{"id":"f259a689-95b4-4634-8239-a01491ee35f7","arxiv_id":"2602.08735","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A 3B multimodal LLM trained with patch-level cross-view alignment plus explicit viewpoint-action reasoning outperforms much larger models on two multi-image spatial reasoning benchmarks.","lead":"This paper teaches a small multimodal language model to reason across multiple photos of the same scene by first aligning corresponding patches across views and then making it write out camera moves before answering. On two of three benchmarks the 3B model beats far larger models, but the third benchmark shows only a small gain and the training data overlaps the first evaluation set.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"HATCH trains on 10,000 samples drawn from SPAR-7M, while the headline benchmark SPAR-Bench-MV is extracted from the same SPAR suite; without a stated train/eval disjointness or matched baseline data, the +17.8 point gain is not a clean measure of the method.","rationale":"The paper's central claim requires that HATCH's empirical superiority reflects the two training objectives, not privileged access to the evaluation benchmark. The weakest point in that argument is the relationship between the training set (SPAR-7M, Section 4.1) and the primary evaluation set (SPAR-Bench-MV, footnote 1/Appendix E.1). No disjointness statement or split specification is provided, and the table footnote introduces an additional comparison mismatch: GPT-5.2's SPAR-Bench-MV number is on the Tiny subset, so 'matching GPT-5.2' is not a like-for-like comparison. This does not disprove the method—the ablations and MindCube-Tiny result offer independent support—but it does mean the headline quantitative claim is not yet fairly testable. The reader's weakest assumption identifies the same issue; my concern is therefore consistent with the CONDITIONAL verdict rather than moving it. A holdout retraining experiment with released sample IDs would settle the matter.","tokens_in":17831,"tokens_out":9068,"duration_ms":105160,"concrete_test":"From the public SPAR-7M and SPAR-Bench releases, build the set of HATCH's 10,000 training scene/question IDs; compute its intersection with SPAR-Bench-MV IDs. Then retrain HATCH on a fresh 10,000-sample SPAR-7M subset that excludes every SPAR-Bench-MV instance (and, ideally, all same-scene near-duplicates) and re-run Table 1. If SPAR-Bench-MV accuracy falls materially below 53.6 or toward the 46.8 SpatialLadder-3B level, the reported advantage is substantially an artifact of distribution overlap; if the number is stable, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 says HATCH's training data are 10,000 multi-image instances 'selected from SPAR-7M'; Appendix E.1 says SPAR-Bench-MV is formed by extracting multi-image samples from SPAR-Bench. The paper never states that the SPAR-Bench-MV evaluation items were excluded from that 10,000-sample training pool, nor that the comparison models (SpatialLadder-3B, Spatial-MLLM-4B, etc.) were trained on matched or absent SPAR-family data. The central quantitative support is the SPAR-Bench-MV jump from 24.9 to 53.6 and the claim of 'matching GPT-5.2' (52.6) on the same benchmark. If even a fraction of the eval items appear in training, or if the evaluation set is simply the same distribution as the training pool, this number conflates data exposure with the proposed PaStA/ActoR objectives. The paper therefore has not established the headline 'clear margin' on its primary benchmark.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HATCH, a two-stage training framework for multi-image spatial reasoning in multimodal large language models. In the first stage, PaStA uses camera poses, intrinsics, and depth maps to build a patch-level spatial-overlap matrix and aligns patch features across views through a cross-entropy objective. In the second stage, ActoR trains the model to emit explicit JSON viewpoint-transition actions before the final answer, using cold-start SFT followed by GRPO with action, answer, and format rewards. Experiments on SPAR-Bench-MV, MindCube-Tiny, and MMSI-Bench report substantial improvements over Qwen2.5-VL-3B and competitive performance against much larger models, together with ablations and single-image benchmark results.","tokens_in":18208,"tokens_out":5716,"duration_ms":64120,"significance":"If the reported numbers are trustworthy, the contribution is practically valuable: explicit geometry-supervised correspondence is a clean and low-cost training signal, and action-then-answer GRPO is a novel, interpretable intermediate objective. The MindCube-Tiny gain (+14.9) is on a benchmark outside the SPAR training suite, and the component ablations are broadly consistent with the paper's mechanistic story. However, the headline SPAR-Bench-MV result is not yet established because the paper does not demonstrate train/eval disjointness, selects a hyperparameter on the evaluation benchmark, and mixes full-benchmark and tiny-subset scores in a headline comparison. These issues must be resolved before the central claim can be accepted.","major_comments":[{"comment":"The paper trains on 10,000 multi-image samples 'selected from SPAR-7M' and evaluates on SPAR-Bench-MV, defined as multi-image samples 'extracted from SPAR-Bench'. Since SPAR-7M and SPAR-Bench come from the same SPAR suite, the paper must state explicitly that the SPAR-Bench-MV evaluation samples were excluded from the 10,000-sample training pool. It also never states whether the same-backbone baselines (SpatialLadder-3B, Spatial-MLLM-4B) were trained on matched, overlapping, or absent SPAR-family data. Without this information, the +17.8 point gain on SPAR-Bench-MV and the claim of matching GPT-5.2 are confounded by possible data exposure. Please provide the exact disjointness procedure and, ideally, report results on a held-out split constructed before training.","section":"§4.1, footnote 1, Appendix E.1"},{"comment":"The PaStA grid resolution n is selected by maximizing average accuracy on SPAR-Bench-MV, which is the same benchmark used for the main result and ablations. This is hyperparameter selection on the evaluation set, and the reported numbers with n=4 are therefore optimistically biased. Please select n on a validation split or fix it a priori; if that is infeasible, report the sensitivity across n for all benchmarks and qualify the headline results accordingly.","section":"Figure 4 / §4.3"},{"comment":"GPT-5.2 and Gemini-3-Pro are evaluated on SPAR-Bench-Tiny-MV (gray italic entries), not on the full SPAR-Bench-MV, yet the text states that 'HATCH (53.6%) matches GPT-5.2 (52.6%) on SPAR-Bench-MV'. These scores are not directly comparable. Either evaluate HATCH on the same SPAR-Bench-Tiny-MV subset, or rephrase the proprietary-model comparison and clearly mark it as approximate.","section":"Table 1 / §4.2"},{"comment":"For image patches with no geometric overlap, S[i,:]=0, so p(j|i) becomes a uniform distribution over all patches in the other image. Unless such rows are masked or explicitly handled, the cross-entropy loss in Eqs. (4)-(5) trains the encoder to produce uniform similarity for non-corresponding patches, which is contrary to the intended alignment. Please specify how zero-overlap rows are treated and provide an ablation or analysis showing the effect of masking them.","section":"§3.3, Eq. (2)"},{"comment":"No error bars, confidence intervals, or multiple-seed results are reported. On MMSI-Bench the overall gain is only +1.4 points (25.6 to 27.0), and several ablation differences are between 0 and 4 points; without variance estimates these differences cannot be interpreted. Please report at least three independent training runs or bootstrap confidence intervals for the key comparisons, especially the small MMSI-Bench gain and the ablation deltas.","section":"Tables 1-4 / §4.1"}],"minor_comments":[{"comment":"The baseline name 'LLaV A-OneVision-4B' contains an awkward spacing; use the standard name consistently.","section":"Table 1 / Table 4"},{"comment":"'Qwen-2.5-VL-3B' appears in one place while the rest of the paper uses 'Qwen2.5-VL-3B'; standardize the notation.","section":"§4.2"},{"comment":"The hyperparameter values for τ1, τ2, t, τt, and τr are not listed in the implementation-details appendix. Please provide these values for reproducibility.","section":"Appendix D"},{"comment":"The caption says 'success and failure cases' but the surrounding text and figure appear to show two successful examples. Clarify whether a failure case is included and what it illustrates.","section":"Figure 5 caption"},{"comment":"The example JSON action list ends with a trailing comma after the final 'turn_down_deg' object. Although likely a formatting artifact, it should be cleaned up to avoid ambiguity in the schema description.","section":"Appendix B.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's central idea is promising and the MindCube-Tiny result is encouraging, but the SPAR-Bench-MV headline is currently unverifiable because of the missing train/eval disjointness statement and the evaluation-set hyperparameter selection. I would ask the authors for a clear data-split disclosure, a matched-baseline comparison, and a re-evaluation with the proprietary models on the same subset. These are fixable within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a sensible two-part training recipe—geometry-supervised patch alignment plus action-then-answer RL—and the ablation work suggests both halves matter. But the central quantitative claim rests on a benchmark that may share data or at least distribution with the training set, and the paper never tells you. That needs to be fixed before the numbers can be taken at face value.\n\nWhat's new: nobody else, as far as the citations show, explicitly supervises both cross-view patch correspondence and viewpoint-action generation in one pipeline. The PaStA objective is a clean use of camera poses and depth at training time only. ActoR with verifiable action and answer rewards is a reasonable extension of GRPO to spatial reasoning. The paper also does the right ablations: removing either component hurts, action-based reasoning beats free-form CoT under the same RL framework, and single-image performance doesn't collapse. MindCube-Tiny, which is outside the SPAR family, shows a solid +12-15 point gain, so there is real signal here independent of SPAR.\n\nWhere it gets soft. The +17.8 point jump on SPAR-Bench-MV is the headline, but the training set (10k samples 'selected from SPAR-7M') and the eval set (SPAR-Bench-MV, 'extracted from SPAR-Bench') are from the same benchmark family. The paper never states that the evaluation items were excluded from the training pool, nor whether the comparison models were trained on SPAR-family data. If the eval items are in the training distribution, the number is inflated by exposure, not just by the method. This is not a claim of fraud—just a missing control that the paper should have provided. Also, the grid size n=4 is chosen based on SPAR-Bench-MV average accuracy (Figure 4), which is hyperparameter selection on the test set. That's a real but minor leak. No error bars anywhere, and no code or data release, so the numbers can't be reproduced.\n\nThe core idea is coherent and the non-SPAR result gives it credibility. But as written, the paper overstates its primary benchmark. It needs a simple fix: specify the exact split, verify no overlap, report multiple seeds, and pick hyperparameters on a validation set.\n\nThis is worth a serious referee—the training recipe is novel enough and the MindCube result is promising. But the referee should push for the disjointness statement and the test-set tuning disclosure. I'd want to see those before trusting the SPAR-Bench-MV claim.","headline":"Potentially useful training recipe, but the headline SPAR-Bench-MV result is not yet cleanly established because training and evaluation data come from the same SPAR family, and the split is never disclosed.","tokens_in":18645,"tokens_out":2598,"would_cite":false,"duration_ms":29003,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that explicit geometry-supervised patch alignment plus action-then-answer reinforcement lifts a 3-billion-parameter multimodal model 14.2 points on average across three multi-image spatial reasoning benchmarks and matches m","keywords":["multi-image spatial reasoning","cross-view correspondence","patch-level alignment","action-then-answer reasoning","GRPO reinforcement learning","viewpoint transformation","multimodal large language models","verifiable rewards"],"falsifier":"Re-run HATCH on the same 10,000-sample split, then evaluate on a held-out multi-view test set from which every training sample and near-duplicate has been removed, while training comparably sized baselines on the same split; if the reported multi-view gain (for example, the +17.8 point rise on the primary benchmark) collapses to noise, the central claim is falsified, and if it persists, the overlap concern is not the source.","tokens_in":17761,"feed_emoji":"🧭","tokens_out":13989,"duration_ms":132628,"temperature":0.7,"pith_summary":"The paper is trying to establish that the missing ingredient in multi-image spatial reasoning for multimodal large language models is not model size but explicit training signals for two human-like mechanisms: seeing which regions in different views are the same place, and composing a sequence of viewpoint changes. To make that case, it introduces HATCH, a training recipe with two stages. The first stage uses training-only camera poses and depth maps to build soft patch-to-patch correspondence targets and drives the image encoder to align features for corresponding regions. The second stage forces the model to write explicit viewpoint-transition actions (turns and moves) before the answer and reinforces both action and answer with verifiable rewards. On three multi-image benchmarks, a 3-billion-parameter model trained this way improves by 14.2 points on average and reaches parity with much larger systems on two benchmarks; a sympathetic reader would care because it suggests spatial reasoning can be bought with better supervision rather than more parameters.","feed_headline":"Cross-view training lifts a 3B model 14.2 points on spatial tasks","feed_subtitle":"A 3B model learns to align views and emit camera actions, rivaling models 10x its size on two benchmarks","key_machinery":"The load-bearing object is the symmetric patch-overlap matrix S: entry S[i,j] measures how consistently patch i in one view projects into patch j in the other, using depth and camera poses. PaStA converts S into a soft target distribution and trains the vision encoder to match it from patch-feature cosine similarities. The second mechanism is the explicit action sequence: the model must output JSON camera operations for every image pair before the answer, and GRPO rewards action accuracy, answer accuracy, and format. Together they force correspondence and stepwise viewpoint composition into the training signal.","core_discovery":"The central claim is that cross-view correspondence can be supervised directly at the patch level, and that explicit action generation then converts alignment into reliable answers. PaStA uses depth maps, camera intrinsics, and poses to build a symmetric patch-overlap matrix, turns it into a soft target, and trains the image encoder so patch-feature similarities reproduce it. ActoR makes the model emit JSON camera actions before the answer and optimizes those actions with group relative policy optimization under answer, action, and format rewards. Reported gains: 24.9→53.6, 37.8→50.2, and 25.6→27.0 on the three benchmarks; ablations show both stages matter, with PaStA more important for view","pith_inferences":["One consequence the authors leave implicit: the same geometry-derived soft target plus verifiable-action recipe may transfer to other perception tasks that decompose into correspondence plus composition, such as object tracking across viewpoints or embodied navigation.","Because the action vocabulary is limited to rotations and translations, a natural extension is object-centric actions (for example, 'move toward object X'), which might help the benchmark categories where all models sit near chance.","The grid-resolution result—accuracy peaks at a 4-by-4 patch grid and falls at finer resolutions—suggests adaptive or attention-based correspondence supervision could push the method further; the paper itself lists attention supervision as future work.","A natural stress test is to re-run with an explicit train/test split and matched baselines, since the training data is drawn from the same benchmark suite as the evaluation."],"forward_implications":["A 3-billion-parameter model can match much larger proprietary systems on multi-view spatial reasoning, so capacity is not the binding constraint.","Camera poses and depth are used only at training time; inference needs only images and text, so the recipe fits existing deployment pipelines.","Explicit JSON actions outperform free-form natural-language chain-of-thought as intermediate reasoning on the tested spatial questions, because actions are quantitative and verifiable.","PaStA and ActoR are complementary: removing PaStA mainly hurts viewpoint-change questions, while removing ActoR mainly hurts depth and distance questions.","The single-image benchmarks improve along with the multi-image ones, so the training does not trade away monocular spatial understanding."],"fun_headline_variants":["HATCH: patch alignment plus camera actions boosts spatial reasoning","Multi-image spatial reasoning gets a HATCH: 3B model beats larger","Patch-level alignment and explicit actions lift spatial AI","How a 3B model rivals 10x larger with cross-view training","Align patches, emit actions: HATCH boosts multi-image reasoning"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The headline gains rest on the unstated assumption that the 10,000 multi-image training samples selected from the source dataset are disjoint from, or at least distributionally distinct from, the evaluation samples, and that the comparison models were trained on matched data; the paper does not report such an exclusion.","fun_headline_variants_meta":{"raw":{"variants":["HATCH: patch alignment plus camera actions boosts spatial reasoning","Multi-image spatial reasoning gets a HATCH: 3B model beats larger","Patch-level alignment and explicit actions lift spatial AI","How a 3B model rivals 10x larger with cross-view training","Align patches, emit actions: HATCH boosts multi-image reasoning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000566,"raw_usage":{"total_tokens":2515,"prompt_tokens":737,"completion_tokens":1778,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":1704}},"tokens_in":481,"tokens_out":1778,"duration_ms":13475,"temperature":1.0,"reasoning_tokens":1704,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T03:10:03.575589+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run HATCH on the same 10,000-sample split, then evaluate on a held-out multi-view test set from which every training sample and near-duplicate has been removed, while training comparably sized baselines on the same split; if the reported multi-view gain (for example, the +17.8 point rise on the primary benchmark) collapses to noise, the central claim is falsified, and if it persists, the overlap concern is not the source.","supporting_citations":[],"review_version":1}