LLM Explanations of Rejected Requests: Correct Reasons ≠ Effective Recourse

The paper proposes recourse validity, showing that a single correct reason explanation is nearly ineffective when multiple rules reject.
When rule-based systems reject requests, LLMs often write explanations, but if multiple rules independently reject, the explanation may cite only one true reason and the user is still rejected after fixing it. The paper calls this missing property recourse validity: changing the cited reason without changing the request itself—can it flip the decision? Using a hitting set argument, the authors predict that when multiple rules reject, one correct reason is almost necessarily ineffective. Testing seven rule systems from access control and US welfare eligibility on 20 LLMs, they find single correct reasons almost never provide recourse, correctness checks are near random, but a strong judge asked about action effectiveness can identify failure. They recommend building explanations from the decision function and judging by action effectiveness.
Title
LLM Explanations of Rejected Requests: Correct Reasons ≠ Effective Recourse
Meta Description
This paper introduces the concept of recourse validity: when an LLM explains a rejected request, pointing out the correct reason may be ineffective—users may still be rejected after making corrections. Experiments show that checking reason correctness is close to random, and explanations should be constructed from the decision function.
Article Content
When rule-based systems reject user requests, language models are increasingly used to write explanations. But if multiple rules independently reject a request, the explanation may point out only one true reason, and the user is still rejected after fixing it. This paper calls this missing property recourse validity: if you change the reason the explanation points to without changing the request itself, can the decision be flipped?

Definition and Computation of Recourse Validity
The paper's core contribution is to precisely define and compute recourse validity. Given a rule system (such as an access control policy or US welfare eligibility rules), an explanation is recourse-valid if and only if the user, after modifying their request according to the explanation, sees the decision change from reject to allow.
The authors use a hitting set argument to predict from the policy itself: when multiple rules reject simultaneously, providing only one correct reason is almost necessarily ineffective. This is because each rejection rule constitutes a constraint that must be satisfied; modifying one reason satisfies only one of the constraints, while the others still block the request.
Hitting set is a classic concept in combinatorics and computer science: given several sets, a hitting set is a set that intersects every one of them. In the context of rule systems, each rejection rule can be viewed as a "constraint set"—the set of attributes the user must modify to satisfy that rule. To flip the decision, the user's modification must "hit" the constraint sets of all rejection rules, i.e., form a hitting set covering all rule constraints. If only one rule rejects, the user only needs to hit that single set; but when multiple rules reject simultaneously, modifying just one reason hits only one of the sets, and the remaining sets are still unsatisfied, so the decision will not flip. This structural argument explains why a single correct reason almost necessarily fails when multiple rules stack.
Experimental Design and Main Findings
The study ported seven rule systems from access control and US welfare eligibility and tested them on 20 LLMs. Key findings include:
- In six of the seven rule systems, when multiple reasons apply simultaneously, a single correct reason almost never provides recourse.
- Checking whether the cited reason is correct (whether computed or asked of an LLM judge) is close to random in identifying this failure, consistent with the prediction of the hitting set argument.
- A strong judge asked "would acting on the explanation succeed?" can identify this failure.
- Stronger and reasoning models perform better, but still fail a lot on rejections with four or more independent reasons.
- Asking to list all rules helps all open models that have room for improvement.
LLM-as-a-judge is a common paradigm for evaluating language model output quality today: using a more capable language model to score or judge another model's output, replacing expensive human annotation. In this study, the judge is used to determine whether the explanation correctly cites the rejection reason, or whether acting on the explanation would flip the decision. The paper finds that when the judge is only asked "is the cited reason correct?", its judgment is close to random—because the reason may indeed be correct yet still not constitute recourse. When the same judge is asked "would acting on the explanation succeed?", it can identify this failure. This shows that judge prompt design is crucial: the question asked must align with the true objective of interest (recourse validity), rather than staying at surface-level correctness.
Why "Correct" Is Not the Same as "Effective"
In traditional explainable AI evaluation, an explanation's "correctness" usually refers to whether it faithfully reflects the model's decision basis. But the paper reveals a more practical problem: even if the explanation correctly points out a rejection reason, the user may still be rejected after acting on it. This is especially prominent in scenarios with multiple stacked rules.
For example, welfare eligibility may be denied simultaneously due to income, assets, and residency status. The explanation mentions only income, and after the user increases income, they are still rejected due to assets or residency status. Although this explanation is "correct", it lacks recourse validity.
Faithfulness in explainable AI (XAI) usually refers to whether an explanation truly reflects the model's internal decision basis—for example, whether a feature attribution method accurately characterizes each feature's contribution to the output. Faithfulness is a traditional core metric of explanation quality, but it measures "the consistency between the explanation and the model's decision basis", not "whether the explanation is useful to the user". Recourse validity belongs to another dimension: it focuses on whether the explanation can guide the user to take effective action. An explanation can be highly faithful yet have no recourse value—for example, accurately pointing out that income is one reason for rejection while omitting the asset condition that also causes rejection. Distinguishing faithfulness and correctness from recourse validity helps explain why existing evaluation metrics cannot capture failure modes in multi-rule scenarios.
Implications for LLM Explanation Generation
The paper suggests that explanations of eligibility decisions should be constructed from the decision function and judged by "whether acting on the explanation is effective". This has direct implications for practical system design:
- Explanation generation should not rely solely on free-form LLM writing, but should incorporate structural information from the rule system.
- When evaluating explanation quality, in addition to correctness, a recourse validity metric should be introduced.
- For multi-rule rejection scenarios, explanations should list all independent reasons where possible, or at least point out the multiple conditions that must be satisfied simultaneously.
Limitations and Future Directions
Experiments focus on rule systems; applicability to more complex policies (such as rules with priorities or time-dependent conditions) remains to be verified. In addition, although a strong judge can identify failures, it is costly, and how to generate recourse-valid explanations at low cost is an engineering challenge.
The paper's conclusion points to a pragmatic direction: the value of an explanation lies not in "saying something correct", but in "whether the user can pass after following it".
One-line Summary
The paper proposes recourse validity, revealing that a single correct reason explanation is almost ineffective when multiple rules reject.
Paragraph Summary
When rule-based systems reject user requests, language models are often used to write explanations, but if multiple rules independently reject the request, the explanation may point out only one true reason, and the user is still rejected after fixing it. The paper calls this missing property recourse validity: if you change the reason the explanation points to without changing the request itself, can the decision be flipped? Using a hitting set argument, the authors predict from the policy itself that when multiple rules reject simultaneously, providing only one correct reason is almost necessarily ineffective, because each rejection rule constitutes a constraint that must be satisfied. The study ported seven rule systems from access control and US welfare eligibility, tested on 20 LLMs, and found that a single correct reason almost never provides recourse, while checking reason correctness is close to random, but a strong judge asked "would acting succeed?" can identify failure. The paper suggests explanations should be constructed from the decision function and judged by action effectiveness.
Related articles

RePlan-VLA: Progress-Aware Replanning Breakthrough for Long-Horizon Robotic Manipulation
The RePlan-VLA framework converts VLA models' short-horizon capabilities into reliable long-horizon execution through progress-aware adaptive replanning, improving success rates by 21.9% and 10.4% on RoboCasa365 and CALVIN, and validating real-world deployment on a Franka Research 3 robot.

Geometric Bridge Flow Matching: A New Breakthrough in Blind Rigid Motion Correction
Geometric bridge flow matching enables blind rigid motion correction in head CBCT, achieving 37.1 dB PSNR and 0.27 mm RPE on simulated data in ~9.1 minutes, outperforming existing methods.

Google Releases Gemini 4 Hydrogen: 2.5x Faster Without Sacrificing Intelligence
Google releases the lightweight Gemini 4 Hydrogen: 2.5x faster task completion, AA Intelligence Index of 54 beating its flagship sibling, 7.3s single-step on Agent Arena faster than Claude and GPT rivals, four thinking levels with pay-as-you-need pricing, now available across AI Studio and more.