects):
1. Default: Full logical trace with all technical attributes.
2. Cluster: Explanations with irrelevant details clustered.
3. Removal: Explanations with irrelevant details removed.
4. Cluster-Removal: Explanations with both operations applied.
Procedure:
- Learning Phase: Participants viewed 2 positive examples with explanations specific to their group.
- Test Phase: Participants classified 6 new instances (3 positive, 3 negative) without explanations, recording their accuracy, answer time, and confidence (0–100 slider).
Participants: 99 valid participants recruited via Prolific, randomly assigned to the four groups.

3. Key Contributions

Formal Relaxation: The authors propose a relaxed notion of abstraction in ASP based on irrelevancy within a specific problem space, moving beyond previous restricted definitions.
Dual-Benefit Hypothesis: They empirically validate that different abstraction strategies serve distinct cognitive functions:
- Clustering aids in structural understanding.
- Removal aids in reducing cognitive load.
Tool Integration: The study integrates formal logic (ASP) with cognitive science experiments using tools like xclingo to generate and visualize abstracted justifications.

4. Results

The study analyzed three dependent variables: Accuracy, Answer Time, and Confidence.

Accuracy (Understanding):
- Significant Improvement: The Clustering group (and the combined Cluster-Removal group) showed significantly higher accuracy compared to the Default group ( $p < 0.05$ ).
- Insight: Grouping features helped participants form clearer mental models of the classification rules, reducing errors.
- Removal Effect: Removal alone did not significantly improve accuracy, possibly because the original explanations were not complex enough to hinder understanding, or the removal focused on specificity rather than abstraction.
Answer Time (Cognitive Effort):
- Significant Reduction: The Removal strategy significantly reduced answer times. The Cluster-Removal group was the fastest, significantly outperforming both the Default and Cluster-only groups.
- Insight: Removing irrelevant details acts as the primary driver for speed, allowing users to ignore undecisive attributes. Clustering did not negatively impact speed.
Confidence:
- No Significant Effect: Abstraction type did not significantly alter self-reported confidence.
- Observation: Participants were generally more confident in correct answers than incorrect ones, regardless of the explanation type. This suggests that while abstraction improves objective performance, it does not necessarily alter the subjective feeling of certainty.

5. Significance and Conclusion

Cognitive Alignment: The paper demonstrates that symbolic AI explanations must be tailored to human cognitive processes. A "technically sound" trace is not necessarily a "cognitively manageable" one.
Strategic Abstraction: The findings highlight a dual role for abstraction:
- Use Clustering to improve comprehension and accuracy (structural organization).
- Use Removal to improve efficiency and speed (reducing cognitive load).
Neuro-Symbolic Implications: This work bridges the gap between formal logic (System 2 reasoning) and human cognitive limitations, providing a blueprint for generating human-centered explanations in Neuro-Symbolic AI systems.
Future Work: The authors suggest exploring how the complexity of the default explanation affects the efficacy of removal and investigating whether $\chi$ -irrelevance captures "specificity" rather than true abstraction.

In summary, the paper provides empirical evidence that formal abstraction is not a one-size-fits-all solution; rather, specific abstraction techniques (clustering vs. removal) should be selected based on whether the goal is to maximize human understanding or minimize cognitive effort.