{"id":"9fe0edb2-bd34-4ad4-8c49-1e9ee154fd44","arxiv_id":"2505.03732","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Explanatory goodness is modeled as the expected utility gain a pragmatically reasoning listener gains from a message, yielding virtues like minimality, proportionality, and normality effects as emergent consequences.","lead":"This paper develops a formal account of causal explanation grounded in conversational pragmatics and structural causal models. It argues that classic explanatory virtues and patterns of causal judgment emerge naturally from treating explanations as useful messages for a listener's decisions.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 'emergent' virtues are not derived from a fixed decision problem: most illustrations in §§4.1–4.5 use a bespoke reward function and cost, so the model can prefer opposite causes when R is changed; the central emergence claim is therefore not yet pinned down.","rationale":"The paper is a serious framework and I do not think the central idea collapses. The RSA-style cascade from literal listener through pragmatic speaker to pragmatic listener is coherent, the manipulation game connects naturally to interventionist themes, and the normality/conjunctive-disjunctive prediction in Section 4.3 is a genuine, potentially testable empirical claim. However, the headline claim that classic explanatory virtues 'emerge' rather than being stipulated requires that the model's outputs are not driven by unconstrained choices made separately for each example. Since (A,R) and Cost(m) are free parameters, the framework currently functions more like a language in which examples can be encoded than a substantive theory of explanation. This is the same concern the Reader identified, though I would sharpen it: it is not merely that decision problems are under-specified, but that Section 4 changes R from example to example, so there is no single model that generates all five virtues. The Eq. 8 expected-utility wording also needs correction before the formal measure bears the weight of the central claim. These issues are fixable: one could commit to the manipulation game as the default, provide a theory of how R is derived from the listener's context, and then fit βL, βS, and Cost to data. Until then the verdict should remain conditional, not rejection; the concrete check above would settle whether the manipulation game alone is enough, or whether additional constraints on R are required.","tokens_in":31824,"tokens_out":11395,"duration_ms":138379,"concrete_test":"Re-run the three main illustrations (roof replacement in §4.1.1, milk theft in §4.4, and pigeon proportionality in §4.5) with the decision problem fixed to the manipulation game of Def. 3 and with one common cost function, instead of each example's custom payoff matrix. If the intuitively preferred messages no longer win under Eq. 6/8, the apparent virtues are artifacts of the hand-picked reward functions, and the account needs an independent theory of which decision problems are operative. In the same recomputation, replace Eq. 8's baseline with the proper prior policy πPrior(a) and use posterior-averaged rewards E_{PL(·|m)}[R(a,·)]; if any ordinal comparison reverses, the formal measure itself must be corrected before the empirical claims can be evaluated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing soft spot is the free choice of the listener's decision problem (A,R) and Cost(m). Eq. 8 defines Goodness relative to (A,R), and Eq. 6 lets the speaker choose messages using US and Cost; no independent procedure fixes these from the explanatory situation. Each illustrative virtue is demonstrated with a different payoff table: roof replacement uses Table 1, milk theft uses a confrontation payoff, the pigeon uses a target-purchasing payoff, and only the normality effect is derived from the manipulation game of Def. 3. Because R is unconstrained, the same formal machinery can prefer opposite causes: Section 4.2 constructs a second decision problem in which citing the drought, rather than the thatched roof, becomes the better message in the same causal structure M∧. So the 'virtues' are not shown to emerge from a communication-first core; they are imported through the choice of decision problem. Additionally, Eq. 8's first term is not the listener's expected utility over K: it omits averaging R(a,·) over the posterior PL(·|m), and the baseline term πPrior(a|m) should be a prior policy πPrior(a), not a distribution conditioned on m. This makes the central measure ambiguous. The paper explicitly disclaims a full defense ('Our goal here is not to offer a full-fledged defence', Section 4), but the gap is not merely confidence: without constraints on R and Cost, the account lacks clear falsifiable content.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a formal, rational-speech-act account of causal explanation. A literal listener updates a prior over a set of causal model/context pairs upon hearing a message \"FACT because X=x\"; a pragmatic speaker selects a message to maximize the listener's expected reward minus a message cost; and a pragmatic listener interprets the speaker's choice using a higher-order model. Explanatory goodness is defined in Eq. (8) as the difference between the listener's expected utility with and without the message. The central claim is that classic explanatory virtues—sensitivity to downstream interests, background knowledge, invariance, minimality/simplicity, proportionality, and normality-based causal selection—emerge from this communication-first core rather than being stipulated. The paper supports this claim with worked examples (roof replacement, milk theft, late meeting, pigeon pecking) and contrasts its account with Halpern and Pearl's treatment of explanation.","tokens_in":32168,"tokens_out":7024,"duration_ms":67576,"significance":"If the framework can be adequately constrained, it would offer a valuable bridge between RSA-style formal pragmatics, interventionist accounts of causal explanation, and the empirical literature on causal selection. The paper is clear about its formal machinery, the contrast with Halpern and Pearl is carefully argued, and the normality-based results in §4.3 are genuine derivations from a single, well-defined manipulation game (Def. 3) rather than from bespoke payoff tables. That said, the paper's central claim that the virtues 'emerge' is not currently established: the examples demonstrate that the framework can reproduce these virtues under favorable choices of decision problem and cost, not that they follow from the communication-first core itself. The framework is therefore best read, at this stage, as a promising formal template with illustrative case studies.","major_comments":[{"comment":"The central measure is mis-specified. Goodness(m, M, u) = Σ_a π_L(a|m)·R(a, M, u) − Σ_a π_Prior(a|m)·R(a, M, u) does not express the listener's expected utility as the prose claims. The first term evaluates rewards at the actual context (M, u) rather than averaging R(a, ·) over the posterior P_L(·|m); the expectation inside the softmax defining π_L (Eq. 3) is not the same as an expectation of final reward. The baseline term π_Prior(a|m) is also problematic: it conditions the prior action distribution on the message, even though it is supposed to represent the listener's behavior without any explanation; the intended object appears to be an unconditional prior policy π_Prior(a). Because Eq. (8) defines the account's central quantity, this ambiguity must be corrected before the examples can be interpreted quantitatively.","section":"§3.4, Eq. (8)"},{"comment":"The 'emergence' claim is underdetermined by the free choice of the listener's decision problem (A, R) and of Cost(m). Each virtue is illustrated with a different bespoke payoff matrix or cost function: Table 1 for roof replacement, the confrontation payoff in the milk-theft example, the target-purchasing payoff in the pigeon example, and only the normality effects in §4.3 are derived from the manipulation game. Since R and Cost are unconstrained, the same machinery can prefer opposite causes in the same causal structure under alternative reward specifications (as the paper itself shows in §4.1.3 with flipped payoffs). Without a principled procedure for fixing (A, R) and Cost from the explanatory context—or a restriction of the general claim to the manipulation-game default—the paper has not shown that the virtues emerge; it has shown that they can be represented.","section":"§4.1–§4.5, Tables 1–2; Eq. (6)"},{"comment":"The derivations of minimality, simplicity, and proportionality rest on inequalities for Cost(m) and on payoff matrices that are chosen to be sensitive to exactly the distinction at issue. In §4.4 the paper acknowledges that the precise payoffs are arbitrary, and in §4.5 the preference for 'red' over 'scarlet' follows from a payoff matrix with a separate action for each causal structure. The paper's disclaimer in Section 4 that it is 'not to offer a full-fledged defence' is honest, but the gap is load-bearing: the selling point of the framework is that virtues emerge rather than being stipulated, and the current examples do not rule out equally natural specifications that would reverse the predicted preferences.","section":"§4.4–§4.5"}],"minor_comments":[{"comment":"The second listed utterance is repeated as 'M = 1 because C = 1' ('Dana took the milk'); it should be 'M = 1 because D = 1'.","section":"§4.4, Example 4"},{"comment":"In the text preceding the example, 'the penguin has been trained to peck at a target' should presumably read 'the pigeon'.","section":"§4.5, Example 7"},{"comment":"The text says the normalizing constant Z is a sum over 'all possible messages m'; it should be a sum over worlds (M, u), since Eq. (2) normalizes a distribution over K.","section":"§3.1, after Eq. (2)"},{"comment":"The quantity π_Prior(a|m) is never formally defined. The paper should either define it explicitly or replace it with an unconditional prior policy π_Prior(a).","section":"§3.4, Eq. (8)"},{"comment":"The weighted decision-problem decomposition R(a, M, u) = Σ_i w_i R_i(a_i, M, u) is mentioned informally, but the action space (as a product) and the treatment of the weights are never specified; a formal definition would help.","section":"§3.3"},{"comment":"The redundancy measure US(m, M, u) − max_{m' ≠ m} US(m, M, u) is called a quantification of redundancy, but it ignores message costs; the text should state clearly that it is a cost-free measure.","section":"§4.4"}],"recommendation":"major_revision","confidential_remarks":"The main threat is not technical inconsistency but underdetermination. I would ask the authors to either prove the emergence results under a default manipulation-game decision problem or substantially soften the claim. The paper fits the journal's interest in formal, multi-agent approaches, and the normative/descriptive gap in the account should be acknowledged more explicitly in the final version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth reading: it gives the most systematic attempt I know to get explanatory virtues out of a decision-theoretic pragmatic framework, rather than stipulating them. The core idea—explanatory goodness as expected utility gain in a rational speech act setting—is coherent and the worked examples are illuminating. The critical discussion of Halpern-Pearl is sharp, and the manipulation game (Def. 3) is a genuinely useful default decision problem. The strongest result is Section 4.3: the normality-structure interaction (abnormal causes in conjunctive structures, normal in disjunctive) follows from the manipulation game plus a prior inequality, not from a bespoke payoff table. That is real derivation.\n\nThe soft spots are real but not fatal. First, Eq. 8, the central definition of Goodness, is formally ambiguous: the first term multiplies π_L(a|m) (which already averages over the posterior) by R(a, M, u) for the actual world, and the baseline uses π_Prior(a|m), a conditional-on-m object that doesn't fit the prose ('if he hadn't received any explanation'). This can be cleaned up, but as written it obscures what is being measured. Second, the 'emergence' claim is weaker than it appears. Each virtue in §§4.1–4.5 is demonstrated with a hand-picked payoff matrix and cost function; Section 4.2 even shows that changing the decision problem flips which cause is better. That's not a bug—context-sensitivity is the point—but it means the virtues emerge only relative to an unconstrained choice of (A,R) and Cost. The manipulation game is a principled default for one case, but the paper doesn't say how to fix these inputs in general. As the authors admit, they are not offering a full defense; the gap matters because without constraints on R and Cost the account has little falsifiable content.\n\nThis is a paper for philosophers of explanation and cognitive scientists who want a precise framework to argue with. It deserves a serious referee. My recommendation: send it out, with the request that the authors clarify Eq. 8, and say something more about where decision problems and costs come from—ideally, a default that covers more cases.","headline":"A serious and readable attempt to derive explanatory virtues from RSA-style pragmatics; the normality result is solid, but the central goodness measure has a formal glitch and the 'emergence' story leans on unconstrained decision problems.","tokens_in":32641,"tokens_out":3535,"would_cite":true,"duration_ms":36445,"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":"The paper proposes that explanatory goodness is the expected-utility gain a pragmatic listener gets from a message, and shows that classic explanatory virtues follow.","keywords":["causal explanation","conversational pragmatics","expected utility","manipulation game","causal selection","normality","proportionality","pragmatic speaker-listener reasoning"],"falsifier":"Run a matched causal-selection study with two causal structures and a listener decision problem whose payoff matrix is sensitive to one cause but not the other, as in the roof-replacement design. The model predicts the speaker will cite the cause the decision problem is sensitive to even when both causes are equally likely and equally informative; if human speakers instead pick the statistically abnormal cause regardless of the decision problem, the central claim fails.","tokens_in":31625,"feed_emoji":"💬","tokens_out":9988,"duration_ms":92913,"temperature":0.7,"pith_summary":"Causal explanations are good, the paper claims, when they help the listener decide and act. The proposed measure of explanatory goodness is the increase in expected utility a listener obtains from a message, compared with acting on what he already believed: a message is good exactly when it moves the listener from worse actions to better ones. From this single quantity, together with a default 'manipulation game' in which the listener picks a variable to intervene on, the paper derives the classic explanatory virtues: attention to the listener's interests and knowledge, invariance across background conditions, minimality and simplicity, proportionality between cause and effect, and normal-vs-abnormal causal selection. Actual causation still anchors the account, but only as the literal content of 'FACT because X=x'; everything else about explanation is treated as communication. If the account succeeds, philosophical questions about explanation and psychological findings about causal judgment are two sides of one expected-utility calculation.","feed_headline":"Explanations are good when they raise the listener's expected payoff","feed_subtitle":"A formal account derives minimality, proportionality, and normality effects from communication and utility, not from definition","key_machinery":"The load-bearing machinery is a pragmatic speaker-listener hierarchy combined with a reward-sensitive decision problem. The literal listener updates on the semantic content that $X=x$ is an actual cause of FACT; the pragmatic speaker chooses an utterance to maximize the expected reward this literal listener will obtain, minus a message cost; the pragmatic listener interprets the message as evidence about what the speaker knew and chose. The key auxiliary object is the manipulation game: a decision problem in which the listener's action is choosing an endogenous variable to intervene on, and the reward is the probability, over background contexts, that intervening on that variable switches the truth of FACT. This game converts the abstract idea that explanations support what-if-things-had-been-different questions into a concrete payoff, and it is what makes normality and causal-structure effects derivable. The central identity defining the account is the goodness equation, which evaluates explanations as the difference between expected utility after and before the message.","core_discovery":"The paper's central claim is that explanatory goodness is a utility increase. Fix a listener with a prior over causal situations, a decision problem with actions $A$ and rewards $R$, and a message $m$ of the form 'FACT because $X=x$'. After hearing $m$, the listener updates to a posterior that reflects both the literal truth that $X=x$ is an actual cause of FACT and the pragmatic fact that the speaker chose $m$. Writing $\\pi_L(a\\mid m)$ for the listener's action policy after the message and $\\pi_{\\mathrm{Prior}}(a\\mid m)$ for the policy before any message, $$\\mathrm{Goodness}(m,M,u)=\\sum_{a\\in A}\\pi_L(a\\mid m)\\,R(a,M,u)-\\sum_{a\\in A}\\pi_{\\mathrm{Prior}}(a\\mid m)\\,R(a,M,u).$$ A message is good exactly when it lets the listener act better than he would have acted. The paper argues that this one quantity, together with a manipulation game—a default decision problem in which the listener chooses an endogenous variable to intervene on so as to change the truth value of FACT—makes the recognized explanatory virtues emerge. Minimality is not assumed but results from message costs; sensitivity to what the listener already knows follows from the speaker's aim to be useful; proportionality follows from the listener's decision problem being sensitive to the level of the cited variable; and the empirical pattern that abnormal causes are cited in conjunctive structures while normal causes are cited in disjunctive structures is derived from the manipulation game's payoff matrix.","pith_inferences":["If goodness is defined by expected-utility gain, then measuring explanation quality in practice requires fixing the listener's decision problem; the paper offers the manipulation game as a default, but any empirical test must decide when that default applies.","The framework suggests a quantitative bridge to causal-selection experiments: the manipulation game's reward weighting matches existing $\\Delta P$ and counterfactual-effect-size measures, so model parameters could be fit to data on causal judgments.","One testable extension is to make the speaker's message cost an explicit function of listener processing difficulty, such as reading time or misinterpretation rate, turning the simplicity discussion into a measurable prediction.","For scientific explanation, the account would need to treat a scientific community as a single listener with aggregated knowledge and aims; the formalism gives a route to do this, but the paper leaves that aggregation unspecified."],"forward_implications":["Explanations become graded and listener-relative: the same sentence can be good for one listener and bad for another, depending only on that listener's decision problem and prior.","Minimality is a cost-driven preference, not a definitional rule: when a longer message removes more relevant uncertainty than its extra cost, the longer explanation is the better one.","Proportionality and level selection follow from decision problems: speakers cite 'red' rather than 'scarlet' when the listener's actions are sensitive to the coarser variable.","The model predicts and explains the causal-selection pattern: abnormal causes in conjunctive structures, normal causes in disjunctive structures, and listeners' ability to infer structure and normality from the speaker's choice.","Known causes can still explain: citing a cause the listener already knows can be informative, because the speaker's choice of that cause reveals which causal structure is actual, as in the late-meeting example."],"supporting_citations":[{"why":"Supplies the actual-cause-based analysis of explanation whose EX2-EX4 conditions the paper argues should emerge from pragmatics rather than be stipulated.","marker":"Halpern and Pearl (2005a)"},{"why":"Provides the interventionist what-if-things-had-been-different idea and the invariance requirement that the manipulation game formalizes.","marker":"Woodward (2003)"},{"why":"States the characterization of explanation as providing resources for answering what-if questions, which the paper takes as its starting point.","marker":"Woodward and Hitchcock (2003)"},{"why":"Gives the basic pragmatic speaker-listener model used to build the literal listener and pragmatic speaker.","marker":"Frank and Goodman (2012)"},{"why":"Provides the review of the speaker-listener hierarchy that the paper's three-level model follows.","marker":"Goodman and Frank (2016b)"},{"why":"Extends the pragmatic speaker model to reward-sensitive listeners, which supplies the decision-theoretic speaker utility.","marker":"Sumers et al. (2023)"},{"why":"Reports the causal superseding pattern that motivates the normality-by-structure predictions.","marker":"Kominsky et al. (2015)"},{"why":"Provides the empirical link between normality and causal strength that the manipulation game is designed to match.","marker":"Icard et al. (2017)"},{"why":"Provides evidence that listeners draw structural and normality inferences from the speaker's choice of cause.","marker":"Kirfel et al. (2022)"},{"why":"Supplies the counterfactual effect size model whose reward structure the manipulation game resembles.","marker":"Quillien and Lucas (2023)"}],"fun_headline_variants":["Explanations are utility boosts for listeners","Good explanations raise the listener's payoff","Why explanations are good: utility increase","A communication-first model of explanation goodness","Explanatory value equals listener's expected gain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every 'why?' question can be paired with a specified decision problem for the listener, or with the manipulation game as a default, and that each message has a known cost; choose different goals or costs and the derived virtues change.","fun_headline_variants_meta":{"raw":{"variants":["Explanations are utility boosts for listeners","Good explanations raise the listener's payoff","Why explanations are good: utility increase","A communication-first model of explanation goodness","Explanatory value equals listener's expected gain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000258,"raw_usage":{"total_tokens":1595,"prompt_tokens":972,"completion_tokens":623,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":559}},"tokens_in":588,"tokens_out":623,"duration_ms":6143,"temperature":1.0,"reasoning_tokens":559,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:43:46.837143+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a matched causal-selection study with two causal structures and a listener decision problem whose payoff matrix is sensitive to one cause but not the other, as in the roof-replacement design. The model predicts the speaker will cite the cause the decision problem is sensitive to even when both causes are equally likely and equally informative; if human speakers instead pick the statistically abnormal cause regardless of the decision problem, the central claim fails.","supporting_citations":[{"cited_title":"Reconciling truthfulness and relevance as epistemic and decision-theoretic utility","cited_arxiv_id":null,"evidence_quote":"Extends the pragmatic speaker model to reward-sensitive listeners, which supplies the decision-theoretic speaker utility."},{"cited_title":"Causal superseding","cited_arxiv_id":null,"evidence_quote":"Reports the causal superseding pattern that motivates the normality-by-structure predictions."},{"cited_title":"Inference from explanation","cited_arxiv_id":null,"evidence_quote":"Provides evidence that listeners draw structural and normality inferences from the speaker's choice of cause."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the counterfactual effect size model whose reward structure the manipulation game resembles."}],"review_version":1}