{"id":"b302c1fd-1364-4a08-9a9e-2e80d0adea20","arxiv_id":"2607.00580","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"Active Spatial Guidance replaces injected positional embeddings in ViTs with a training-only 2D coordinate regression loss on final-layer tokens, yielding better results than learned absolute or rotary embeddings on ImageNet-100, ADE20K, and Hypersim under matched training.","lead":"The paper proposes Active Spatial Guidance, a training-only auxiliary 2D coordinate regression loss that lets Vision Transformers learn spatial structure from data without any injected positional embeddings or mechanisms. If this holds, it could simplify ViT architectures by removing a core design choice while maintaining or improving results on standard vision tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Auxiliary 2D regression loss may improve optimization independently of inducing spatial token organization","rationale":"The reader's weakest_assumption already isolates the exact causal attribution issue; the proposed control directly tests whether that assumption is necessary. No other internal inconsistency (e.g., in the from-scratch protocol or backbone choice) appears more load-bearing on the evidence provided.","tokens_in":1753,"tokens_out":313,"duration_ms":12013,"concrete_test":"Train an additional control arm that uses the identical guidance head architecture and loss weight but regresses to randomly permuted or constant 2D targets instead of true patch coordinates; compare final task metrics on ImageNet-100 against both the true-Guidance and the injected-positional baselines under the same protocol. If the random-target arm matches true Guidance, the spatial-organization mechanism is not required for the observed gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that gains arise specifically because the coordinate-regression objective organizes patch tokens spatially in the absence of any injected positional mechanism. The method adds an auxiliary head and loss during training only; the injected baselines receive no equivalent auxiliary supervision. If the performance delta is driven by the presence of any auxiliary loss term (or by altered gradient flow/regularization) rather than by the 2D spatial target itself, the claim that spatial inductive bias can be shaped purely through training-time supervision does not hold. The abstract and reader note do not indicate a control that isolates this variable.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes Active Spatial Guidance, a training-only auxiliary 2D coordinate-regression loss on final-layer patch tokens of a positional-injection-free ViT. Under matched from-scratch training with DINOv3 backbones, Guidance outperforms injected baselines (learned absolute and rotary positional embeddings) on ImageNet-100 classification, ADE20K segmentation, and Hypersim depth estimation; the head is removed at inference. Additional results cover resolution transfer and multi-resolution training, with public code released.","tokens_in":1880,"tokens_out":454,"duration_ms":19635,"significance":"If the central empirical claim holds after controls, the work shows that spatial inductive bias need not be architecturally injected and can instead be shaped by training-time supervision, which would simplify ViT design and potentially improve robustness. Public code is a clear strength for reproducibility.","major_comments":[{"comment":"The interpretation that gains arise specifically from learned spatial organization (rather than from the mere presence of any auxiliary loss) requires an ablation with a non-spatial auxiliary target (e.g., regression to random coordinates or a mismatched task). No such control is described in the abstract or experimental sections; without it the causal link between the 2D spatial objective and the observed deltas remains unisolated.","section":"Abstract, §3 (method), §4 (experiments)"},{"comment":"§4.1 and Table 1: the reported gains are stated to be under identical training protocols, yet no values are given for the auxiliary-loss weight, number of random seeds, or statistical testing; these omissions make it impossible to assess whether the performance deltas are robust or sensitive to hyper-parameter choices.","section":"§4.1, Table 1"}],"minor_comments":[{"comment":"Notation for the guidance head and its removal at inference could be formalized with a short equation or diagram in §3 to avoid ambiguity.","section":"§3"},{"comment":"Figure captions should explicitly list the exact baseline configurations (e.g., “RoPE with same training schedule”) for quick cross-reference.","section":"Figures 2–4"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the detailed and constructive comments. We address each major point below and will revise the manuscript to incorporate the suggested improvements for clarity and rigor.","responses":[{"response":"We agree that an ablation with a non-spatial auxiliary target (such as regression to random coordinates) would help isolate whether the performance gains are specifically due to the spatial 2D coordinate objective rather than the presence of any auxiliary loss. While the design of Active Spatial Guidance is motivated by spatial supervision, this control experiment was not included in the original submission. We will add the requested ablation to the revised experimental section to strengthen the causal claim.","revision_made":"yes","referee_comment":"[Abstract, §3 (method), §4 (experiments)] The interpretation that gains arise specifically from learned spatial organization (rather than from the mere presence of any auxiliary loss) requires an ablation with a non-spatial auxiliary target (e.g., regression to random coordinates or a mismatched task). No such control is described in the abstract or experimental sections; without it the causal link between the 2D spatial objective and the observed deltas remains unisolated."},{"response":"We acknowledge that the auxiliary-loss weight, number of random seeds, and any statistical testing details were not explicitly reported in §4.1 or Table 1, which limits assessment of robustness. These values are fixed across all compared methods and documented in the released code, but we agree they should be stated in the paper. We will add the specific hyper-parameter values, seed counts, and relevant statistics to the revised §4.1 and Table 1.","revision_made":"yes","referee_comment":"[§4.1, Table 1] §4.1 and Table 1: the reported gains are stated to be under identical training protocols, yet no values are given for the auxiliary-loss weight, number of random seeds, or statistical testing; these omissions make it impossible to assess whether the performance deltas are robust or sensitive to hyper-parameter choices."}],"tokens_in":1412,"tokens_out":440,"duration_ms":15395,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core idea is straightforward: drop all injected positional mechanisms from a ViT, add an auxiliary head that regresses 2D patch coordinates from the final-layer tokens, train with that loss, then remove the head at inference. They run this on DINOv3 backbones from scratch and report better numbers than absolute or rotary embeddings on ImageNet-100 classification, ADE20K segmentation, and Hypersim depth.\n\nWhat stands out is the controlled from-scratch protocol across three tasks and the public code. That setup lets readers check whether the auxiliary objective can substitute for architectural injection, which is a clean question.\n\nThe main weakness is the missing control the stress-test note flags. The baselines get no auxiliary loss at all, so any extra supervision or gradient signal could explain the delta rather than the specific 2D coordinate target. Without an ablation that adds a non-spatial auxiliary loss to the injected baselines, the claim that spatial organization is induced purely through supervision rests on an assumption rather than a direct test. Loss weighting and statistical details are also thin in the abstract.\n\nThis is worth a referee for groups working on transformer inductive biases and positional encodings. The experimental scope is reasonable and the question is well-posed, even if the causal link needs tightening. I would send it out rather than desk-reject.","headline":"The paper shows you can train ViTs without positional embeddings by adding a coordinate regression loss at training time, but the evidence does not yet isolate whether the spatial target itself drives the gains.","tokens_in":2332,"tokens_out":334,"would_cite":false,"duration_ms":7388,"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":"Vision transformers can learn spatial organization from an auxiliary coordinate loss without any positional embeddings.","keywords":["vision transformers","positional embeddings","spatial guidance","auxiliary loss","image classification","semantic segmentation","depth estimation"],"falsifier":"A controlled re-run in which models trained with Guidance show equal or lower accuracy than matched models that retain absolute or rotary embeddings on the same three tasks would falsify the central claim.","tokens_in":2662,"feed_emoji":"","tokens_out":697,"duration_ms":18348,"temperature":0.7,"pith_summary":"The paper asks whether vision transformers need injected positional mechanisms to capture spatial structure in images, or whether such organization can arise from data supervision alone. It introduces Active Spatial Guidance, a training-only auxiliary loss that regresses 2D coordinates from final-layer patch tokens while disabling all positional injection. Under matched from-scratch training on DINOv3 backbones, the resulting positional-injection-free models outperform strong baselines that use learned absolute or rotary embeddings on ImageNet-100 classification, ADE20K segmentation, and Hypersim depth estimation. The guidance head is removed at inference, leaving a standard ViT encoder plus task head. The results indicate that spatial inductive bias need not be architecturally supplied but can instead be shaped by training-time supervision.","feed_headline":"Auxiliary loss replaces positional embeddings in vision transformers","feed_subtitle":"Coordinate regression during training organizes patch tokens and raises accuracy on classification, segmentation, and depth tasks.","key_machinery":"Active Spatial Guidance: a training-only auxiliary 2D coordinate-regression loss on final-layer patch tokens that induces spatial organization without any positional injection mechanism.","core_discovery":"Active Spatial Guidance disables positional injection entirely and applies an auxiliary 2D coordinate-regression loss only to the final-layer patch tokens during training. The guidance head is discarded at inference, so the deployed model is a pure positional-injection-free ViT encoder plus task module. On ImageNet-100, ADE20K, and Hypersim, this yields higher accuracy than matched models that use learned absolute positional embeddings or rotary positional embeddings, with further gains in resolution-transfer robustness when multi-resolution training is added.","pith_inferences":["If the coordinate loss truly organizes tokens, analogous auxiliary supervision might allow removal of positional encodings from other transformer domains such as language modeling.","The approach could simplify architecture search by reducing the need to design or tune positional mechanisms.","The method suggests that self-attention's permutation invariance can be countered by data-driven supervision rather than by explicit position signals."],"forward_implications":["Guidance improves ImageNet-100 top-1 accuracy over learned absolute and rotary positional embeddings under identical training.","Guidance raises mIoU on ADE20K semantic segmentation and reduces depth error on Hypersim relative to the same injected baselines.","Models trained with Guidance exhibit better accuracy when test resolution differs from training resolution.","Multi-resolution training further improves Guidance models across a range of input sizes.","Spatial inductive bias in ViTs can be supplied by training supervision instead of architectural injection."],"fun_headline_variants":["Coordinate regression replaces injected positional embeddings in ViTs","Training-only loss enables positional-injection-free vision transformers","ViTs gain spatial regularity from auxiliary 2D regression loss","Active guidance shapes spatial bias without architectural injection"],"cache_read_input_tokens":64,"weakest_assumption_plain":"The performance gains are caused by genuine spatial organization learned via the auxiliary loss rather than by uncontrolled differences in optimization or regularization.","fun_headline_variants_meta":{"raw":{"variants":["Coordinate regression replaces injected positional embeddings in ViTs","Training-only loss enables positional-injection-free vision transformers","ViTs gain spatial regularity from auxiliary 2D regression loss","Active guidance shapes spatial bias without architectural injection"]},"model":"grok-4.3","cost_usd":0.004911,"raw_usage":{"total_tokens":2426,"prompt_tokens":708,"num_sources_used":0,"completion_tokens":59,"cost_in_usd_ticks":49112000,"prompt_tokens_details":{"text_tokens":708,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1659,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":708,"tokens_out":59,"duration_ms":13059,"temperature":1.0,"reasoning_tokens":1659,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-02T14:50:59.364737+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A controlled re-run in which models trained with Guidance show equal or lower accuracy than matched models that retain absolute or rotary embeddings on the same three tasks would falsify the central claim.","supporting_citations":[],"review_version":1}