{"id":"84f57956-0bf2-44f4-8d32-4ff796be871a","arxiv_id":"2507.21571","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"AI explanations should present the least-supported beliefs in the agent's reasoning, because those are the likely source of user surprise when behavior is unexpected.","lead":"This paper argues that explanations from AI agents used repeatedly over time should surface the specific beliefs or facts where agent and user disagree, not just describe the model's internal reasoning. It proposes a memory-based model that tracks which parts of the agent's knowledge have no support from past interactions and presents those to the user.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Least-support selection ignores counterfactual relevance: a novel-but-inert fact can be chosen over the actual difference-maker, so the explanation need not resolve the user's surprise.","rationale":"The reader's verdict identified support reliability as the weakest assumption. That is real, but it is not the most load-bearing. Even granting perfect knowledge of what the user does not know, the algorithm can pick an epistemically new but causally irrelevant premise. The paper defines the relevant uncommon ground counterfactually in Section 3, then operationalizes it in Section 4 as 'lowest support', without proving that low support implies counterfactual relevance. A simple formal countermodel in the paper's own defeasible setting settles the point. This does not refute the overall extrospective-explanation idea; it shows the model is incomplete and needs a selection step that checks which unsupported items, when revised to the user's belief, actually change the decision. Hence the verdict remains CONDITIONAL: the conceptual proposal is plausible but the central claim is only supported if such a counterfactual relevance filter is added and tested. We partially agree with the reader: the selection rule is the weak point, but the specific failure mode is different and more fundamental.","tokens_in":11188,"tokens_out":7032,"duration_ms":79151,"concrete_test":"Build the birthday-cake example (Figure 4) in a small defeasible-logic reasoner with priorities. Let the robot's derivation of 'store on patio' use two unsupported premises: A = 'outside temperature is cool' and B = 'the cake has chocolate frosting' (B is relevant to serving, not to storage). Give B support 0 and A support 1. Apply the paper's rule: the explanation is B. Then test the Section 3 counterfactual condition: if B were replaced by the user's belief (no chocolate frosting), the robot's conclusion remains 'store on patio'; if A were replaced by the user's belief (outside is warm), the conclusion flips to the user's expected 'store in fridge'. Since the least-supported item fails the counterfactual test while the more-supported item passes it, the selection rule is insufficient for the stated goal.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Section 4, Figure 5) is that for a surprising behavior the agent should present the part of its knowledge that lies in the uncommon ground and has the lowest support from the user/discourse context. Two distinct properties are being conflated. Section 3 correctly defines the explanatory content counterfactually: the item must be such that if it were changed to the user's belief, the agent's prediction would match the user's expectation ('if this had been different and had aligned with the belief of the user, then the prediction of the agent would have been the same as the prediction of the user'). But the generation algorithm in Section 4 never checks this condition; it selects purely on support. Support estimates only whether the user is likely to know/accept an item; it says nothing about whether that item is a difference-maker for the current decision. In a defeasible-rule system, an agent's derivation can contain premises that are not necessary for the chosen conclusion (e.g., redundant support or side conditions). If such an inert fact happens to have the lowest support, the algorithm explains the wrong thing: the user learns a new fact but still cannot see why the agent acted as it did. The paper explicitly acknowledges (Section 4) that several facts/rules can have no support, yet offers no criterion for selecting among them beyond support. Thus the framework does not establish that 'least support' identifies the uncommon ground that explains the surprise; it only identifies content the user probably does not know. This gap holds even if support were a perfect oracle of user knowledge, so it is more fundamental than the reliability of interaction-based support estimates.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a conceptual framework for generating personalized explanations in long-term human–AI interaction. It introduces \"extrospective explanations,\" which are explanations selected not only from the agent's internal reasoning but also from what the agent has learned about the user through prior interactions. The central idea is that, when a user is surprised by an AI behavior, the agent should identify the \"uncommon ground\" — the fact or rule in its reasoning that the user does not share — and present the piece of knowledge with the lowest support from the user and discourse contexts. The framework is built on the SUDO model with situational, user, discourse, and ontological contexts, and is illustrated with a defeasible-logic household robotics example involving a birthday cake. The paper is a position/proposal: it contains no formal algorithm, implementation, or user study.","tokens_in":11471,"tokens_out":4086,"duration_ms":50969,"significance":"If the framework is developed further, it addresses a genuine gap in XAI: local explanations for non-expert users in repeated, personalized interactions. The paper correctly identifies that surprise-driven explanation requests are implicitly counterfactual, and it makes a reasonable case that explanations should be tailored to the user's presumed knowledge. It also has strengths: it explicitly builds on a formal context model (SUDO), uses defeasible logic rather than black-box approximations, and candidly acknowledges several limitations in Section 4. However, the central selection rule — least-supported fact or rule — is not justified as a way to identify the counterfactual difference-maker, and the support model itself is under-specified. The manuscript is more of a well-motivated position statement than a validated model, and its claims should be scaled accordingly.","major_comments":[{"comment":"The selection rule \"present the piece of information with the lowest support\" conflates the likelihood that the user does not know a fact with the counterfactual relevance of that fact. Section 3 defines the uncommon ground counterfactually: the explanation should be an item such that, if it were changed to the user's belief, the agent's prediction would match the user's expectation. But support, as described in Figure 5, only estimates whether the user is aware of or agrees with an item; it says nothing about whether that item is necessary for the agent's conclusion. In a defeasible derivation, a premise can be non-essential — for example, a side condition or a redundant support for a conclusion already entailed by other rules. If such an inert fact happens to have low support, the algorithm would present a novel but irrelevant fact, and the user would still not understand why the agent acted as it did. The paper itself notes (Section 4) that several facts or rules can have no support, but it offers no criterion for choosing among them beyond support. The central claim therefore needs either an additional relevance check (e.g., verify that changing the candidate item flips the agent's conclusion) or a careful restriction to cases where the least-supported item is on the derivation path.","section":"Section 4, Figure 5"},{"comment":"The support model is not formally defined. The star levels in Figure 5 are described in prose (three stars for explicit communication, two for perceived facts, one for agreed-upon facts in past interactions), but there is no precise account of how support is accumulated, updated, combined across user and discourse contexts, or decayed over time. More importantly, the framework treats absence of support as evidence that the user does not know a fact. This is an inference from silence: a fact may have no support simply because it was never mentioned, yet the user may still know it from general world knowledge. Conversely, a user may have privately disagreed during a previous interaction while remaining silent, which would give an unsupported item an incorrect rank. Because the entire selection mechanism rests on this support-as-proxy assumption, the model needs either a formalization of support with explicit update rules or a statement that the support estimate is a heuristic assumption requiring empirical validation.","section":"Section 4, Figure 5"},{"comment":"The running example appears internally inconsistent about what the user knows. The text says the user is not aware that the outside temperature is cool, but also says the user expected the cake to be placed in the hallway \"which they thought was just as cool as the outside.\" If the user believes the hallway is just as cool as the outside, then the user's knowledge includes a belief about hallway temperature that is not listed in the agent's knowledge, and the uncommon ground is not limited to the outside-temperature fact. The Figure 4 caption states that \"the only difference in their respective worldview is that the AI agent is aware that it is quite cool outside today,\" which appears to contradict the narrative. This matters because the example is used to motivate the least-support rule, and the ambiguity makes it unclear which fact the framework is supposed to select.","section":"Section 3, Figure 4"},{"comment":"The paper makes prescriptive claims about the helpfulness of extrospective explanations without empirical support. Section 4 states that \"our approach presents a viable first step at presenting a helpful explanation tailored to the user,\" and Section 5 claims that \"extrospective explanations can provide more helpful information for end-users.\" Since the paper contains no user study, simulation, or formal argument establishing these properties, these claims should be framed as hypotheses or design goals rather than established results. This is a load-bearing issue only to the extent that the paper presents itself as a model; if the authors intend this as a position paper, the claims should be softened accordingly.","section":"Section 4, Section 5"}],"minor_comments":[{"comment":"The affiliation line contains a typo: \"Univserity of Bremen\" should be \"University of Bremen.\"","section":"Title page"},{"comment":"The paper mentions several strategies for discovering the implied counterfactual (explicitly asking, predicting, or using a mental model) but does not connect these to the proposed framework. A brief explanation of why the framework chooses the last option would improve readability.","section":"Section 2"},{"comment":"The paper uses \"most specific rules\" and \"last logical step\" as examples of introspective selection, but it does not define these terms formally in the defeasible-logic setting. A sentence clarifying the intended priority or specificity semantics would help.","section":"Section 4"},{"comment":"The caption says \"three stars in the situational context indicate that a fact has been communicated to the robot by the user,\" but the text in Section 4 says \"The other facts and rules have received varying levels of support, depending on whether or not the information was explicitly communicated by the human.\" The relationship between star levels and the user/discourse context distinction could be stated more explicitly.","section":"Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is a well-written conceptual proposal, but its central contribution — the least-support selection rule — is currently not well-founded. The authors should either formalize the support model and add a relevance test, or reposition the paper as a research proposal with clearly labeled hypotheses. The absence of any evaluation is acceptable for a position paper only if the claims are softened. The self-citation of SUDO is appropriate and not problematic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: a clear, honest conceptual paper that names a real problem and a plausible direction, but the central selection rule conflates rarity with relevance. The stress-test concern lands.\n\nWhat is actual: the 'extrospective' framing—explanation generated from the agent's memory of this user, rather than from its own reasoning alone—is a useful reframing of contrastive explanation. The cake example is effective in showing why salience fails: the most influential fact may already be common ground. Connecting Miller's counterfactual account to Porzel's SUDO context model is a legitimate extension, and the support-star idea is a coherent way to think about accumulating evidence for shared knowledge. The paper is well structured and unusually honest about its limits: it explicitly says several facts or rules can have no support, and it calls its approach a 'viable first step.' The self-citation of SUDO is appropriate—it is the starting framework, not the evidence for the central claim.\n\nSoft spot, and it is the load-bearing one: Section 3 defines the right explanatory content counterfactually—the item whose change would align the agent's prediction with the user's expectation—but Section 4's generation rule selects purely on support. In a defeasible-rule derivation, not every premise is a necessary premise. A fact can be both unknown to the user and irrelevant to why the agent acted. If that inert fact has the lowest support, the algorithm explains the wrong thing. The paper acknowledges ties in support but offers no criterion that checks whether the item is actually a difference-maker. This gap holds even if support is a perfect oracle for user knowledge, so it is more fundamental than the reliability of interaction-based estimates. Fixing it seems doable: restrict candidates to those that appear in an actual derivation of the conclusion, then among those pick the least-supported. But as written, the central claim is not established.\n\nAlso, there is no formal model, implementation, or user study. That is acceptable for a position paper, and the paper is appropriately modest, but it means the contribution is in the framing and the open problem rather than a validated method.\n\nWho should read it: people working on XAI personalization, human-robot explanation, or conversational explanation. They will find a useful vocabulary and a clear problem statement.\n\nRecommendation: I would send this to peer review. It deserves serious referee time, with the selection rule as the main pressure point. It needs revision, not rejection.","headline":"A clear, honest conceptual proposal for personalizing explanations through interaction memory, but its core selection rule needs a counterfactual-relevance check before it does what it claims.","tokens_in":12002,"tokens_out":3219,"would_cite":true,"duration_ms":39505,"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":"For surprising AI behavior, explain the fact or rule the user is least likely to share.","keywords":["extrospective explanations","uncommon ground","personalized explainable AI","SUDO model","counterfactual explanations","human-centered AI","household robotics","user context"],"falsifier":"An empirical study in which users interact with a household agent while privately holding a belief they never express, and the agent later acts on that belief, would test the claim: the least-supported selection would point to some other unsupported item, and users would remain confused about the actual disagreement. Repeatedly observing this confusion would falsify the claim that accumulated support ranks the uncommon ground.","tokens_in":11028,"feed_emoji":"🤖","tokens_out":4607,"duration_ms":51853,"temperature":0.7,"pith_summary":"Explaining a surprising AI decision to a non-expert should not list the features that most influenced the model; it should surface the one belief or fact the user does not already share. The paper calls this the \"uncommon ground\" between user and agent, and argues that an agent can locate it from a dynamic memory of previous interactions with that same user. It proposes an extrospective model of explanation in which the agent's knowledge is organized into situational, user, discourse, and ontological contexts, and pieces of knowledge accumulate \"support\" whenever interactions suggest the user shares them. The resulting rule is that, when asked for an explanation, the agent presents the fact or inference with the least support. If this works, the same AI behavior can be explained differently to different users, and users can immediately see whether they were missing information or disagreeing with the AI.","feed_headline":"For surprising AI, explain the fact the user already knows least","feed_subtitle":"A new model ranks what the agent remembers about you and explains with the least-supported fact.","key_machinery":"The load-bearing object is the SUDO model of the agent's worldview, which separates knowledge into four contexts: situational context (facts about the current state, the Abox), user context (knowledge about the user's abilities, preferences, and state), discourse context (memory of previous requests, actions, and reactions), and ontological context (general rules and commonsense knowledge, the Tbox). User and discourse contexts attach \"support\" to facts and rules in the other two contexts whenever an interaction suggests the user shares that item, such as explicitly communicated facts, facts the user has perceived, or rules used in successful interactions. Explanation generation then becomes a selection problem: among the knowledge used in the reasoning chain that led to the behavior, choose the item with the least support. The paper also uses defeasible logic, in which specific rules can outrank general ones, to model the reasoning itself.","core_discovery":"The paper's central claim is that, for repeated human-AI interaction in the home, the purpose of an explanation is to locate the uncommon ground: the set of facts and rules that the AI used in its reasoning but that the user does not know or accept. Because the exact counterfactual the user had in mind is usually only implied, the agent cannot rely on being told what was expected. Instead, the agent should maintain a personal memory of the user, and explain by presenting the knowledge item with the lowest accumulated support from prior user and discourse contexts. The authors work through a household-robot scenario in which a robot stores a birthday cake outside because the fridge is too small and the patio is cool; which fact is \"uncommon\" differs per user, and only the support model, not salience or rule priority, picks the right one. The paper frames this as a first step: the agent still cannot know the user's mind, and several unsupported items may remain.","pith_inferences":["A testable extension would measure explanation quality by the number of follow-up questions a user needs before the uncommon ground becomes clear.","The support signal conflates \"user did not object\" with \"user agrees\"; weighting silent acceptance less than explicit confirmation would make the ranking safer.","The same uncommon-ground selection could be applied to black-box models by treating feature-attribution or influence scores as the reasoning chain and interaction logs as support.","A user who privately disagrees without saying so is the hard case: the model will rank their disputed belief as supported, so an explanation study should include users instructed to stay silent while disagreeing."],"forward_implications":["For the same agent action, two users can receive different explanations, each pointing at the fact or rule that is new to that user.","The agent does not need to know the user's expected outcome in advance; the least-supported item in the used reasoning chain is a candidate explanation without solving the counterfactual.","Users can tell quickly whether they lacked information or disagree with the agent, and can then correct the agent's knowledge base.","In domains where consistency or completeness matters, such as medical or legal decisions, introspective explanations remain necessary; the approach is aimed at local, personalized surprise explanations.","The support model can be built from interaction history without explicit user profiles, avoiding the configuration burden of direct user input."],"supporting_citations":[{"why":"Supplies the SUDO model of contextual computing that the paper adapts into the agent's four-context personal memory.","marker":"[Porzel, 2010]"},{"why":"Provides the notion of common ground in language use from which the paper derives its concept of uncommon ground.","marker":"[Clark, 1996]"},{"why":"Supports the claim that explanation requests carry an implied counterfactual, i.e., \"why not what I expected\".","marker":"[Miller et al., 2017]"},{"why":"Provides the taxonomy showing that non-expert explanation needs typically arise from surprise.","marker":"[Liao et al., 2020]"},{"why":"Defines the introspective XAI methods that this paper contrasts with extrospective explanations.","marker":"[Adadi and Berrada, 2018]"}],"fun_headline_variants":["Explain what the user knows least, not how the AI thinks","AI explains the fact you don't know yet","Tailor AI explanations to your knowledge gap","Explain the least-known fact to each user","Uncommon ground: explain what the user lacks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The selection rule assumes that how much support a fact or rule accumulated from past interactions reliably indicates whether this user already knows or accepts it; a user can be silent or polite during a successful interaction while privately disagreeing, which would give the wrong item high support.","fun_headline_variants_meta":{"raw":{"variants":["Explain what the user knows least, not how the AI thinks","AI explains the fact you don't know yet","Tailor AI explanations to your knowledge gap","Explain the least-known fact to each user","Uncommon ground: explain what the user lacks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0005,"raw_usage":{"total_tokens":2409,"prompt_tokens":872,"completion_tokens":1537,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":488,"completion_tokens_details":{"reasoning_tokens":1465}},"tokens_in":488,"tokens_out":1537,"duration_ms":11520,"temperature":1.0,"reasoning_tokens":1465,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:36:06.424241+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An empirical study in which users interact with a household agent while privately holding a belief they never express, and the agent later acts on that belief, would test the claim: the least-supported selection would point to some other unsupported item, and users would remain confused about the actual disagreement. Repeatedly observing this confusion would falsify the claim that accumulated support ranks the uncommon ground.","supporting_citations":[{"cited_title":"Using language","cited_arxiv_id":null,"evidence_quote":"Provides the notion of common ground in language use from which the paper derives its concept of uncommon ground."},{"cited_title":"Peeking Inside the Black - Box : A Survey on Explainable Artificial Intelligence ( XAI )","cited_arxiv_id":null,"evidence_quote":"Defines the introspective XAI methods that this paper contrasts with extrospective explanations."}],"review_version":1}