{"id":"c5193c58-99df-4e72-b016-08caf263af07","arxiv_id":"2507.17761","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An interactive agent co-constructs provenance-grounded explanations with users, evaluated via LLM-simulated users and an LLM-as-a-judge.","lead":"The authors propose an interactive chatbot that co-builds explanations of AI outputs with users, grounding the conversation in data provenance. The paper presents a prototype and an evaluation framework that uses simulated users and an LLM-as-a-judge to scale up testing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The evaluation framework never measures provenance grounding: none of the seven judge criteria in §4 checks agent claims against the provenance metadata, so Table 1 cannot support the 'grounded in data provenance' half of the central claim.","rationale":"The reader's weakest assumption concerns the reliability of LLM personas and the LLM judge. That is a real external-validity issue. My concern is more specific and more damaging to the central claim: even granting perfect judge reliability, the framework lacks a construct-valid measure of provenance grounding. In this sense, it is the single most load-bearing issue because it blocks the 'grounded in data provenance' half of the claim independently of simulation fidelity. The paper is otherwise honest: it calls itself a vision paper, open-sources its code, and lists future validation steps. The prototype and the dialogue trace in Figure 3 are suggestive but not evidence for the second conjunct. A CONDITIONAL verdict remains right: the vision is worth pursuing, but the evaluation framework must be extended with a provenance-grounding metric and validated, before the reported scores can support the central claim.","tokens_in":7012,"tokens_out":3757,"duration_ms":34717,"concrete_test":"Add a provenance-grounding check to the framework: for each assistant turn, extract all provenance-dependent claims (class expression, positive/negative examples, data sources, criteria), and automatically compare them against the metadata that was injected into the prompt; compute claim-level precision/recall. Then correlate these grounding scores with the LLM judge's 'Correctness & Fidelity' ratings across the 60 dialogues (6 personas × 10 class expressions). If the correlation is weak or grounding accuracy is low, the evaluation framework does not validate the 'grounded in provenance' part of the central claim, and Table 1 should be reinterpreted as measuring dialogue quality only.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two parts: the co-constructed explanation is useful to the user and grounded in data provenance. The evaluation framework in §4 only operationalizes the first part. The seven judge criteria — Clarity & Structure, Depth & Completeness, Correctness & Fidelity, Relevance & Focus, Appropriateness for the Persona, Transparency, Engagement & Intuition — never ask whether assistant statements match the provenance metadata that was given as input. 'Correctness & Fidelity' is defined as accurate facts and faithful to 'the original query/context,' not faithful to the provenance data. The judge is a separate LLM; nothing in the description says it receives the metadata as ground truth. Consequently, Table 1's high 'Correctness & Fidelity' scores can at most indicate fluent dialogue, not provenance-grounded explanation. The paper's own caveats — §5 says personas are 'meant to broadly indicate a role and not necessarily substitute for true human behaviour,' and §6 defers correlation with application-grounded metrics to future work — show the framework's validity is unestablished. This is an internal gap: the central claim's second conjunct is not measured at all.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an interactive explainer agent that co-constructs explanations of an AI system's outputs with the user, using data provenance as the grounding source. It describes a prototype ENEXA chatbot for neural class expression learning, in which provenance metadata is added to prompts and an LLM generates responses. It also introduces an evaluation framework based on LLM-simulated user personas and an LLM-as-a-judge, and reports an experiment with six personas, ten class expressions, and three-turn dialogues, scored on seven criteria. The authors position the work as a vision statement with an initial prototype and a scalable evaluation framework, not as a finished deployment.","tokens_in":7162,"tokens_out":2999,"duration_ms":30851,"significance":"If the framework were validated, this work would be a useful contribution to trace-based XAI by operationalizing co-construction of explanations and by proposing a scalable, reproducible evaluation method that combines user simulation with LLM judging. The paper makes a good-faith effort to connect to existing evaluation taxonomies and to acknowledge limitations; it also publicly links the source code and prompts, which supports reproducibility. The prototype illustrates a plausible mechanism for tailoring provenance-based explanations to different user roles. However, the current evaluation does not establish the central two-part claim, because provenance grounding is never directly measured and the LLM-based evaluation instruments are not calibrated against human judgements.","major_comments":[{"comment":"None of the seven criteria in Section 4 (Clarity & Structure, Depth & Completeness, Correctness & Fidelity, Relevance & Focus, Appropriateness for the Persona, Transparency, Engagement & Intuition) checks whether the assistant's statements are consistent with the provenance metadata. The definition of 'Correctness & Fidelity' is faithfulness to 'the original query/context,' not to the provenance data, and the judge is not described as receiving the metadata as ground truth. Consequently, the high 'Correctness & Fidelity' scores in Table 1 support at most dialogue fluency and topical relevance, not the 'grounded in data provenance' half of the abstract's central claim. The evaluation protocol should either add an explicit provenance-grounding criterion, have the judge verify statements against a provenance store, or both.","section":"Section 4 and Table 1"},{"comment":"The evaluation framework's validity depends on two unvalidated proxies: LLM-simulated personas and an LLM-as-a-judge. Section 4 cites prior work showing agreement between LLM judges and humans and cites user-simulation methods, but the paper does not calibrate its own judge or personas against human judgements, and it does not include a baseline or comparison system. The experiment itself is small (six personas, ten class expressions, three turns) and Table 1 reports only means and standard deviations without any uncertainty or significance assessment. The paper's own caveats in Section 5 ('not necessarily substitute for true human behaviour') and Section 6 (deferring correlation with application-grounded metrics to future work) are appropriate, but they mean the results should be described as illustrative, not as evidence that co-construction works.","section":"Sections 4 and 5"},{"comment":"The claim that the framework enables 'large-scale reproducible evaluations' is stronger than the evidence supports. Reproducibility is aided by the public code, but the evaluation criteria lack a rubric tying each 5-point score to observable dialogue properties, and the framework's validity for ranking different explanation approaches is not established. Adding a concrete scoring rubric, a small human-judgement validation study, and at least one baseline would make the framework's claims testable.","section":"Section 6"}],"minor_comments":[{"comment":"The word 'interatively' in the third sentence of the introduction appears to be a typo for 'interactively.'","section":"Section 1"},{"comment":"The table header contains 'Relevances,' which should be 'Relevance' to match the other criterion names.","section":"Table 1"},{"comment":"The seven criteria are listed with one-sentence definitions, but no rubric is provided for assigning values on the 5-point scale; specifying what distinguishes a score of 3 from a 4 on each criterion would improve reproducibility.","section":"Section 4"},{"comment":"In the example trace, the assistant states that information 'was verified from Wikipedia,' but the provenance metadata mentioned in Section 3 concerns the information extraction pipeline; the example should make clear which provenance records support this claim.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuine vision paper with a working prototype, but the evaluation framework doesn't measure what the abstract claims. The idea is worth engaging; the evidence is not.\n\nWhat's new: building on the co-construction account from [22] and provenance-based XAI from [14], the paper contributes a concrete chatbot that mixes provenance metadata with LLM context to tailor explanations through dialogue. The specific system and the evaluation approach — LLM-simulated personas interacting with an explainer, scored by an LLM judge — go beyond prior work. The prototype is real, the code and prompts are open, and the example dialogue in Figure 3 shows the intended interaction clearly. The paper is also honest about its limits: it says personas \"broadly indicate a role\" and defers correlation with application-grounded metrics to future work.\n\nThe soft spot is load-bearing. The central claim is that explanations are both \"useful to the user\" and \"grounded in data provenance\". The evaluation framework in Section 4 operationalizes the first half but not the second. None of the seven judge criteria ask whether assistant claims match the provenance metadata. 'Correctness & Fidelity' is defined as accurate facts and faithfulness to the original query/context, not to the provenance input. The judge is a separate LLM, and nothing in the description says it receives the metadata as ground truth. So Table 1's high Correctness scores can at most indicate fluent dialogue, not provenance grounding. This is an internal gap: the paper's own setup cannot support half of its abstract claim.\n\nOther issues are in proportion. The sample is tiny — six personas, ten class expressions, three turns. There is no baseline, no human comparison, and the LLM judge is not calibrated against human ratings. The paper cites the LLM-as-a-judge literature but doesn't address how its own judge might inherit biases. These are known limitations of the framework, and the paper acknowledges most of them, but it still calls the framework \"scalable reproducible evaluations\" without evidence that the scores mean anything.\n\nWho is this for? Researchers working on provenance-based XAI or interactive explanation systems. A reader interested in the vision and the prototype will get value; a reader looking for validated evaluation methodology will not. I'd send it to peer review — the idea deserves serious referee time and the prototype is a real starting point — but I'd expect heavy revision. The evaluation framework needs either human validation or a direct check of assistant claims against the provenance store before the paper's central claim is supported.","headline":"Honest vision paper with a working prototype, but the evaluation framework never actually checks the 'grounded in provenance' half of the central claim.","tokens_in":7758,"tokens_out":1477,"would_cite":false,"duration_ms":15650,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that AI explanations should be co-constructed in dialogue, anchored by data provenance so they stay faithful to how the system actually produced its result.","keywords":["co-construction of explanations","provenance","explainable AI","LLM-as-a-judge","user simulation","dialogue evaluation","trace-based explanations","interactive explanation agent"],"falsifier":"Run a validation study where the same chat traces are scored by both the LLM judge and a panel of human raters matched to the six persona profiles, then check the correlation on each criterion. Alternatively, deploy the chatbot with real users in those roles and measure whether the explanations improve task success or user understanding; low correlation or poor real-user performance would undercut the central claim.","tokens_in":6751,"feed_emoji":"💬","tokens_out":3327,"duration_ms":30917,"temperature":0.7,"pith_summary":"The paper argues that explanations of complex AI systems should be co-constructed in dialogue rather than delivered as a single output, and that data provenance should anchor that dialogue so the resulting explanation is both tailored to the user and faithful to how the system actually produced its result. To make the idea concrete, the authors built a prototype chatbot that explains the outputs of a neural class expression learner by feeding provenance into an LLM prompt and letting users ask follow-up questions. They also propose a scalable evaluation framework in which LLM-simulated user personas chat with the explainer and an LLM judge scores the traces. If the approach works, provenance-based explanations could be adapted to many user types and many machine-learning tasks without expensive human evaluation. The paper is a vision-and-prototype contribution: it establishes a scenario and a measurement method, not a field-tested system.","feed_headline":"Dialogue co-builds AI explanations from data provenance","feed_subtitle":"Prototype and LLM-based evaluation aim to tailor explanations to each user without losing the workflow trace.","key_machinery":"The central object is the ENEXA explanation chatbot's dialog engine, which mediates between the user and an LLM by maintaining chat history and injecting provenance metadata, namely the underlying data sources and the information extraction procedure used to learn the class expression, into the prompt. This lets the LLM mix provenance with its own background knowledge and tailor the level of detail to the user. The evaluation machinery is the automated framework that pairs LLM-simulated user personas with a separate LLM judge scoring seven criteria (clarity, depth, correctness and fidelity, relevance and focus, persona appropriateness, transparency, and engagement) on a five-point scale.","core_discovery":"On the paper's own terms, the central claim is that provenance can be the grounding for an interactive co-construction process: the user and an explanation agent together shape what is explained, while the underlying workflow trace keeps the explanation accurate. The authors demonstrate this with ENEXA, a chatbot that verbalizes class expressions learned from positive and negative examples, injects provenance about data sources and the information extraction procedure into the LLM prompt, and answers follow-up questions so the explanation narrows onto what the user actually needs. They report that in automated evaluation over six personas and ten class expressions, the explainer scored well on correctness and fidelity and on relevance and focus, with more mixed results on engagement and depth. The paper's second claim is that the combination of LLM user simulation and an LLM-as-a-judge offers a route to large-scale, reproducible evaluation of such co-constructive explanation systems.","pith_inferences":["A testable extension is to validate the LLM judge against human raters on the same chat traces; if agreement is high, the framework becomes a practical screening tool for explanation design.","The persona set could grow into a reusable battery of hundreds of personas, as the paper's vision section anticipates, enabling standardized comparison across different explanation agents.","The approach likely transfers beyond class expression learning to any pipeline with recorded provenance, such as data preparation and feature engineering workflows, but the paper does not demonstrate that transfer.","Because the chatbot mixes provenance with the LLM's background knowledge, there is a risk of drift from the record; a human-factors study would reveal whether users notice such drift and whether it matters."],"forward_implications":["If the central claim is correct, provenance-based explanations can be personalized in dialogue without sacrificing faithfulness to the underlying computation.","The evaluation framework makes it practical to compare explanation strategies across many personas and many tasks at scale.","The approach extends trace-based explanation to interactive settings, letting developers and product owners interrogate the entire data-to-result pipeline rather than just a model's features.","The seven scoring criteria offer a concrete template for measuring the quality of co-constructed explanations.","The identified weaknesses in engagement and depth point to longer dialogues and richer provenance as the immediate next improvements."],"supporting_citations":[{"why":"Supplies the LLM-as-a-judge method that the evaluation framework adopts, with reported high agreement between LLM agents and human experts in multi-turn dialogues.","marker":"[28]"},{"why":"Provides the systematic review of dialogue evaluation methods and user simulation that motivates the automated evaluation approach.","marker":"[4]"},{"why":"Defines neural class expression learning, the machine-learning scenario used in the prototype and experiments.","marker":"[7]"},{"why":"Establishes trace-based explanations, the type of explanation this work builds on and extends to dialogue.","marker":"[5]"},{"why":"Supplies the conceptual framing of explanation as a co-constructed social practice, which the paper's vision is built around.","marker":"[22]"},{"why":"The closest prior work on explanation as exploration of data pipelines, which the paper distinguishes by its focus on user tailoring.","marker":"[6]"},{"why":"Supplies the dialogue-systems evaluation background, including task success and user simulation, that the framework draws on.","marker":"[8]"}],"fun_headline_variants":["Provenance-grounded dialogue builds AI explanations","Interactive agent co-builds explanations from provenance","Provenance chats tailor AI explanations to users","Co-constructing AI explanations with provenance-based dialogue","LLM-agent co-construction grounds explanations in provenance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation framework's scores are assumed to stand in for real human judgments; if LLM-simulated personas and the LLM judge do not reflect actual users, the reported successes do not show that co-construction works for humans.","fun_headline_variants_meta":{"raw":{"variants":["Provenance-grounded dialogue builds AI explanations","Interactive agent co-builds explanations from provenance","Provenance chats tailor AI explanations to users","Co-constructing AI explanations with provenance-based dialogue","LLM-agent co-construction grounds explanations in provenance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00081,"raw_usage":{"total_tokens":3490,"prompt_tokens":819,"completion_tokens":2671,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":435,"completion_tokens_details":{"reasoning_tokens":2599}},"tokens_in":435,"tokens_out":2671,"duration_ms":18013,"temperature":1.0,"reasoning_tokens":2599,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:59:06.196192+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a validation study where the same chat traces are scored by both the LLM judge and a panel of human raters matched to the six persona profiles, then check the correlation on each criterion. Alternatively, deploy the chatbot with real users in those roles and measure whether the explanations improve task success or user understanding; low correlation or poor real-user performance would undercut the central claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LLM-as-a-judge method that the evaluation framework adopts, with reported high agreement between LLM agents and human experts in multi-turn dialogues."},{"cited_title":"Evaluating Task-oriented Dialogue Systems: A Systematic Review of Measures, Constructs and their Operationalisations","cited_arxiv_id":"2312.13871","evidence_quote":"Provides the systematic review of dialogue evaluation methods and user simulation that motivates the automated evaluation approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes trace-based explanations, the type of explanation this work builds on and extends to dialogue."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the conceptual framing of explanation as a co-constructed social practice, which the paper's vision is built around."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The closest prior work on explanation as exploration of data pipelines, which the paper distinguishes by its focus on user tailoring."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the dialogue-systems evaluation background, including task success and user simulation, that the framework draws on."}],"review_version":1}