{"id":"258356b3-d150-42f5-802d-ca59ebf2a56e","arxiv_id":"2506.12240","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An in-context LLM framework that produces dual expert and non-expert explanations, evaluated on well-being clustering with a user study and LIME-alignment metrics.","lead":"This paper proposes an LLM-based framework that generates both expert-level technical explanations and simple, human-friendly explanations from black-box AI models in a single response. It tests the framework on a well-being clustering task and reports that users find the explanations clearer than a standard XAI method, while the technical outputs correlate with LIME feature rankings.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Content-quality claim rests on a circular metric: LIME supplies the few-shot demonstrations in Section 4 and is then the 'ground truth' in Section 5.2, so Spearman=0.92 measures imitation, not correctness.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing weakness in the paper's central quantitative claim. The framework has two independently testable contributions: generating expert-facing feature-importance explanations and generating non-expert-friendly text. The second is supported by the paired-comparison user study, which, despite its modest size and possible ordering effects, shows a large, significant pragmatic-quality advantage and is a genuine behavioral outcome. The first, however, is validated only by agreement with LIME, and LIME is simultaneously the source of the few-shot demonstrations. This makes Spearman=0.92 an expected consequence of in-context learning rather than an external check. The paper's own Section 6 discusses hallucinations and the black-box nature of LLM feature-importance computation, but it treats Table 4 as evidence that hallucinations are reduced; that argument inherits the same circularity. I therefore agree with the conditional verdict. The concern does not require rejection because the human-friendliness claim stands on independent user data and the framework is reproducible with code. It does require either an external ground-truth comparison or a clear claim downgrade to 'imitates LIME under few-shot prompting.' The proposed SHAP/independent-surrogate recomputation is feasible and would settle the issue.","tokens_in":19183,"tokens_out":4446,"duration_ms":56144,"concrete_test":"Recompute Table 4's Spearman correlations against an independent explanation method that was never used as a demonstration (e.g., SHAP values or a fresh local sparse linear surrogate computed on the same 20 test instances), and verify in the released code that the evaluation instances are disjoint from the few-shot demonstrations. If Spearman against the independent method is materially below 0.92 (say, below 0.6), then 0.92 reflects imitation of the prompt examples and the content-quality claim must be downgraded; if the independent correlation remains high, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim of high content quality (Spearman rank correlation=0.92; Abstract, Section 5.2, Table 4) is supported by an evaluation that is circular. In Section 3.2, LIME is selected as the ground-truth XAI method because of its 0.93 fidelity, and its instance-level feature rankings form the 'contextual thesaurus.' In Section 4, those same LIME explanations are inserted into the few-shot prompt as demonstration examples. In Section 5.2, the framework's outputs are then compared against LIME on 20 randomly selected instances. A model shown a handful of LIME outputs and asked to produce feature-importance rankings is very likely to reproduce the format and content of those examples, especially if the evaluation instances are not explicitly disjoint from the demonstrations; the paper does not state that they are. The zero-to-few-shot trajectory in Table 4 makes the mechanism concrete: LLaMA3 Spearman rises from 0.01 with no demonstrations to 0.92 with few-shot LIME examples. That is evidence of successful in-context imitation, not evidence that the explanations are correct. The problem is compounded because LIME fidelity measures surrogate agreement with the classifier, not whether the features are genuinely responsible for well-being clusters; if LIME is inaccurate for this problem, both the demonstrations and the scoring metric inherit the same error. The user study in Section 5.1 is independent evidence for human-friendliness in a paired comparison, but it does not test technical correctness. Therefore the claim that the framework delivers 'high-quality technical explanations aligned with foundational XAI methods' is overstated; at most, the evidence shows alignment with LIME under few-shot prompting.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an LLM-based framework for human-centered explainable AI (HCXAI) that uses in-context learning to generate, in a single response, both technical explanations for experts and human-friendly explanations for non-experts. The framework is instantiated on a well-being clustering task using the LifeSnaps dataset, with a ``contextual thesaurus'' built from LIME explanations over 20 instances. The authors evaluate the framework through a user study (N=56) comparing their explanations with LIME, and through automatic structure and content quality metrics on 20 instances. They report a Spearman rank correlation of 0.92 with LIME ground truth and significant improvements in pragmatic quality over LIME in a paired comparison, concluding that LLMs can serve as HCXAI enablers.","tokens_in":19496,"tokens_out":2069,"duration_ms":24917,"significance":"If the claims were fully supported, the work would be a useful contribution to the growing literature on using LLMs to make XAI accessible, and the reproducible code release and detailed comparison of prompting techniques are commendable. The user study provides real, if partial, evidence that the framework's explanations are preferred by non-experts over raw LIME output in a paired comparison, particularly on pragmatic quality. However, the central quantitative claim of high content quality (Spearman=0.92, Abstract and Section 5.2) is undermined by a circular evaluation setup: the same LIME explanations that serve as few-shot demonstrations in the prompt are also used as the scoring ground truth. The reported numbers therefore measure how well the LLM imitates its in-context examples, not whether the explanations are correct in any external sense. The paper is well-structured and transparent about its methodology, which makes the circularity easy to identify, but the main evidence for technical content quality needs substantial rework before the central claim can be accepted.","major_comments":[{"comment":"The Spearman=0.92 result is not a valid measure of content quality because of circularity. In Section 3.2, LIME explanations are selected as the ground-truth ``thesaurus''; in Section 4, the few-shot prompt includes those same LIME explanations as demonstration examples; and in Section 5.2, the framework's outputs are compared against LIME on 20 randomly selected instances. The paper does not state whether the 20 evaluation instances are disjoint from the few-shot demonstration instances. If they are not disjoint, the comparison reduces to testing how well the LLM copies its input examples. Even if they are disjoint, the zero-to-few-shot trajectory in Table 4 (LLaMA3 Spearman rising from 0.01 to 0.92) is exactly what would be expected from in-context imitation of the output format and feature-name style, and does not establish correctness with respect to any external notion of feature importance. The authors should either (a) evaluate against a ground truth that is independent of the demonstrations (e.g., human expert feature judgments on the same instances), or (b) at minimum demonstrate that the evaluation instances are disjoint from the demonstrations and that the correlation persists when the demonstrations are replaced by LIME explanations on different instances, and clearly report this separation.","section":"Sections 3.2, 4, and 5.2, Table 4"},{"comment":"The designation of LIME as ground truth rests on a fidelity of 0.93, but LIME fidelity measures how well the local surrogate agrees with the classifier's predictions, not whether the identified features are causally responsible for the well-being clusters. The authors reformulate clustering as a classification task (with 99% accuracy) and then apply LIME to the classifier, so any errors in the surrogate or in the reformulation are inherited by both the demonstrations and the evaluation metric. This is a correctness-risk concern, not just a circularity issue: if LIME is inaccurate for this clustering problem, the ``high content quality'' claim collapses regardless of the Spearman value. The authors should provide an independent validation of the feature importance rankings, or explicitly temper the claim from ``high content quality'' to ``high similarity to LIME'' and discuss what that does and does not establish.","section":"Section 3.2, Table 1"},{"comment":"The summary statement that the framework ``significantly outperforms LIME in a user study'' is stronger than the results support. In the A/B test (Table 2), the confidence intervals all include zero for every item and overall scale, and the authors correctly note the absence of statistically significant differences there. The significant advantage appears only in the paired comparison (Table 3), where it is confined to Pragmatic Quality and its four items; the Hedonic Quality overall difference (0.11 vs. -0.27) is not statistically significant, and the confidence intervals for several HQ items include zero. The abstract and conclusion should be reworded to state that the paired comparison showed significant improvements in pragmatic quality, while hedonic quality improvements were not significant and the independent A/B test showed only non-significant trends. This does not invalidate the user-study evidence, but it narrows the claim considerably.","section":"Section 5.1, Tables 2 and 3"}],"minor_comments":[{"comment":"The phrase ``Alarming evidence about AI's unprecedented human values risks'' is grammatically unclear; consider revising to ``risks to human values'' or similar.","section":"Abstract and Section 1"},{"comment":"The colors (light blue, orange, light green) used to mark configurable parts in the prompt and system templates are not visually annotated in the text, making it hard to follow which parts are data, model, and explanation-specific. Please add a legend or describe the color coding more explicitly in the caption.","section":"Section 4, Figure 5"},{"comment":"The tables are dense; adding a note that confidence intervals are 95% confidence intervals and stating the sample size per group in the caption (N=25 for LIME, N=31 for framework in Table 2) would improve interpretability.","section":"Section 5.1, Tables 2 and 3"},{"comment":"The row labels ``structure'' and ``content'' are ambiguous; the table header reads ``structure Content''. Clarify that structure quality comprises coherence, grammar errors, readability, and sentiment consistency, while content quality comprises Spearman, NDCG difference, and Euclidean distance.","section":"Section 5.2, Table 4"},{"comment":"The limitation that ``the current implementation supports only realistic features'' is valuable, but the discussion of hallucinations equates low NDCG difference with reduced hallucination; this is an indirect proxy and should be stated as such rather than as a direct measure of factual accuracy.","section":"Section 6, Limitations"}],"recommendation":"major_revision","confidential_remarks":"The circularity of the content-quality evaluation is the main obstacle to acceptance; the user study and the overall framework design are publishable pieces if the content-quality claim is re-framed or supported with an independent ground truth. I would also suggest the authors compare more directly with x-[plAIn] (Ref. [37]) since that work is very close in spirit, and clarify the novelty of their contribution relative to it in the related-work section."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the configuration is new: feeding XAI method outputs as few-shot demonstrations and returning dual expert/non-expert explanations in one LLM response. I haven't seen that exact setup before, and it is a sensible, practical idea. Second, the main quantitative claim—content quality via Spearman rank correlation 0.92 with LIME—is less convincing than the abstract suggests, for the reason the stress-test note identifies: LIME explanations are the few-shot demonstrations in Section 4 and also the \"ground truth\" in Section 5.2. That is not a clean external benchmark; it is at least partially circular.\n\nWhat the paper does well: the framework is clearly described, the code is on GitHub, and the authors are honest about limitations in Section 6. The clustering benchmark is thorough (six algorithms, multiple variants, many metrics). The user study (N=56) is real independent evidence that the non-expert explanations are more usable than raw LIME in a paired comparison — the pragmatic quality numbers are respectable and the A/B test, though not significant, points in the same direction. The authors also correctly note that structure quality is mixed and depends more on the LLM than on prompting technique.\n\nSoft spots, in rough order of severity:\n1. The content-quality claim is overstated. LIME is not \"ground truth\" for feature importance; it is a surrogate-based method whose fidelity (0.93) says nothing about whether the features are genuinely responsible for the clusters. Showing that the LLM imitates LIME under few-shot prompting is expected, not strong evidence of \"high-quality technical explanations.\" The zero-to-few-shot trajectory in Table 4 (Spearman 0.01 → 0.92) makes the imitation story concrete. The paper does not state whether the 20 evaluation instances are disjoint from the demonstrations—another small but addressable gap.\n2. Generalizability claims exceed evidence. The abstract and introduction promise domain-, model-, explanation-agnostic behavior, but evaluation covers one domain, one clustering model, and one XAI method. The \"agnostic\" design is only a prompt template; there is no empirical support for transfer.\n3. The user study, while useful, is small, mostly STEM participants, and the significant effect appears only in the paired comparison, not the A/B test. That is fine as a preliminary result, but the abstract's \"improved interpretability\" overstates it.\n\nBottom line: this paper deserves a serious referee. It is a real contribution to the XAI/HCI literature and the framework idea is worth building on, but the authors need to reframe the content-quality claim (alignment with LIME, not ground-truth correctness) and ideally test on an external benchmark or an additional XAI method. For a reading group, I'd say yes—it is a good candidate for discussing evaluation design in LLM+XAI. I would not cite it in my own work in the next year, mainly because the core quantitative claim is not yet solid, though I might revisit after revision.","headline":"Worth engaging: a genuinely new LLM+XAI framework with a real user study, but the headline Spearman=0.92 is weaker than it looks because LIME both supplies the in-context examples and serves as the scoring target.","tokens_in":20072,"tokens_out":2460,"would_cite":false,"duration_ms":32118,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A structured LLM prompt with a few LIME examples produces expert-grade feature rankings and clear explanations for non-experts in one response.","keywords":["Explainable AI (XAI)","Large Language Models","Human-centered explainability","In-context learning","Feature importance","User study","Well-being clustering","LIME"],"falsifier":"Replace the few-shot demonstrations with a deliberately permuted LIME ranking for the same instances while keeping the prompt otherwise identical; if the LLM's output correlates with the permuted ranking instead of the original LIME ranking, then the reported Spearman=0.92 measures imitation of the demonstrations, not correctness of the explanation.","tokens_in":18995,"feed_emoji":"🤖","tokens_out":6455,"duration_ms":100738,"temperature":0.7,"pith_summary":"The paper claims that a carefully structured LLM prompt, fed with a handful of LIME-style explanation examples, can produce both expert-grade feature-importance rankings and plain-language explanations for non-experts in a single response. The framework is built around a ground-truth contextual 'thesaurus' derived from a well-being clustering analysis, and it is meant to close three gaps in current XAI: black-box transparency, explanations that make no sense to non-experts, and the costly post-processing step needed to humanize raw XAI output. If the claim holds, high-stakes domains like healthcare could skip the standard three-step pipeline of model inference, explanation extraction, and manual interpretation, and instead hand both audiences one immediately usable explanation. The paper reports a Spearman rank correlation of 0.92 with LIME feature rankings (LLaMA3, few-shot) and a user study (N=56) in which the framework's explanations scored significantly higher than LIME on pragmatic quality.","feed_headline":"LLM framework matches LIME rankings, wins on human clarity","feed_subtitle":"A single LLM response gives experts LIME-level rankings and non-experts text users rate clearer than LIME.","key_machinery":"The load-bearing mechanism is the ground-truth contextual 'thesaurus': a benchmark set of well-being cluster assignments and LIME feature-importance explanations, built from over 40 data, model, and XAI combinations, which is fed into the LLM through structured prompt and system templates. The prompt supports zero-, one-, and few-shot modes; the few-shot mode inserts several demonstration examples of the form output-features-explanation. The system template has three parts: 'Setting the Scene' tells the LLM to compute feature importance; 'Expert Response' forces a JSON list of features with importance values; 'Non-expert Response' turns those values into a plain-language paragraph. Everything hinges on in-context learning transferring the LIME-style ranking procedure from the demonstrations to new instances, which is what moves Spearman correlation from near zero in zero-shot to 0.92 in few-shot.","core_discovery":"The central claim is that LLMs, when given in-context demonstrations of a foundational XAI method, can act as a full human-centered explainability layer: they reproduce the technical feature-importance rankings an expert would get from LIME, and simultaneously translate those rankings into a human-friendly paragraph a non-expert can act on. The paper constructs the demonstrations by clustering the LifeSnaps well-being data with k-means, converting the cluster assignments into a classification problem (99% accuracy with a support vector classifier), and generating LIME explanations on those predictions; the best clusterings and LIME explanations form the thesaurus that is injected into the prompt. In the few-shot setting LLaMA3 reaches Spearman rank correlation 0.92 with LIME's rankings, with NDCG difference 0.001 and Euclidean distance 0.02, while Mistral reaches 0.74, and zero-shot performance is close to zero, showing that the contextual demonstrations, not the LLM's prior knowledge, carry the content quality. In a paired user study, the framework's human-centered explanations outperformed LIME on overall pragmatic quality (1.00 vs -0.16) and on every pragmatic item, with hedonic gains present but not statistically significant. The paper positions this as evidence that LLMs are viable enablers of human-centered explainable AI.","pith_inferences":["A natural test of whether the LLM is computing feature importance or imitating its examples is to swap the few-shot demonstrations to a different XAI method (e.g., SHAP or Anchor): if the outputs track the new method, the framework is a translation layer for whatever explanation it is shown rather than an independent reasoner.","The user-study gain may partly reflect modality rather than LLM reasoning, since the comparison is text-plus-JSON against LIME's raw visualization; a human-written plain-language translation of the same LIME output would isolate how much the LLM adds.","The agnosticism claim could be stress-tested by rebuilding the thesaurus in a different high-stakes domain, such as credit decisions or clinical risk scores, and checking whether the Spearman correlation and user preference survive the transfer; degradation would show the thesaurus, not the templates, is the bottleneck."],"forward_implications":["Few-shot demonstrations are what make the framework work: LLaMA3's rank correlation with LIME jumps from 0.01 zero-shot to 0.92 few-shot, so the contextual examples, not the pre-trained model alone, carry the explanation quality.","The framework collapses the usual three-step pipeline (model inference, explanation extraction, post-processing) into a single LLM response for new instances.","For non-experts, the framework's explanations beat raw LIME on user-rated clarity, ease of use, efficiency, and supportiveness, with the paired overall pragmatic-quality score rising from -0.16 to 1.00.","Because the prompt and system templates are designed to be data-, model-, and explanation-agnostic, the same scaffolding can be rebuilt for other domains and tasks by supplying a new thesaurus.","Content quality is measured against LIME, so the framework's 'high-quality technical explanation' claim is bounded by LIME's fidelity (0.93 on the benchmarking sample) as the reference standard."],"supporting_citations":[{"why":"Supplies LIME, the foundational XAI method whose explanations serve as both the ground-truth thesaurus and the few-shot demonstrations.","marker":"[47]"},{"why":"Establishes the in-context few-shot learning capability that the framework relies on to transfer explanation knowledge.","marker":"[6]"},{"why":"Provides the User Experience Questionnaire (UEQ) with pragmatic and hedonic quality scales used in the N=56 study.","marker":"[28]"},{"why":"Supplies the LifeSnaps multi-modal well-being dataset used for the clustering analysis and thesaurus construction.","marker":"[65]"},{"why":"Motivates the training/validation feature split used to derive and interpret well-being clusters.","marker":"[27]"},{"why":"Supports the paper's observation that few-shot prompting improves LLM output quality, aligning with the reported content-quality gains.","marker":"[8]"},{"why":"One of the benchmarked XAI methods (Anchor) whose explanations are compared when building the thesaurus, showing the benchmarking breadth behind the ground truth.","marker":"[48]"}],"fun_headline_variants":["LLM explains like LIME, clearer to humans","LLM reproduces LIME, humans favor its clarity","One LLM response: expert rankings plus plain talk","LLM bridges XAI gap for experts and laypeople"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that LIME's feature-importance rankings are the correct explanation, and because those same rankings are also the few-shot demonstrations, the measured correlation measures imitation of the examples rather than external correctness.","fun_headline_variants_meta":{"raw":{"variants":["LLM explains like LIME, clearer to humans","LLM reproduces LIME, humans favor its clarity","One LLM response: expert rankings plus plain talk","LLM bridges XAI gap for experts and laypeople"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000377,"raw_usage":{"total_tokens":2106,"prompt_tokens":1142,"completion_tokens":964,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":758,"completion_tokens_details":{"reasoning_tokens":908}},"tokens_in":758,"tokens_out":964,"duration_ms":101538,"temperature":1.0,"reasoning_tokens":908,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:56:18.582657+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the few-shot demonstrations with a deliberately permuted LIME ranking for the same instances while keeping the prompt otherwise identical; if the LLM's output correlates with the permuted ranking instead of the original LIME ranking, then the reported Spearman=0.92 measures imitation of the demonstrations, not correctness of the explanation.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the User Experience Questionnaire (UEQ) with pragmatic and hedonic quality scales used in the N=56 study."},{"cited_title":"Scientific Data9(1), 663 (2022)","cited_arxiv_id":null,"evidence_quote":"Supplies the LifeSnaps multi-modal well-being dataset used for the clustering analysis and thesaurus construction."},{"cited_title":"In: Proceedings of the AAAI conference on artificial intelligence","cited_arxiv_id":null,"evidence_quote":"One of the benchmarked XAI methods (Anchor) whose explanations are compared when building the thesaurus, showing the benchmarking breadth behind the ground truth."}],"review_version":1}