{"id":"cb303afd-db01-410d-9f57-81706d009750","arxiv_id":"2607.03425","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A HITL Bayesian procedure estimates user-specific linear SCMs from noisy intervention feedback and yields lower-cost, more valid causal recourse than non-causal baselines under linear synthetic settings.","lead":"The paper proposes a human-in-the-loop Bayesian method that estimates each user's structural causal model from intervention queries, then uses that model to generate personalized algorithmic recourse. It matters because standard recourse either ignores causality or assumes a shared known causal model that rarely exists for real individuals.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The soft likelihood in Section 4 can accept many wrong SCMs that still match noisy responses within σ, so estimated-model recourse may track the oracle only under the paper's own generative assumptions.","rationale":"The Reader correctly flags the restrictive linear-ANM + known-topology assumption as the weakest modelling premise and therefore issues a CONDITIONAL verdict. That premise is indeed necessary, yet it is not the most immediate threat to the strongest claim. Even inside the linear-Gaussian class the soft likelihood of Section 4 only enforces approximate matching of a fraction of the observed responses; it does not guarantee that the recovered structural coefficients are close enough to \theta GT for the counterfactual map used by Eq. (2) to remain faithful. Because the reported simulations generate responses from the same family that is later used for evaluation, they cannot detect this identification failure. The concrete test isolates exactly that failure mode while remaining fully reproducible with the released code. If the test shows that a non-negligible posterior mass yields high parameter error yet still produces high validity under M̃ (and low validity under MGT), the central empirical claim weakens and the paper stays conditional for a sharper reason. If the test fails to find such mass, the Reader's original concern about model-class mismatch remains the binding limitation and the verdict is unchanged. In either case the appropriate recommendation is still CONDITIONAL; the stress-test merely sharpens the precise condition that must be verified.","tokens_in":11294,"tokens_out":902,"duration_ms":8321,"concrete_test":"Fix M1, \theta GT, eta=0.1, \tau=0.2 (or the exact σ used for Fig. 2), T=10. Draw 200 independent response sets D under the paper's mixture model with α=0. Draw 500 posterior samples \thetã via Zeus. For each sample compute both (i) parameter error ||\thetã-\theta GT|| and (ii) the validity of the recourse action obtained from that \thetã when the action is executed on the true MGT. Report the fraction of posterior mass for which validity on MGT falls below 0.8 while ||\thetã-\theta GT||>0.5. If that fraction exceeds 20%, the soft likelihood does not concentrate on SCMs that preserve oracle-level recourse.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (Definition 1 + RQ1) is that posterior samples of a linear ANM recovered from T noisy intervention-response pairs produce recourse whose validity and cost, when re-evaluated on the unobserved ground-truth SCM, closely track the oracle and beat the non-causal prior. The inference step that must deliver this is the soft likelihood of Section 4: P(D|\thetã) = 1{ (1/N)∑ 1(Δk(\thetã)≤σ) ≥ 1-β }. Because the indicator only requires that a (1-β) fraction of the reported descendant values lie inside an ℓ2-ball of radius σ around the model's prediction, any parameter vector whose soft interventions stay inside that ball is assigned equal likelihood. Under the paper's own user-response model (mixture of \theta GT and adversarial \theta*), many such vectors exist that are far from \theta GT yet still satisfy the indicator for the modest T=10 and the σ values used in the figures. Consequently the MCMC posterior can place mass on SCMs whose induced counterfactual map CF(·,a;M̃) differs from the true map; the subsequent gradient-based solver of Eq. (2) then returns actions that look good under M̃ but need not remain valid or low-cost once transferred to MGT. The linear-Gaussian experiments (Figs. 2-3) never isolate this gap: the same generative process that produces the responses is also used to evaluate the recovered model, so agreement is partly by construction. The claim therefore rests on an untested identification property of the soft likelihood rather than on demonstrated recovery of the causal parameters that actually govern recourse.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The paper formalizes personalized causal recourse (Definition 1) under uncertainty about a user's true SCM and proposes a HITL Bayesian pipeline that estimates a surrogate linear additive-noise model from T noisy intervention-response pairs via MCMC (soft likelihood in Section 4), then solves robust causal recourse (Eq. 2) on the estimated model. As a controlled proof-of-concept it evaluates the pipeline on three synthetic three-variable SCMs (linear M1, mixture-noise M2, non-linear M3) with simulated user responses under a Bernoulli mixture model (Eq. 4). For linear M1 the estimated SCM produces validity and cost that track the ground-truth oracle and improve over a non-causal prior baseline (Figs. 2-3); performance degrades under model mismatch while still often remaining cheaper than the prior; intervention patterns are partially recovered (Figs. 4-5). Code is released.","tokens_in":11733,"tokens_out":1206,"duration_ms":22560,"significance":"If the approach scales beyond the current restrictive assumptions, it would remove a central practical obstacle in causal algorithmic recourse—the requirement of a known global SCM—by eliciting user-specific causal mechanisms and thereby producing more plausible, lower-cost interventions. The formalization of the task, the honest reporting of degradation under mixture noise and non-linearity, and the public code release are genuine strengths. The contribution is a clean proof-of-concept for HITL causal personalization; its immediate significance is bounded by the three-variable linear-Gaussian setting and purely synthetic evaluation, but the direction is valuable for high-stakes XAI.","major_comments":[{"comment":"The soft likelihood (Section 4) is defined as the indicator P(D|θ̃)=1{(1/N)∑1(Δk(θ̃)≤σ)≥1-β}. Any parameter vector whose soft-intervention predictions stay inside an ℓ2-ball of radius σ for a (1-β) fraction of the T=10 responses receives identical likelihood. Under the paper's own mixture response model (Eq. 4) many such vectors exist that are distant from θGT yet still satisfy the indicator. Consequently the MCMC posterior need not concentrate on models whose counterfactual map CF(·,a;M̃) coincides with the true map; the subsequent solver of Eq. (2) can therefore return actions that look good under M̃ but lose validity or cost once transferred to MGT. The linear experiments (Figs. 2-3) never isolate this gap because the same generative process supplies both the responses and the evaluation oracle. An ablation that injects held-out response noise or constructs alternative SCMs that match","section":"Section 4, soft likelihood"},{"comment":"The surrogate (structural equations (3)) and the estimation procedure assume known topological order of the ground-truth SCM, independent standard-Gaussian exogenous noise, and no unobserved confounders. These assumptions are load-bearing: when they are violated (M2 mixture noise, M3 non-linear mechanisms) validity drops markedly while the prior baseline gap narrows. The central empirical claim of RQ1 therefore holds primarily inside the paper's own generative class rather than more generally. Either the surrogate class must be relaxed or additional experiments must delineate the regimes in which the linear-Gaussian approximation remains useful for recourse.","section":"Section 4, structural equations (3)"},{"comment":"All feedback is generated by the stylized mixture model (Eq. 4); no real-user study is performed. While the manuscript correctly labels the work a proof-of-concept, the validity of the response model itself (Bernoulli mixture of true and adversarial parameters) is untested against actual human causal judgments, which frequently exhibit structured biases, anchoring, or coherent but incorrect beliefs. This leaves the practical utility of the HITL loop empirically unsupported beyond the synthetic regime.","section":"Section 5 / User response model"}],"minor_comments":[{"comment":"Paolo Giudici's ORCID is given as 0000-0000-0000-0000, an obvious placeholder that should be corrected or removed.","section":"Author list"},{"comment":"Section 5.1 describes qualitative recovery of θ but supplies no quantitative coefficient error (e.g., MSE or posterior credible intervals) or posterior visualizations; adding these would tighten the link between estimation quality and the recourse results of RQ1.","section":"Section 5.1"},{"comment":"Minor typographical issues: 'apriori' → 'a priori'; missing spaces after 'e.g.,' and 'i.e.,'; occasional run-on sentences in the discussion of M3.","section":"Throughout"},{"comment":"Figures 2-5 report means ± std over five runs; adding the raw per-run values or a small table of numerical validity/cost would aid reproducibility checks.","section":"Figures 2-5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a short, honest proof-of-concept; it would fit a workshop or short-paper track more naturally than a full journal article, but the formalization is clean and the code release is a plus. The third-author ORCID placeholder is odd and should be fixed before any public version. Novelty relative to PEAR/HIP-CORE (causal mechanisms versus cost/preference elicitation) is adequately distinguished."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing worth knowing is that this paper cleanly formalizes personalized causal recourse under unknown user SCMs and shows a workable Bayesian HITL estimator that, for linear additive-noise models, recovers actions whose validity and cost on the held-out ground-truth SCM track the oracle and beat a non-causal prior. That combination is new relative to Karimi-style causal recourse (known SCM) and PEAR/HIP-CORE-style personalization (costs/preferences, not structural equations).\n\nWhat they do well: Definition 1 is crisp, the evaluation protocol is honest (actions are always re-applied to M_GT, not to the estimated model), and they report the degradation under mixture noise (M2) and non-linearity (M3) without spinning it. Code and data are released. The three-variable linear case (Figs. 2–3) is a legitimate existence proof that interactive soft interventions plus MCMC can improve over a pure prior baseline.\n\nSoft spots, in proportion. The soft likelihood (indicator that a 1-β fraction of descendant responses lie inside an ℓ2 ball of radius σ) is under-identified; many wrong θ vectors can satisfy it for T=10, so posterior mass need not concentrate on the parameters that actually govern the counterfactual map. The linear-Gaussian experiments never isolate this gap because the same generative process supplies both queries and evaluation. Topology is assumed known, exogenous noise is fixed standard Gaussian, and all “humans” are simulated mixtures of θ_GT and adversarial θ*. Those are real limits, but the paper labels itself a proof-of-concept and does not claim otherwise. Free parameters (α, σ, β, T, ε, γ, θ bounds) are numerous yet standard for this literature.\n\nThis is for people already working on causal or preference-based recourse who want a concrete next experiment, not for someone looking for a ready-to-deploy method. Math and citations look solid; no circularity in the main metrics. I would send it to peer review: the gap is real, the formalization is useful, and the linear results plus honest failure modes give referees something concrete to push on (real users, richer SCMs, query selection). Worth engaging if you care about the subfield; not a must-read outside it.","headline":"Clean proof-of-concept that personalizes causal recourse via HITL Bayesian ANM estimation; linear sims work, soft likelihood and known-topology assumptions keep it provisional.","tokens_in":12333,"tokens_out":553,"would_cite":false,"duration_ms":4900,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Querying a user about interventions can recover a personal causal model that yields valid, lower-cost recourse.","keywords":["algorithmic recourse","causal recourse","human-in-the-loop","structural causal models","Bayesian inference","counterfactual explanations","personalized interventions"],"falsifier":"Run the same query-and-estimate pipeline on real users whose feature dependencies are known to be non-linear or whose noise is non-Gaussian; if the resulting actions fail to flip the decision when applied in the real world at rates near the linear synthetic case, the central claim does not hold outside the assumed model class.","tokens_in":12167,"feed_emoji":"🔄","tokens_out":754,"duration_ms":6795,"temperature":0.7,"pith_summary":"Standard algorithmic recourse either ignores how features cause one another or pretends every person shares the same known causal model. This paper argues that you can instead treat each user as a noisy oracle: ask them what would happen if you changed one of their features, then use Bayesian inference to recover a personal structural causal model. Recourse actions computed on that recovered model, when applied to the true (unobserved) process, stay nearly as valid and cheap as oracle recourse for linear synthetic users and beat a non-causal baseline. The same pipeline degrades under non-linear mechanisms and mismatched noise, so the authors present the method as a controlled proof of concept rather than a finished system.","feed_headline":"User queries recover personal causal models for better recourse","feed_subtitle":"Bayesian estimation from intervention feedback yields lower-cost, more valid actions than non-causal baselines","key_machinery":"Personalized Causal Recourse (Definition 1): MCMC estimation of structural coefficients in a linear ANM surrogate from intervention–response pairs, followed by gradient-based robust optimization of the cheapest intervention that flips the classifier under the estimated model.","core_discovery":"A human-in-the-loop Bayesian procedure that estimates a user-specific linear additive-noise model from noisy soft-intervention responses produces personalized causal recourse whose validity and cost, measured on the ground-truth SCM, closely track the oracle and improve over a non-causal prior baseline for linear synthetic data.","pith_inferences":["If the topological order itself must also be elicited, the query budget and MCMC design would need to expand substantially beyond the current fixed-order setting.","The same intervention-response protocol could be used to personalize cost models or preference structures rather than only structural equations.","Real-user studies would immediately reveal whether the stylized mixture-noise response model underestimates structured human biases."],"forward_implications":["Recourse systems need not assume a single shared causal model; individual intervention queries can substitute for that knowledge.","When the linear-ANM class matches the user, estimated recourse stays close to oracle validity and cost while beating non-causal baselines.","Model mismatch (mixture noise, non-linear mechanisms) measurably degrades both estimation accuracy and recourse validity.","Robustness constraints (epsilon-balls) reduce validity even under the true model, exposing a feasibility–robustness trade-off that must be managed."],"fun_headline_variants":["Human feedback builds personal SCMs for valid lower-cost recourse","Bayesian queries recover user causal models for personalized actions","Interactive soft interventions estimate SCM for better recourse","HITL Bayesian inference personalizes causal recourse over baselines","Noisy intervention responses yield user-specific causal recourse"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The user's true causal process can be well approximated by a linear additive-noise model whose graph order is already known and whose noise is independent standard Gaussian.","fun_headline_variants_meta":{"raw":{"variants":["Human feedback builds personal SCMs for valid lower-cost recourse","Bayesian queries recover user causal models for personalized actions","Interactive soft interventions estimate SCM for better recourse","HITL Bayesian inference personalizes causal recourse over baselines","Noisy intervention responses yield user-specific causal recourse"]},"model":"grok-4.5","effort":"low","cost_usd":0.002556,"raw_usage":{"total_tokens":952,"prompt_tokens":693,"num_sources_used":0,"completion_tokens":59,"cost_in_usd_ticks":25560000,"prompt_tokens_details":{"text_tokens":693,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":200,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":693,"tokens_out":59,"duration_ms":2415,"temperature":1.0,"reasoning_tokens":200,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T02:33:50.972240+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the same query-and-estimate pipeline on real users whose feature dependencies are known to be non-linear or whose noise is non-Gaussian; if the resulting actions fail to flip the decision when applied in the real world at rates near the linear synthetic case, the central claim does not hold outside the assumed model class.","supporting_citations":[],"review_version":1}