{"id":"c21a6ab7-9051-48a9-a7ce-5c4e0bddd582","arxiv_id":"2607.04133","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PCG-CD recovers directed edges and latent confounders by comparing LNML code-lengths of four pairwise models and thresholding a Δ-pseudo-collinearity gap.","lead":"The paper introduces PCG-CD, an MDL method that scores nonlinear causal models with Gaussian processes and flags latent confounders via a code-length gap called Δ-pseudo-collinearity. It matters because most practical causal graphs mix nonlinear mechanisms with hidden common causes, and few methods handle both without strong parametric assumptions.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Heuristic residual proxy for latent code-length is the load-bearing soft spot under nonlinear mechanisms","rationale":"The reader correctly isolates the residual-proxy approximation (§4.1–4.3) as the weakest assumption supporting the strongest claim. My stress-test confirms that this is load-bearing: without a reliable mapping from unobserved Z to the residual pairwise score, both the Δ-threshold rule and the subsequent greedy skeleton construction lose their justification under the nonlinear mechanisms the paper advertises. The linear-only theorem and the authors’ own “heuristic” language make the gap explicit; the proposed concrete test directly measures whether the proxy tracks the true latent score on the same nonlinear data used for evaluation. No stronger internal inconsistency appears, so the verdict remains CONDITIONAL rather than REJECT; the concern simply sharpens the conditions already noted by the reader (nonlinear theory, ε sensitivity, broader baselines).","tokens_in":13209,"tokens_out":673,"duration_ms":5947,"concrete_test":"Generate 50 independent 6-node ER graphs with one latent confounder under the paper’s cubic nonlinear SEM (§5.2), compute both the true residual-conditioned LNML gap (using the known Z) and the residual-proxy Δ of Eq. 10 for every confounded and non-confounded pair; report the AUC of proxy-Δ versus true-gap for distinguishing A↔B from directed edges. If AUC falls below ~0.75 (or the linear-case AUC of Table 1), the heuristic fails for the claimed nonlinear regime and the strongest claim weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that PCG-CD recovers directed edges and latent confounders rests on deciding among four pairwise models by LNML scores, with the confounder case identified by Δ-pseudo-collinearity (Eq. 10: |Δ| ≤ nε). Section 4.1 explicitly replaces the true latent score L(A_res | Z) + L(B_res | Z) by the residual pairwise proxy L(A_res ↔ B_res) = L(A_res | B_res) + L(B_res | A_res), calling the substitution “heuristic in nature” and noting that it is used only to make the objective computable, not as a principled MDL decision. Theorem 1 supplies an O(n C(|coefA-coefB|,σ Z,…)) bound solely for linear mechanisms; the nonlinear discussion is informal (“if approximately linear or range small”). If the residual proxy fails to track true confounding under the cubic (or other) nonlinear mechanisms used in the experiments, both the orientation rule and the bidirected decisions become unreliable, undermining the abstract claim. The synthetic multi-node results (Table 2) and Auto-MPG example therefore cannot be taken as confirmation that the same proxy works for realistic nonlinear latent confounding.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes an MDL framework for causal discovery under nonlinear mechanisms and latent confounders. Causal mechanisms are modeled by Gaussian processes; pairwise relationships among residualized variables are scored by an LNML-style code-length (Eqs. 5–8). Four candidate structures (A\to B, A←B, A↔B, independence) are compared, and latent confounding is declared when the directional gap satisfies Δ-pseudo-collinearity (Eq. 10: |Δ|≤nε). A greedy algorithm (PCG-CD) starts from a complete bidirected graph, orients or removes edges by these scores, then breaks cycles. Theorem 1 bounds Δ for linear latent mechanisms; experiments on synthetic pairs/graphs and Auto MPG report improved Z-hit/E-hit versus FCI variants.","tokens_in":13567,"tokens_out":1287,"duration_ms":10118,"significance":"If the residual proxy and thresholding rule are reliable, the work would be a useful addition to MDL-based causal discovery: it is mechanism-agnostic (GP), explicitly handles latent confounders, and reduces edge evaluation complexity relative to GES-style search. Strengths include an explicit LNML formulation under GPs, a named detection criterion (Δ-pseudo-collinearity) with a linear-case bound, a concrete O(|V|²) greedy procedure, and public code. The contribution is incremental rather than foundational, but the combination of nonlinear mechanisms and latent-confounder scoring under MDL is of genuine interest to the causal-discovery community.","major_comments":[{"comment":"§4.1 explicitly replaces the true latent score L(A_res|Z)+L(B_res|Z) by the residual pairwise proxy L(A_res↔B_res)=L(A_res|B_res)+L(B_res|A_res), calling the substitution “heuristic in nature” and stating it is used only to make the objective computable, not as a principled MDL decision. The abstract and §1.3 claim that PCG-CD “effectively detects latent confounders” under nonlinear mechanisms. Because the proxy is the sole operational definition of the confounder model, the central claim rests on an unvalidated approximation. A controlled check (e.g., synthetic nonlinear Z→A, Z→B with known ground-truth code-lengths, or comparison against a sampled latent model) is needed before the multi-node results can be read as confirmation.","section":null},{"comment":"Theorem 1 (§4.3) supplies an O(n C(|coefA−coefB|,σ_Z,…)) bound only for linear mechanisms. The nonlinear discussion is informal (“if approximately linear or range small”). Experiments in §5.2 use cubic polynomial mechanisms, yet no corresponding bound or diagnostic is given for when Δ remains O(n) under those mechanisms. Without this, the fixed threshold rule Δ≤nε (ε=5) lacks theoretical support precisely in the regime claimed by the abstract.","section":null},{"comment":"§5 sets ε=5 from “preliminary experiments on synthetic graphs with similar sparsity” and keeps it fixed for all sample sizes and both linear/nonlinear settings. Table 1 shows that AUC of Δ for distinguishing direct edges from confounding varies substantially with n and σ (e.g., linear n=1000, σ=0.2 yields AUC≈0.55). A sensitivity analysis or data-driven selection of ε is required; otherwise the reported Z-hit/E-hit in Table 2 may be tuned to the same synthetic family used to choose the threshold.","section":null},{"comment":"Table 2 evaluates only 6-node Erdős–Rényi graphs with a single latent confounder and 20 replicates. FCI is the sole baseline; stronger latent-aware nonlinear methods cited in §1.2 (e.g., NOCADILAC, CoCa) are not compared. Sim1/Sim2 favor sparse graphs, so FCI’s higher similarity scores are dismissed as “false negatives,” but no denser or multi-confounder regime is tested. The multi-node evidence is therefore too narrow to support the general claim of accurate recovery under nonlinear latent confounding.","section":null}],"minor_comments":[{"comment":"§3.2 and Eq. (8) present an upper-bound Bayesian/GP score as “the resulting computable LNML objective.” Clarify the precise relationship to the LNML definition (Eqs. 1–2) and whether the parametric-complexity term is fully accounted for.","section":null},{"comment":"Algorithm 1 Stage 2 modifies edges inside cycles by “minimum loss” but does not specify how loss is computed when converting → to ↔ versus deleting the edge; a short formal definition would help reproducibility.","section":null},{"comment":"Figure 2 annotates Δ values on PCG-CD edges; the caption and text should state the sample size used for the Auto MPG run so that the threshold nε can be verified.","section":null},{"comment":"Notation for residual components (A_res) appears in §4 without an explicit formula; a one-line definition would improve readability.","section":null},{"comment":"Related-work coverage of MDL latent methods (DNML, CDHC) is brief; a sentence contrasting the GP residual proxy with those approaches would situate the novelty more clearly.","section":null}],"recommendation":"major_revision","confidential_remarks":"The load-bearing soft spot identified by the stress-test (heuristic residual proxy under nonlinear mechanisms) is real and is acknowledged by the authors themselves in §4.1. I do not view it as an irrecoverable flaw, but it must be either theoretically tightened or empirically stress-tested before acceptance. The paper is otherwise clearly written and the code release is a plus. Fit for a solid ML/causal-discovery venue after revision; not yet ready for a top-tier journal without the additional validation."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The new piece here is a concrete greedy algorithm (PCG-CD) that scores GP mechanisms with LNML, decides among four pairwise relations via code-length, and flags latent confounding with a named Δ-pseudo-collinearity gap. That combination is not just a rehash of GLOBE, CoCa, or FCI; it is a practical deletion-based O(|V|²) search that starts from a complete bidirected graph and never re-adds edges.\n\nWhat works: the GP LNML score under sufficiency is standard and cleanly written. The residual reformulation makes the latent-aware objective computable without sampling Z. Theorem 1 gives a usable linear-case bound that motivates the nε threshold. Synthetic pairwise AUCs look strong, multi-node Z-hit/E-hit beat FCI on the authors’ metrics, and the Auto-MPG example is readable. Code is linked. Citations cover the right MDL and latent-FCM lines.\n\nSoft spots are real but not hidden. Section 4.1 itself calls the residual proxy L(A_res ↔ B_res) for the true latent score “heuristic.” Theory stops at linear mechanisms; the nonlinear discussion is informal. ε=5 is fixed from preliminary runs on similar graphs. Multi-node experiments stay at six nodes with one planted confounder and custom metrics that punish sparse FCI. No RCD/NOCADILAC/GES-style baselines, no uncertainty, no ε sensitivity. Those limits keep the abstract claim from being fully secured, but they do not make the method incoherent.\n\nThis is for people already working on MDL causal discovery or needing a nonlinear continuous method that can output bidirected edges. A serious referee should see it; the contribution is clear enough and the gaps are fixable. I would engage, cite the algorithm and the pseudo-collinearity idea if I am writing in this area, and bring it to reading group if we are covering latent-aware continuous methods.","headline":"Solid mid-subfield MDL method for nonlinear continuous data with latents; the residual proxy is openly heuristic and theory is linear-only, but the algorithm and results are still worth a look.","tokens_in":14149,"tokens_out":496,"would_cite":true,"duration_ms":4612,"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":"An MDL score that treats latent common causes as first-class options lets nonlinear causal graphs be recovered without assuming causal sufficiency.","keywords":["causal discovery","latent confounders","minimum description length","LNML","Gaussian processes","pseudo-collinearity","nonlinear mechanisms"],"falsifier":"Generate nonlinear data from a known latent confounder whose mapping to the two children is highly non-invertible or high-noise; if the residual Δ-gap still falls below the threshold while a true directed edge produces a large gap, the criterion fails.","tokens_in":14087,"feed_emoji":"↔️","tokens_out":608,"duration_ms":5489,"temperature":0.7,"pith_summary":"Causal discovery usually either assumes every relevant variable is observed or forces mechanisms to be linear. This paper drops both restrictions by scoring every pair of variables under four competing models—A causes B, B causes A, a shared latent cause, or independence—using the luckiness-normalized maximum-likelihood code length of a Gaussian-process regression. When the two directed scores are almost equal, the gap is called Δ-pseudo-collinearity and is taken as evidence of a latent confounder rather than a direct edge. A greedy algorithm, PCG-CD, starts from a complete bidirected graph, orients or deletes edges according to these scores, and finally breaks cycles by the cheapest local change. On synthetic nonlinear data and a real automobile data set the method recovers directed edges and flags confounded pairs more reliably than the standard partial-ancestral-graph baseline. If the approach holds up, practitioners can search for causal structure in continuous, nonlinear systems even when unobserved common causes are known to exist.","feed_headline":"MDL scores flag latent common causes in nonlinear data","feed_subtitle":"A code-length gap called Δ-pseudo-collinearity orients edges or marks confounders without linear assumptions","key_machinery":"Δ-pseudo-collinearity: the absolute difference between the two directed residual code lengths; when it stays below a linear threshold the pair is declared bidirected (latent-confounded).","core_discovery":"Minimizing the LNML code length of Gaussian-process structural equations, together with a threshold on the pairwise directional gap Δ ≤ nε (Δ-pseudo-collinearity), identifies both directed causal edges and latent confounders without requiring linear mechanisms or causal sufficiency.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["LNML code lengths reveal directed edges and latent confounders","MDL via LNML orients pairs and flags Δ-pseudo-collinear confounders","PCG-CD greedily recovers nonlinear causes without sufficiency","Shortest LNML models detect latent common causes in nonlinear data","Δ gap on GP LNML scores marks confounders or causal directions"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The unobserved common cause can be replaced, for scoring purposes, by the residuals of the two observed variables themselves; the paper calls this substitution a heuristic.","fun_headline_variants_meta":{"raw":{"variants":["LNML code lengths reveal directed edges and latent confounders","MDL via LNML orients pairs and flags Δ-pseudo-collinear confounders","PCG-CD greedily recovers nonlinear causes without sufficiency","Shortest LNML models detect latent common causes in nonlinear data","Δ gap on GP LNML scores marks confounders or causal directions"]},"model":"grok-4.5","effort":"low","cost_usd":0.005634,"raw_usage":{"total_tokens":1443,"prompt_tokens":660,"num_sources_used":0,"completion_tokens":92,"cost_in_usd_ticks":56340000,"prompt_tokens_details":{"text_tokens":660,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":691,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":660,"tokens_out":92,"duration_ms":6812,"temperature":1.0,"reasoning_tokens":691,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T21:26:24.072779+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Generate nonlinear data from a known latent confounder whose mapping to the two children is highly non-invertible or high-noise; if the residual Δ-gap still falls below the threshold while a true directed edge produces a large gap, the criterion fails.","supporting_citations":[],"review_version":1}