{"id":"cb1e6caa-80cb-401b-9617-b5f5affa76ba","arxiv_id":"2606.17513","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"REEF-GP fits a Gaussian process to residuals of a frozen neural operator using its internal embeddings to deliver geometry-aware post-hoc uncertainty quantification for PDEs on unstructured domains.","lead":"The paper introduces REEF-GP, a post-hoc method to add uncertainty estimates to neural operators for PDEs by fitting a Gaussian process to the model's residuals using its learned internal embeddings as the feature space for the kernel. This could allow fast PDE surrogates to provide calibrated, geometry-aware uncertainties at lower cost than ensembles, supporting more reliable use in applications with varying shapes.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Suitability of frozen operator embeddings as stable GP kernel features under geometric shift remains an untested assumption","rationale":"The reader's weakest_assumption directly identifies the same load-bearing point. Because the full manuscript was inaccessible to the reader and the abstract supplies only high-level claims without the promised ablations or stability analysis, the concern stands as the primary unverified link; no stronger internal inconsistency is visible from the given material.","tokens_in":1724,"tokens_out":322,"duration_ms":17570,"concrete_test":"On one benchmark exhibiting geometric shift (e.g., the shock-front case), replace the operator embeddings with either (a) random Fourier features of the same dimension or (b) a separately trained feature map; recompute calibration metrics (e.g., ECE or negative log-likelihood) and uncertainty concentration on physically meaningful regions. If the embedding-based version loses its reported advantage, the assumption does not hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the neural operator's internal coordinate-feature embeddings form a suitable, stable feature space for the residual GP kernel without extra feature learning. The abstract states that REEF-GP 'adapts the operator's intrinsic coordinate-feature representations' and adds spectral normalization plus heteroscedastic noise for stability, yet provides no derivation or ablation showing that these embeddings remain geometry-aware and non-degenerate when the input geometry distribution shifts. If the embeddings collapse or become misaligned with the residual structure on unseen domains, the post-hoc GP calibration guarantee fails even if predictive accuracy is preserved.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes REEF-GP, a post-hoc UQ framework for neural operators that fits a GP to the residuals of a frozen operator, using its internal coordinate-feature embeddings to define the kernel (with spectral normalization, heteroscedastic noise, and subset training for stability). It claims that this preserves predictive accuracy while delivering calibrated uncertainties competitive with deep ensembles on five PDE benchmarks with varying geometries, at lower cost, and remains robust under geometric distribution shift with uncertainty concentrating in physically meaningful regions.","tokens_in":1867,"tokens_out":468,"duration_ms":24935,"significance":"If the central claims hold, the work would provide an efficient, geometry-aware alternative to parameter-centric UQ methods for operator learning, leveraging existing embeddings rather than ensembles or sampling; this could be practically useful for scalable uncertainty in PDE surrogates on unstructured domains.","major_comments":[{"comment":"Abstract: the robustness claim under geometric distribution shift ('Our approach remains robust under geometric distribution shift') rests on the untested assumption that the frozen operator's internal embeddings remain suitable, stable, and non-degenerate features for the GP kernel; no ablation, analysis, or derivation is referenced showing that embeddings do not collapse or misalign on shifted domains, which is load-bearing for the post-hoc calibration guarantee.","section":"Abstract"},{"comment":"Abstract (methods description): the claim that spectral-normalized projections plus heteroscedastic geometry-aware noise ensure stability and scalability on unstructured domains is stated without any equation, bound, or empirical verification that these modifications preserve geometry-awareness of the embedding kernel when input geometries shift; this directly affects whether the GP fit yields reliable uncertainties.","section":"Abstract"}],"minor_comments":[{"comment":"Abstract: the statement that results are 'competitive with deep ensembles but at a fraction of their cost' would benefit from explicit quantification of the cost ratio and error-bar details on the five benchmarks to allow direct comparison.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about embedding stability under shift is directly reflected in the manuscript's lack of targeted validation; this is a correctness-risk issue rather than a consensus disagreement, and targeted experiments (e.g., embedding similarity metrics or ablation on shifted geometries) would be needed to resolve it."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the careful reading and constructive feedback on the abstract claims. We address each major comment below and will revise the manuscript accordingly to strengthen the presentation of the robustness and stability arguments.","responses":[{"response":"We agree that the robustness claim would be strengthened by explicit analysis of embedding stability. In the revised manuscript we will add a dedicated subsection (with ablations and quantitative metrics) demonstrating that the frozen operator's internal embeddings remain non-degenerate and aligned on the geometrically shifted domains appearing in our benchmarks. This analysis will directly support the post-hoc calibration guarantee.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the robustness claim under geometric distribution shift ('Our approach remains robust under geometric distribution shift') rests on the untested assumption that the frozen operator's internal embeddings remain suitable, stable, and non-degenerate features for the GP kernel; no ablation, analysis, or derivation is referenced showing that embeddings do not collapse or misalign on shifted domains, which is load-bearing for the post-hoc calibration guarantee."},{"response":"We will expand the methods section to include the explicit equations governing the spectral-normalized projections and heteroscedastic noise model. We will also add empirical verification (including controlled comparisons with and without these components) showing that geometry-awareness of the embedding kernel is preserved under the geometric shifts present in the benchmarks.","revision_made":"yes","referee_comment":"[Abstract] Abstract (methods description): the claim that spectral-normalized projections plus heteroscedastic geometry-aware noise ensure stability and scalability on unstructured domains is stated without any equation, bound, or empirical verification that these modifications preserve geometry-awareness of the embedding kernel when input geometries shift; this directly affects whether the GP fit yields reliable uncertainties."}],"tokens_in":1356,"tokens_out":385,"duration_ms":19685,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper's main move is to freeze a trained neural operator and fit a Gaussian process directly to its residuals, using the operator's internal coordinate-feature embeddings to define the kernel. They layer on spectral normalization, heteroscedastic geometry-aware noise, and subset training to make it scale on unstructured meshes. This is presented as a distinct post-hoc route that avoids parameter sampling or separate feature learning.\n\nWhat stands out is the direct reuse of the learned embeddings for geometry-aware uncertainty without extra machinery. The five PDE benchmarks with varying geometries are a reasonable test bed, and the claim that predictive accuracy is preserved while uncertainty calibration matches deep ensembles at lower cost is the practical hook. Uncertainty concentrating near shocks is the kind of qualitative check that matters in these problems.\n\nThe soft spot is the central assumption that those embeddings stay stable and aligned with residuals when the input geometry distribution shifts. The abstract mentions the added stabilizations but does not show ablations or diagnostics that would confirm the embeddings do not collapse or misalign on unseen domains. If that happens, the GP calibration guarantee weakens even if mean predictions hold. Without the full methods section, data splits, and error-bar details it is also hard to judge how much the reported competitiveness depends on hyperparameter choices or benchmark specifics.\n\nThis is for people already using neural operators on PDEs who want a cheap UQ add-on rather than retraining ensembles. A reader focused on practical calibration under domain shift would find the construction worth examining.\n\nIt deserves peer review. The idea is clear enough and the empirical framing addresses a real need; referees can check whether the embedding-stability claim holds with the actual experiments and code.","headline":"REEF-GP repurposes a neural operator's own embeddings for a post-hoc residual GP with some stability tweaks, but the geometric-shift robustness claim rests on an assumption that still needs direct checks.","tokens_in":2389,"tokens_out":418,"would_cite":false,"duration_ms":21839,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Fitting a Gaussian process to the residuals of a frozen neural operator using its own internal embeddings produces calibrated, geometry-aware uncertainty estimates at low cost.","keywords":["neural operators","uncertainty quantification","Gaussian processes","post-hoc methods","geometry-aware modeling","PDE surrogates","residual learning","embedding features"],"falsifier":"On a new collection of PDE problems that include previously unseen geometric variations, if the uncertainty intervals produced by REEF-GP fail to cover the true errors at the nominal rate or if the uncertainty maps do not concentrate near known physical features, the central claim would be falsified.","tokens_in":2634,"feed_emoji":"","tokens_out":734,"duration_ms":34577,"temperature":0.7,"pith_summary":"Neural operators deliver fast approximations to PDE solutions across varying domain shapes, yet their point predictions leave users without reliable measures of error, especially when geometry changes. The paper shows that one can keep a trained operator fixed and build a Gaussian process directly on its prediction errors by treating the operator's learned internal representations as the kernel's feature space. This avoids learning any new feature map and incorporates spectral normalization plus subset training to handle unstructured domains. On five benchmarks the resulting uncertainties match the calibration quality of deep ensembles while preserving accuracy and using far less computation. The estimates also stay reliable when the domain geometry shifts, with higher uncertainty appearing near physically relevant features such as shock fronts.","feed_headline":"Residual GP on operator embeddings yields calibrated PDE uncertainties","feed_subtitle":"Matches ensemble accuracy at lower cost and stays reliable when domain geometry changes.","key_machinery":"REEF-GP (Residual on Embedded Features Gaussian Process), which constructs the GP kernel from the neural operator's intrinsic embeddings to produce geometry-aware uncertainties without retraining the operator.","core_discovery":"REEF-GP fits a Gaussian process to the residuals of a frozen neural operator, defining the kernel via the operator's internal coordinate-feature embeddings rather than a separately learned map. Spectral-normalized projections, heteroscedastic geometry-aware noise, and efficient subset-based training ensure stability and scalability. Across five PDE benchmarks with varying geometries, the method preserves predictive accuracy, yields uncertainty estimates competitive with deep ensembles at lower cost, and remains robust under geometric distribution shift, with uncertainty concentrating in physically meaningful regions.","pith_inferences":["The same embedding-based residual GP could be tested on operator architectures not included in the five benchmarks to check broader applicability.","If the method's calibration holds under geometric shift, it may reduce the need for explicit geometry-augmented training data in surrogate modeling pipelines.","Uncertainty concentration near shocks could be used to drive adaptive sampling or mesh refinement in downstream simulation loops.","Pre-computed embeddings from one operator might serve as a shared feature basis for multiple related PDE tasks."],"forward_implications":["Existing trained neural operators can receive calibrated uncertainty estimates without any parameter updates.","Uncertainty quantification automatically adapts to changes in domain geometry through the operator's learned representations.","Computational overhead remains close to a single model evaluation plus a lightweight GP fit rather than ensemble training.","Uncertainties naturally highlight regions of physical interest such as discontinuities without explicit supervision.","The approach applies directly to unstructured meshes common in real engineering geometries."],"fun_headline_variants":["REEF-GP uses operator embeddings for GP-based PDE uncertainty","Residual GP on frozen operator embeddings calibrates geometry-aware uncertainties","GP residuals from neural operator features quantify PDE uncertainties post-hoc","Scalable post-hoc UQ for neural operators via embedding-defined GPs","Geometry shift robust UQ from residuals in operator learned features"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The embeddings already learned by the neural operator form a stable and suitable feature space for a Gaussian process to model the residuals accurately.","fun_headline_variants_meta":{"raw":{"variants":["REEF-GP uses operator embeddings for GP-based PDE uncertainty","Residual GP on frozen operator embeddings calibrates geometry-aware uncertainties","GP residuals from neural operator features quantify PDE uncertainties post-hoc","Scalable post-hoc UQ for neural operators via embedding-defined GPs","Geometry shift robust UQ from residuals in operator learned features"]},"model":"grok-4.3","cost_usd":0.006262,"raw_usage":{"total_tokens":2951,"prompt_tokens":677,"num_sources_used":0,"completion_tokens":82,"cost_in_usd_ticks":62624500,"prompt_tokens_details":{"text_tokens":677,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2192,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":677,"tokens_out":82,"duration_ms":20659,"temperature":1.0,"reasoning_tokens":2192,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T01:39:48.825875+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"On a new collection of PDE problems that include previously unseen geometric variations, if the uncertainty intervals produced by REEF-GP fail to cover the true errors at the nominal rate or if the uncertainty maps do not concentrate near known physical features, the central claim would be falsified.","supporting_citations":[],"review_version":1}