{"id":"64eb2d47-96ab-42c3-8df1-a4637c8d7e44","arxiv_id":"2605.06355","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Introduces MO-ARM framework for training order-agnostic autoregressive models directly on incomplete data, showing implicit MCAR imputation in standard training and outperforming baselines on benchmarks.","lead":"This paper reinterprets order-agnostic autoregressive models as performing implicit imputation on complete data and introduces MO-ARM, a framework to train them directly on incomplete data under general missingness while also enabling active variable selection. A smart generalist might read it to see practical improvements for handling missing values in real-world prediction and generative modeling tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"Reader correctly flagged that full text was unavailable, preventing technical scrutiny of the weakest assumption. No additional concern can be raised without the derivations or tables.","tokens_in":1616,"tokens_out":212,"duration_ms":13759,"concrete_test":"Obtain the full manuscript and inspect §3 (training procedure) and §4 (experiments) to confirm whether the MCAR-imputation equivalence holds exactly and whether the general-missingness objective reduces to a standard likelihood under the stated assumptions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract outlines a coherent reinterpretation of order-agnostic autoregressive models for missing data, including an implicit MCAR imputation effect from standard training and a new framework for general missingness. No internal inconsistency, unsupported derivation, or experimental design flaw is detectable from the provided summary. The central outperformance claim rests on benchmarks whose details are not visible, but the high-level logic does not contain an obvious load-bearing gap.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript reinterprets order-agnostic autoregressive models for missing-data settings. It argues that standard training on fully observed data implicitly performs imputation under an MCAR mechanism, yielding robust out-of-sample imputation. It then introduces a principled training framework for incomplete data under general missingness mechanisms and applies the resulting amortized conditional densities to active information acquisition (sequentially choosing the most informative missing variables). The proposed MO-ARM model is reported to outperform established imputation baselines across real-world benchmarks.","tokens_in":1683,"tokens_out":332,"duration_ms":23813,"significance":"If the empirical claims are substantiated, the work supplies a coherent extension of order-agnostic autoregressive models to missing data, together with an implicit-imputation observation and an application to active acquisition. These elements could be useful for generative modeling pipelines that routinely encounter incomplete observations. The manuscript does not report machine-checked proofs or fully reproducible code artifacts, but the high-level logic appears internally consistent.","major_comments":[],"minor_comments":[{"comment":"Abstract: the specific real-world benchmarks, missingness rates, and quantitative margins of improvement are not stated; these details belong in the experimental section (presumably §4 or §5) to permit assessment of the outperformance claim.","section":"Abstract"},{"comment":"The notation used for the missingness mechanisms and the precise definition of the new training objective should be introduced with explicit equations early in the methods section to avoid ambiguity with standard missing-data terminology.","section":null}],"recommendation":"uncertain","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their review of our manuscript. We appreciate the recognition that the high-level logic appears internally consistent and that the elements could be useful for generative modeling pipelines encountering incomplete observations. The report does not enumerate specific major comments, so we have no individual points to rebut at this stage. We remain available to provide further clarification, additional experiments, or revisions should the referee or editor request them.","responses":[],"tokens_in":1152,"tokens_out":99,"duration_ms":12132,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main thing to know is that this work reinterprets order-agnostic autoregressive models so they can be trained straight on incomplete observations instead of requiring a separate imputation step first. They show that the usual complete-data training already behaves like imputation under missing completely at random, which explains some robustness at test time, and then extend it to a framework that handles arbitrary missingness patterns. They also use the conditional densities for active selection of which missing variables to observe next.\n\nWhat is actually new is the explicit training procedure on incomplete data and the active acquisition use case. The logic follows from the autoregressive factorization, so the extension feels natural rather than forced. The benchmarks on real-world data are a plus, and the consistent outperformance over standard imputation baselines is the kind of result that matters for practitioners.\n\nThe soft spots are mostly around verification. The abstract claims this is the first principled approach, so the paper needs to show clearly how it differs from other recent generative models that handle missing data. The experimental details matter here: whether the missingness mechanisms were simulated correctly, whether baselines were given the same information, and whether gains hold across different missing rates and patterns. If those controls are solid, the central claim holds up; if not, the advantage could shrink.\n\nThis paper is for people working on deep generative models or imputation in settings where data is rarely complete. A reader who already uses autoregressive density estimators would see immediate value in the missing-data extension. It has enough new technical content and empirical grounding to deserve a serious referee rather than a desk reject.","headline":"The paper gives a direct training method for order-agnostic autoregressive models on incomplete data under general missingness, with the claim that standard training already handles MCAR implicitly.","tokens_in":2138,"tokens_out":398,"would_cite":false,"duration_ms":18186,"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":"Order-agnostic autoregressive models can be trained directly on incomplete data under general missingness and outperform standard imputation.","keywords":["order-agnostic autoregressive models","missing data","imputation","missingness mechanisms","active information acquisition","deep generative models","incomplete datasets"],"falsifier":"On a dataset generated with a known non-random missingness pattern, compare MO-ARM imputation error against a baseline trained only on complete data; if the errors are statistically indistinguishable, the advantage of the new framework is falsified.","tokens_in":2527,"feed_emoji":"","tokens_out":646,"duration_ms":24875,"temperature":0.7,"pith_summary":"The paper reinterprets order-agnostic autoregressive models as already performing imputation when trained on complete data, because their variable ordering acts like random masking under a missing completely at random process. It then supplies the first training procedure that works on incomplete observations without assuming the data are fully observed. The same models support active selection of which missing entry to observe next by using their conditional densities to rank information gain. Across real benchmarks the resulting MO-ARM version beats established imputation methods. Readers would care because many practical datasets arrive with missing entries and embedding the handling inside the generative model removes an extra preprocessing stage.","feed_headline":"Autoregressive models train directly on incomplete data","feed_subtitle":"MO-ARM adjusts order-agnostic training for any missingness pattern and beats standard imputation on real benchmarks.","key_machinery":"The Missingness-Aware Order-Agnostic Autoregressive Model (MO-ARM), which folds the observed missing pattern into the autoregressive factorization so that each conditional is trained only on the actually observed preceding variables.","core_discovery":"Order-agnostic autoregressive models trained on fully observed data implicitly perform imputation under a missing completely at random mechanism and retain robust performance at high missing rates. The authors introduce a missingness-aware training framework that extends the same models to arbitrary missingness mechanisms, allowing direct learning from incomplete data while preserving the ability to perform amortized conditional density estimation for active information acquisition.","pith_inferences":["Generative modeling pipelines could skip separate imputation steps and train end-to-end on raw incomplete records.","The active acquisition routine could be paired with budgeted data collection where each observation carries a cost.","Similar missingness adjustments might apply to other order-agnostic sequence models beyond the architectures tested here."],"forward_implications":["Standard training on complete data already produces competitive imputation under high missing rates via the order-agnostic property.","A single training procedure works for any missingness mechanism without requiring complete observations.","The learned conditionals enable sequential choice of the next most informative variable to observe.","MO-ARM records lower imputation error than common baselines on multiple real-world datasets."],"fun_headline_variants":["Order-agnostic autoregressive models handle incomplete data","MO-ARM trains directly on datasets with missing values","Implicit imputation in order-agnostic autoregressive training","Missingness-aware order-agnostic autoregressive models","Order-agnostic models enable training on arbitrary missingness"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Missingness can be separated from the data-generating process so that the autoregressive conditionals remain unbiased when trained only on observed entries.","fun_headline_variants_meta":{"raw":{"variants":["Order-agnostic autoregressive models handle incomplete data","MO-ARM trains directly on datasets with missing values","Implicit imputation in order-agnostic autoregressive training","Missingness-aware order-agnostic autoregressive models","Order-agnostic models enable training on arbitrary missingness"]},"model":"grok-4.3","cost_usd":0.004614,"raw_usage":{"total_tokens":2243,"prompt_tokens":580,"num_sources_used":0,"completion_tokens":75,"cost_in_usd_ticks":46137000,"prompt_tokens_details":{"text_tokens":580,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1588,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":580,"tokens_out":75,"duration_ms":11262,"temperature":1.0,"reasoning_tokens":1588,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-30T23:23:57.569063+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"On a dataset generated with a known non-random missingness pattern, compare MO-ARM imputation error against a baseline trained only on complete data; if the errors are statistically indistinguishable, the advantage of the new framework is falsified.","supporting_citations":[],"review_version":2}