Opening — Why this matters now

Explainable AI has spent years chasing a mirage: explanations that feel intuitive to humans but are generated by machines that have no intuition at all. As models creep further into regulated, safety‑critical, or user‑facing domains, the cost of a bad explanation isn’t just annoyance—it’s lost trust, rejected automation, or outright regulatory non‑compliance.

Logic puzzles—Sudoku, logic grids, scheduling mini‑problems—may seem trivial, but they are controlled environments where explanation methods can be dissected without the messy ambiguity of real‑world data. This new paper proposes something unusually practical: instead of guessing what humans prefer in an explanation, learn their preference structure directly through interactions, using a method charmingly named MACHOP.

It’s not about making machines smarter. It’s about making explanations closer to how humans reason—step by step, constraint by constraint.

Background — Context and prior art

Constraint Programming (CP) has long struggled with explainability. Traditional approaches aim for “minimal explanations,” usually counting how many constraints are needed to justify a step. Intuitive, yes, but often misleading: a short explanation is not always the clearest.

Prior work tried to manually craft linear objective functions that score explanation quality. The problem? Human interpretability is not linear. And hand‑tuning these objectives requires the sort of empathy and domain intuition machines definitely don’t have.

This is where Constructive Preference Elicitation (CPE) enters: show the user two alternatives, learn from the comparison, and iteratively estimate the weighting of different sub‑objectives. In theory, this allows explanation systems to adjust themselves to different learners, contexts, or even domains.

But applying CPE straight to explanation steps—especially messy ones involving dozens of constraint types—breaks quickly. Differences in feature scales distort learning. Query suggestions become redundant. User wait time balloons.

MACHOP was built to fix that.

Analysis — What this paper contributes

The authors extend the CPE framework to step‑wise explanations and introduce several innovations:

1. Dynamic feature normalization

Explanation steps are multi‑objective: number of facts, number of row constraints, block constraints, adjacency, etc. Their scales vary wildly. Blindly mixing them leads to unstable learning.

The paper proposes two normalization schemes:

  • Cumulative: scale relative to the maximum seen so far.
  • Local: scale relative to the most recent comparison.

The latter performs best—simple, robust, and responsive.

2. MACHOP: A new query generation strategy

Traditional CPE suffers from redundant suggestions (two nearly identical explanations). MACHOP fixes this with two tools:

  • Non‑domination constraint: ensures y2 is meaningfully different from y1.
  • UCB‑based diversification: borrowed from multi‑armed bandits, it pushes exploration into under‑examined sub-objectives while still weighing user‑important ones.

This balances exploration and exploitation—ideal for human‑in‑the‑loop learning.

3. Practical runtime improvements

Computing optimal explanation steps can be slow, especially when every possible fact must be considered. They offer:

  • Online fact selection: best but slow.
  • Offline SES sequencing: faster and nearly as good.

For business applications, this distinction matters: users won’t wait 50 seconds for a query.

Findings — Results with visualization

1. MACHOP reduces regret by ~80%

Across Sudoku and logic‑grid puzzles, MACHOP consistently outperformed the standard Choice Perceptron.

Method Avg. Regret (Sudoku) Avg. Regret (Logic Grid)
Choice Perceptron 2.0 3.8
MACHOP 0.4 0.9

2. Real users preferred MACHOP explanations

After 30 queries:

  • MACHOP explanations were preferred over SES in 70.7% of cases.
  • Choice Perceptron managed only 44.8%.

3. Query generation time becomes feasible

Human experiments show MACHOP queries can be generated in 1–3 seconds. That’s well within acceptable UX boundaries.

Implications — Next steps and significance

The technical contribution is interesting, but the implications extend far beyond puzzles.

1. Regulators will eventually demand explainability preferences

Not everyone understands explanations the same way. Learning‑based explanation systems could adapt to:

  • novice vs expert users
  • compliance auditors vs operators
  • safety‑critical vs everyday automation

2. Preference elicitation may become a required component of enterprise AI

Rather than guessing how a bank auditor wants a fraud decision explained, the system can learn it.

3. MACHOP is a step toward personalized explainability

This aligns naturally with agentic workflows: explanation agents that adapt to the user over time.

4. In Business Automation (Cognaptus’ world), this matters

Imagine an AI workflow assistant explaining why it routed a ticket, rejected an invoice, or scheduled an action. Today these chains are opaque. Tomorrow they may be personalized, preference‑aligned, and dynamically learned.

MACHOP gives us the scaffolding: structured features, interactive comparisons, and a stable learning loop.

Conclusion

This paper offers a rare combination: theoretical elegance with practical usability. MACHOP takes a tedious piece of explainable AI—choosing how to explain—and turns it into a learnable interaction. The prototypes focus on puzzles, but the trajectory points toward real business relevance: adaptive explainability for automation systems, compliance tooling, and interactive agent UX.

As enterprises adopt deeper AI pipelines, the ability to tune not just decisions but how decisions are explained will be a competitive advantage.

Cognaptus: Automate the Present, Incubate the Future.