{"id":"c3db9a0f-6d87-4087-8f25-06d7dd98438b","arxiv_id":"2607.09461","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"CEFOL solves recursive-utility dynamic programs by learning a state-control certainty equivalent and minimizing Bellman plus first-order/KKT residuals with neural networks.","lead":"The paper introduces CEFOL, a deep-learning solver for discrete-time dynamic programs with recursive utility that learns the nonlinear certainty equivalent with its own network and trains policies via first-order and KKT residuals. It matters because recursive preferences are standard in asset pricing and macro, yet hard to solve when constraints bind and states are multi-dimensional.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Residual diagnostics partly re-use training objectives; VFI agreement is limited to low-dimensional cases, so the accuracy claim rests on a soft validation standard.","rationale":"The reader correctly identifies residual-based evaluation under nonconvex training, without a convergence theorem and with only partial VFI coverage, as the weakest assumption behind the accuracy claim. That is the right load-bearing concern: the method is carefully derived and the numerical suite is broader than a toy example, but the evidence that small residuals imply an accurate solution to the true recursive program is soft outside the VFI-checked small-noise case. I do not see a stronger internal inconsistency or a derivation error that would overturn the methods contribution. The concern therefore supports keeping CONDITIONAL rather than moving to ACCEPT or REJECT. A multi-seed / multi-architecture residual-and-policy stability check on the non-VFI models would settle whether residual smallness is robust enough to underwrite the headline accuracy language.","tokens_in":40784,"tokens_out":628,"duration_ms":6195,"concrete_test":"For the Epstein–Zin consumption-saving model with (β,γ,ρ)=(0.95,20,0.5) and the DSGE calibration (§4.2, §4.4), recompute out-of-sample Bellman and FOC residuals and policy functions under at least three independent random seeds and two materially different network widths/depths (or learning-rate schedules), holding the evaluation grid fixed. If median residuals or policy values shift by more than about one order of magnitude (or >5–10% relative policy difference) across runs, residual smallness is not a stable accuracy certificate and the strongest claim should be narrowed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that CEFOL solves recursive-utility DPs with residuals of order 10^{-4}–10^{-3} and close VFI match when available. That claim is load-bearing on residual smallness plus optional VFI agreement as sufficient evidence of an accurate solution to the true program. The paper itself flags nonconvex training with no global optimality guarantee (§1). The reported out-of-sample Bellman and FOC/Euler residuals (§4) are the same residual classes minimized in training (§3.2–3.3: certainty-equivalent, Bellman, stationarity, Fischer–Burmeister). Nested simulation for diagnostics is more accurate than the training estimator, but still measures consistency of the learned fixed point rather than distance to the true value/policy. VFI agreement is available only for the small-noise robust-control model (§4.3); risk-sensitive, Epstein–Zin consumption-saving, and DSGE cases rely on internal consistency (CEFOL vs CEFOL-td vs CEFOL-vc) and residual size. Larger residuals near binding constraints are expected for interior Euler diagnostics, but without independent high-accuracy benchmarks or multi-seed/hyperparameter sensitivity, residual smallness alone does not tightly pin down solution accuracy for the harder models.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes CEFOL, a deep learning algorithm for discrete-time infinite-horizon dynamic programs with recursive utility. The central device is a dedicated neural network for the state-control certainty-equivalent value C(s,c), which is then inserted into Bellman targets and into model-specific first-order/KKT residuals used to train policy and multiplier networks. Value, certainty-equivalent, and (when used) multiplier networks are trained on separate residual losses, with stabilizations including target value networks, delayed FOC updates, exploratory control perturbation, and an independent sample-mean product estimator for squared conditional stationarity residuals. The framework is written for general equality/inequality constraints and occasionally binding constraints via Fischer–Burmeister KKT terms. Numerical applications cover risk-sensitive and Epstein–Zin consumption-saving models, a small-noise robust-control problem (with VFI benchmarks), and an Epstein–Zin DSGE model with stochastic volatility. Reported out-of-sample Bellman and FOC/Euler residuals are typically of order 10^{-4}–10^{-3} over relevant state regions, with larger residuals mainly near binding constraints, and close agreement with VFI when available.","tokens_in":41216,"tokens_out":1699,"duration_ms":33882,"significance":"If the method works as claimed, it is a useful computational contribution for recursive-utility dynamic programming, where nonlinear certainty equivalents enter both the Bellman equation and optimality conditions and are hard to evaluate inside policy updates. The careful FOC/KKT derivation under recursive utility (§2.2), including the structural condition that can eliminate future value gradients, and the modular treatment of vector controls and occasionally binding constraints without penalty reformulations, are genuine strengths relative to expected-utility residual methods and to model-specific deep equilibrium systems. The paper also situates CEFOL relative to Maliar et al. (2021), Friedl et al. (2023), and the companion CEL algorithm, and shows that expected utility is nested. For a computational journal, residual diagnostics plus VFI agreement on a low-dimensional nonlinear case are a recognizable evaluation package; the main open issue is how tightly residual smallness pins down solution accuracy for the harder recursive models.","major_comments":[{"comment":"§3.2–3.3 and §4: The headline accuracy claim (Abstract; §4) rests on out-of-sample Bellman and FOC/Euler residuals of order 10^{-4}–10^{-3}, plus VFI agreement only in the small-noise robust-control model (§4.3). Those residual classes are essentially the same objectives minimized in training (certainty-equivalent, Bellman, stationarity, Fischer–Burmeister). Nested simulation improves the diagnostic estimator relative to the training product estimator, but still measures consistency of a learned fixed point rather than distance to the true value/policy. For the risk-sensitive, Epstein–Zin consumption-saving, and DSGE applications, the paper relies on internal CEFOL/CEFOL-td/CEFOL-vc alignment and residual size. Given the paper’s own acknowledgment of nonconvex training without global optimality guarantees (§1), the accuracy claim needs stronger independent grounding: multi-seed/hyperpara","section":"§4 Numerical Results; §3.2–3.3"},{"comment":"§2.2, Eqs. (18)–(27): The FOC simplification that eliminates ∂V(s_{t+1})/∂s_{t+1} depends on a structural condition that the transition factors through a scalar y_{t+1}. This is load-bearing for the practical residual forms used later. The applications in §4.1–4.4 implement specialized Euler/KKT residuals, but the manuscript does not systematically verify that (18) holds in each environment or state what is done when it fails (direct automatic differentiation of C and V, alternative envelope forms, etc.). Please state explicitly, for each numerical model, whether the structural condition is used, how F_{t+1,k} is obtained, and whether any residual involves future value gradients.","section":"§2.2; applications in §4.1–4.4"},{"comment":"§3.1 and §4: Three architectures are developed (four-, five-, and three-network), yet the numerical section appears to use only the four-network baseline. Without ablations, the five-network decomposition (Ve, D) and the compact three-network variant remain untested design options rather than demonstrated tools. Either report comparative residuals/stability/cost for the alternative architectures on at least one application, or narrow the main text to the architecture actually used and move the others to an appendix as optional variants. Relatedly, CEFOL is positioned as complementary to CEL (Peng and Guo, 2026) and distinct from Friedl et al. (2023), but §4 contains no head-to-head accuracy or cost comparison on a shared recursive-utility problem; such a comparison would substantially strengthen the methods contribution.","section":"§3.1; §4; literature discussion of CEL/Friedl"}],"minor_comments":[{"comment":"Figures in §4 often evaluate along a one-dimensional slice (cash-on-hand or capital) with other states fixed at steady state. Please state this limitation more prominently when interpreting residual magnitudes, and consider reporting residual quantiles on simulated ergodic samples as a complement.","section":"§4 figures and captions"},{"comment":"Hyperparameters (network widths/depths, learning rates, delay d, τ, residual weights λ_S/λ_FB, exploration ζ_k, N_z, etc.) are numerous but only partially documented in the main text. A compact table of training settings per application would aid reproducibility.","section":"§3.4–3.5; §4"},{"comment":"In the risk-sensitive examples, absolute Bellman errors are used because V crosses zero; in other examples relative Bellman errors are used. A short note unifying the diagnostic definitions and normalizations (including the Euler residual (106)–(108)) would help cross-model comparison.","section":"§4 opening; §4.1.3"},{"comment":"Notation occasionally shifts between C(s,c), C(s,c;θ_C), and script C; keep a single convention for the true certainty equivalent versus the network approximator.","section":"§2–§3"},{"comment":"The manuscript date line reads “July 13, 2026”; confirm versioning/arXiv metadata consistency before journal submission.","section":"title page"},{"comment":"Some figure panels mark horizontal thresholds (e.g., 10^{-4}, 10^{-3}) without stating whether these are formal tolerances or visual guides; clarify in captions.","section":"§4 figures"}],"recommendation":"major_revision","confidential_remarks":"Fit is appropriate for a computational economics/finance methods venue. The contribution is incremental relative to residual deep learning for expected-utility models and to Friedl et al./CEL, but the recursive-utility FOC packaging and constraint handling are real. I would not reject on novelty grounds. The main editorial risk is overstated accuracy language in the Abstract if the authors do not strengthen independent validation; with clearer caveats and the requested checks, the paper can be a solid methods piece. Note the companion CEL paper (arXiv:2607.04278) is closely related; ensure the journal is comfortable with the split and that CEFOL’s distinct FOC-based policy update is demonstrated, not only described."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful move here is treating the state-control certainty equivalent as its own network and feeding it into model-specific FOC/KKT residuals for policy and multipliers, rather than burying the nonlinear CE inside a joint residual soup or only using Bellman improvement. That is the real difference from Maliar-style residual methods and from Friedl et al., and it is spelled out cleanly.\n\nWhat they do well: the §2.2 derivation is careful, including the structural condition that kills future value gradients when the transition factors through a scalar y. The modular four/five/three-network layouts, delayed FOC updates, target value net, and Fischer–Burmeister KKT handling are practical and general enough for occasionally binding constraints without penalty hacks. The numerical suite is broader than a toy: risk-sensitive and EZ consumption-saving, small-noise robust control with VFI, and a recursive DSGE with stochastic volatility. Residuals of order 10^{-4}–10^{-3} over interior regions, internal CEFOL/CEFOL-td/CEFOL-vc agreement, and close VFI match where available are honest computational evidence for this class of method.\n\nSoft spots, in proportion: no convergence theorem and nonconvex training are flagged by the authors themselves; residual diagnostics partly re-evaluate the same classes minimized in training, so they measure fixed-point consistency more than distance to the true solution. Nested simulation helps, but VFI is only for the small-noise model. No code, incomplete hyperparameter reporting, free parameters everywhere. That is standard for this literature, not a hidden collapse of the claim.\n\nThis is for people who actually solve recursive-utility DPs with constraints and want a mesh-free FOC-based recipe. It does not reorganize theory. I would send it to peer review; the construction is clear enough and the experiments serious enough to deserve referee time. Worth engaging if you work on computational macro/finance solvers.","headline":"Solid methods paper: separate CE network into FOC/KKT residual learning for constrained recursive-utility DP; residual evidence is real but soft, VFI only on one low-dim case.","tokens_in":41843,"tokens_out":486,"would_cite":true,"duration_ms":7679,"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":"A dedicated neural network for the certainty equivalent lets deep learning solve recursive-utility dynamic programs, including occasionally binding constraints, via first-order and KKT residuals.","keywords":["deep learning","dynamic programming","recursive utility","certainty equivalent","first-order conditions","KKT residuals","Epstein–Zin","occasionally binding constraints"],"falsifier":"On any of the reported models, recompute nested Monte Carlo Bellman and Euler/KKT residuals (or a high-resolution VFI benchmark) on an independent dense test set; if residuals remain large or systematically biased away from the training residual classes, or if policies diverge from VFI where VFI is reliable, the claim fails.","tokens_in":41642,"feed_emoji":"📈","tokens_out":700,"duration_ms":7495,"temperature":0.7,"pith_summary":"Dynamic programs with recursive preferences (risk-sensitive, Epstein–Zin, robust control) are hard because a nonlinear certainty-equivalent of next-period value sits inside both the Bellman equation and the optimality conditions, and sample averages do not recover it. This paper claims that learning that state-control certainty equivalent with its own neural network, then training policy and multiplier networks by minimizing model-specific first-order and KKT residuals (while separate networks keep Bellman and certainty-equivalent consistency), yields accurate policies and values without grids, penalty methods, or problem-specific rewrites. On consumption-saving, small-noise robust-control, and DSGE examples, out-of-sample Bellman and optimality residuals are typically 10⁻⁴–10⁻³ over the relevant state region and match value-function-iteration benchmarks where those exist. Expected utility is recovered as the special case of a linear certainty equivalent, so the same architecture covers both settings. A sympathetic reader cares because many modern asset-pricing and macro models use recursive preferences and constraints that classical grid methods cannot scale to.","feed_headline":"Neural certainty equivalent unlocks recursive-utility DP","feed_subtitle":"First-order and KKT residuals solve constrained models with 10⁻⁴–10⁻³ errors","key_machinery":"The certainty-equivalent first-order learning (CEFOL) scheme: a certainty-equivalent network C(s,c) that amortizes the nonlinear transformation f⁻¹(E[f(V(s′))]), combined with residual losses built from stationarity (including Euler), Fischer–Burmeister complementarity, and equality constraints, trained with target networks and delayed policy updates.","core_discovery":"Representing the state-control certainty equivalent by a dedicated neural network, and feeding that approximation into model-specific first-order and KKT residual losses for the policy and multipliers, produces a practical deep-learning solver for discrete-time recursive-utility dynamic programming that handles general equality and inequality constraints, including occasionally binding ones, with small residual diagnostics and close agreement with VFI when available.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["CE net enables FOC-trained recursive-utility DP solvers","Dedicated certainty-equivalent net handles constrained recursive DP","CEFOL trains policies via FOC and KKT residuals for recursive utility","Neural CE approximation solves recursive DP with occasionally binding constraints","First-order residual nets crack recursive-utility dynamic programs"],"cache_read_input_tokens":32896,"weakest_assumption_plain":"That small out-of-sample residual diagnostics under nonconvex neural training, plus optional agreement with grid VFI on low-dimensional cases, are enough to treat the learned networks as accurate solutions of the true recursive program.","fun_headline_variants_meta":{"raw":{"variants":["CE net enables FOC-trained recursive-utility DP solvers","Dedicated certainty-equivalent net handles constrained recursive DP","CEFOL trains policies via FOC and KKT residuals for recursive utility","Neural CE approximation solves recursive DP with occasionally binding constraints","First-order residual nets crack recursive-utility dynamic programs"]},"model":"grok-4.5","effort":"low","cost_usd":0.005642,"raw_usage":{"total_tokens":1565,"prompt_tokens":841,"num_sources_used":0,"completion_tokens":65,"cost_in_usd_ticks":56420000,"prompt_tokens_details":{"text_tokens":841,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":659,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":841,"tokens_out":65,"duration_ms":6942,"temperature":1.0,"reasoning_tokens":659,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T02:51:27.185706+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On any of the reported models, recompute nested Monte Carlo Bellman and Euler/KKT residuals (or a high-resolution VFI benchmark) on an independent dense test set; if residuals remain large or systematically biased away from the training residual classes, or if policies diverge from VFI where VFI is reliable, the claim fails.","supporting_citations":[],"review_version":1}