{"id":"9ec56272-7a57-4fb3-a30f-8a513f4f2da7","arxiv_id":"2606.24178","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Reframing test-time canonicalization as OOD detection allows systematic use of ~20 OOD scores and 9 search algorithms, with distance-based scores plus random search plus local refinement performing best, plus a gated mechanism to preserve in-distribution accuracy.","lead":"This paper shows that test-time canonicalization of transformed images can be achieved by minimizing out-of-distribution scores over geometric transformations, without retraining the classifier. A smart generalist might read it for a practical way to add robustness to existing vision models against rotations, scales, and shears.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"The reader's weakest assumption is precisely what the reported experiments measure (accuracy after canonicalization vs. original). No hidden modeling assumption, unstated identifiability condition, or evaluation circularity is visible in the described protocol that would undermine the empirical ranking of scoring functions and optimizers.","tokens_in":1749,"tokens_out":278,"duration_ms":26649,"concrete_test":"Recompute the main accuracy tables using only the subset of inputs whose original OOD score already lies below the gating threshold; if accuracy on this subset drops by more than 1-2 points after forced canonicalization, the gate's necessity is confirmed and the core correlation holds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is an empirical result: among ~20 OOD scores and 9 search procedures, distance-based scores with random search plus local refinement yield the highest downstream classification accuracy on transformed inputs across character, sketch, image, and point-cloud benchmarks. The paper directly measures this accuracy (not merely OOD score reduction) and introduces a gate that withholds transformation when the original input already has low OOD score. Because the evaluation is end-to-end on the quantity the claim cares about, the assumption that lower OOD implies better canonical alignment is tested rather than merely posited.","agreement_with_reader":"disagree"},"referee_report":{"model":"grok-4.3","summary":"The paper reframes test-time canonicalization of affine-transformed inputs (rotations, scales, shears) as minimization of an out-of-distribution (OOD) score, allowing any existing OOD detector to serve as the objective without modifying the downstream classifier. It reports a systematic comparison of approximately twenty OOD scores and nine search procedures across benchmarks from handwritten characters and sketches to natural images and 3D point clouds, concluding that distance-based scores combined with random search plus local refinement yield the highest downstream classification accuracy. A gated mechanism is added that applies the transformation only when the original input's OOD score indicates misalignment is present.","tokens_in":1845,"tokens_out":448,"duration_ms":19548,"significance":"If the reported ordering of methods holds, the work supplies a practical, model-agnostic route to robustness that reuses off-the-shelf OOD detectors and requires no retraining or architectural changes. The systematic enumeration of the design space, the end-to-end measurement of classification accuracy rather than proxy OOD reduction alone, and the introduction of the gate to protect in-distribution performance are the primary contributions. Public code release aids verification and extension.","major_comments":[],"minor_comments":[{"comment":"Abstract states 'around twenty' OOD scores; the main experimental section should list the exact set of scores and the precise search algorithms evaluated so that the performance ordering can be reproduced without ambiguity.","section":"Abstract"},{"comment":"The description of the gated mechanism indicates that transformation occurs only when the OOD score 'indicates this is needed,' but the exact threshold selection procedure, its sensitivity analysis, and the fraction of inputs transformed on in-distribution data should be reported with quantitative detail.","section":null},{"comment":"Benchmarks are summarized at a high level (handwritten characters, sketches, natural images, 3D point clouds); the manuscript should include a table or appendix specifying the exact datasets, the ranges and distributions of the applied affine transformations, and any exclusion criteria used for the reported results.","section":null}],"recommendation":"minor_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the detailed summary of our work and the positive evaluation of its contributions, including the systematic comparison, end-to-end accuracy measurements, gated mechanism, and code release. The recommendation for minor revision is appreciated. No major comments were listed in the report, so we have no specific points requiring rebuttal or revision at this stage. We remain available to address any minor suggestions or clarifications that may arise.","responses":[],"tokens_in":1300,"tokens_out":103,"duration_ms":5810,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main takeaway is that treating canonicalization as OOD score minimization works when you pick distance-based scores, use random search with local refinement, and add the gate to skip already-aligned inputs.\n\nThe paper opens the design space by mapping the problem to any OOD score instead of limiting to logit energies. The sweep over roughly twenty scores and nine optimizers is new relative to earlier canonicalization papers, and they measure actual downstream classification accuracy rather than just score reduction. The gated mechanism is a practical addition that preserves clean accuracy. Evaluation spans handwritten characters, sketches, natural images, and 3D point clouds, which is a reasonable spread, and code is released.\n\nSoft spots are limited. The link between lower OOD score and better alignment is tested end-to-end via accuracy, so the central assumption is not merely posited. Still, the size of the gains and their consistency would be clearer with the full tables and error bars. It would also help to see more on when the correlation between OOD score and classifier improvement breaks.\n\nThis is for researchers working on model-agnostic robustness fixes in vision. Readers focused on OOD methods or test-time adaptation will find the comparisons useful. It has enough new empirical ground and clean framing to deserve a serious referee.","headline":"Distance-based OOD scores with random search plus local refinement and a gate give the strongest test-time canonicalization results across the tested benchmarks.","tokens_in":2344,"tokens_out":313,"would_cite":false,"duration_ms":19737,"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":"Reframing test-time canonicalization as out-of-distribution scoring allows any OOD score to guide search for an affine transformation that aligns an input with the training distribution.","keywords":["test-time canonicalization","out-of-distribution detection","affine transformations","zero-shot robustness","image classification","point cloud processing","gated transformation"],"falsifier":"Finding a benchmark where the transformation that minimizes the chosen OOD score produces lower classification accuracy than the untransformed input would falsify the approach.","tokens_in":2645,"feed_emoji":"🔄","tokens_out":636,"duration_ms":15703,"temperature":0.7,"pith_summary":"The paper shows that pretrained classifiers can be made robust to rotations, scales, and shears at inference time by searching for a transformation that minimizes an out-of-distribution score, without altering the model itself. This turns canonicalization into a generic optimization problem where any OOD detector serves as the objective. Across handwritten characters, sketches, natural images, and 3D point clouds, distance-based scores combined with random search and local refinement give the strongest gains. A gating step applies the transformation only when the original input scores as out-of-distribution, avoiding accuracy drops on already-aligned examples.","feed_headline":"OOD scores guide zero-shot canonicalization of images and clouds","feed_subtitle":"Distance-based detectors with random search plus local refinement restore accuracy on rotated and scaled inputs while a gate protects in-dis","key_machinery":"Minimization of an out-of-distribution score over affine transformations of the input, using a gated threshold to decide whether to apply the result.","core_discovery":"By treating the search for a canonical form as minimization of an out-of-distribution score over candidate affine transformations, the method enables zero-shot robustness using unmodified pretrained classifiers. Systematic evaluation of around twenty scores and nine search procedures identifies distance-based scores with random search plus local refinement as the strongest combination across benchmarks, while a simple gate on the original input's score prevents harmful transformations on in-distribution data.","pith_inferences":["The approach could be tested on non-affine transforms such as perspective changes or elastic deformations if suitable OOD scores exist.","It raises the possibility that OOD detection and geometric robustness share deeper structure than separate research lines suggest.","Replacing random search with a learned optimizer might further reduce the number of forward passes needed per input."],"forward_implications":["Any pretrained vision model gains robustness to affine transforms without retraining or architectural changes.","The space of usable scoring functions expands to all existing OOD detectors rather than being limited to logit-based energies.","A single gated mechanism preserves accuracy on aligned inputs while recovering performance on transformed ones.","The same search procedure works across 2D images and 3D point clouds without domain-specific tuning."],"fun_headline_variants":["OOD scores enable zero-shot canonicalization","Canonicalization by minimizing OOD scores","Distance-based OOD scores canonicalize inputs","Gated scoring protects in-distribution accuracy"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"A lower out-of-distribution score after transformation reliably signals a form that improves the downstream classifier's accuracy.","fun_headline_variants_meta":{"raw":{"variants":["OOD scores enable zero-shot canonicalization","Canonicalization by minimizing OOD scores","Distance-based OOD scores canonicalize inputs","Gated scoring protects in-distribution accuracy"]},"model":"grok-4.3","cost_usd":0.007464,"raw_usage":{"total_tokens":3435,"prompt_tokens":684,"num_sources_used":0,"completion_tokens":50,"cost_in_usd_ticks":74637000,"prompt_tokens_details":{"text_tokens":684,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2701,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":684,"tokens_out":50,"duration_ms":20205,"temperature":1.0,"reasoning_tokens":2701,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T00:56:22.860926+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Finding a benchmark where the transformation that minimizes the chosen OOD score produces lower classification accuracy than the untransformed input would falsify the approach.","supporting_citations":[],"review_version":1}