{"id":"ffd6f592-3b06-40c1-a1d3-f0d611a05180","arxiv_id":"2608.00502","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SpatialAfford improves affordance grounding in a 4B VLM by first supervising cross-modal attention with the ground-truth region and then applying GRPO, outperforming several 7B+ baselines.","lead":"This paper presents SpatialAfford, a two-stage training method that aligns a compact vision-language model's attention to affordance regions before optimizing its bounding-box predictions, reporting large gains on affordance benchmarks with a 4B model. The finding matters because it suggests that explicitly teaching small models where to look can close part of the gap with much larger models in embodied AI tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing control: SAA may improve metrics as an unstructured attention regularizer, not through spatial alignment; a shifted-target training experiment would test causal role.","rationale":"The reader's weakest_assumption is exactly the load-bearing point: the SAA loss must actually shape attention in a way that mediates coordinate grounding, rather than being an uninterpretable regularizer. I agree with that assessment. The manuscript provides good empirical evidence that adding SAA improves metrics, but the mechanism is underdetermined. The proposed shifted-target experiment is the minimal control that separates 'spatial alignment' from 'generic regularization'. If the control shows no sensitivity to the target location, the paper's central mechanism collapses, even though the empirical gains might persist; that would change the contribution from a novel attention-grounded training paradigm to a minor loss-function tweak. If the control shows sensitivity, the core claim is substantially validated. Other issues—missing appendix details, lack of error bars, baseline number inconsistencies—are real but secondary; they affect reproducibility and comparison rigor, not the conceptual claim. Since the reader already issued a CONDITIONAL verdict, my recommendation is to keep it unchanged, with the concrete test as a condition for full acceptance.","tokens_in":12270,"tokens_out":4600,"duration_ms":58696,"concrete_test":"Train Stage 1 (SAA+SFT) with the SAA target PM constructed from a spatially incorrect box—e.g., the GT box shifted by 30% of image width, or a randomly permuted grid—while keeping LSFT, α, and all other hyperparameters unchanged. Then run Stage 2 GRPO as in the paper. Compare gIoU/P@50 on ShareRobot and AGD20K against the GT-aligned SAA. If performance stays within a few points of the GT-aligned run, SAA is acting as a generic regularizer and the causal mechanism fails. If performance drops substantially or predicted boxes systematically shift toward the shifted target, then attention alignment is causally involved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that aligning cross-modal attention to the ground-truth affordance box (LSAA, Eq. 7) causally improves affordance grounding, converting the objective into 'attention-grounded spatial reasoning.' Table IV shows SFT+SAA > SFT and > GRPO alone, and Fig. 5 shows sharper attention maps. However, this does not establish that the specific spatial location of the attention target is what matters. LSAA has two inseparable effects: (i) a spatial prior concentrated on the affordance region, and (ii) a generic regularizer that reduces attention entropy by pushing PS toward a peaked distribution PM. The α sensitivity sweep (Table V) and the coverage-loss comparison are consistent with a tunable regularizer; they do not isolate the spatial prior. Moreover, the extra LSAA term is added on top of LSFT, so any improvement could stem from additional attention-level supervision rather than from the 'where to look' semantics. The paper's own OOD diagnostic (Table III) shows KLD actually worsens slightly vs. Qwen3-VL-4B (9.35 vs. 9.29), which further undermines a straightforward 'attention becomes more aligned' story. Thus, the core mechanistic interpretation—that teaching the model to look at the affordance region is what improves grounding—is not yet supported. Without a control that uses an incorrect attention target, the improvement could be a regularizer artifact, and the central contribution would reduce to an uninterpretable auxiliary loss.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SpatialAfford, a two-stage training framework for compact vision-language models (VLMs) on affordance grounding. Stage 1, Spatial Attention Alignment (SAA), adds to the SFT objective a KL-divergence loss that pushes the model's cross-modal attention over visual tokens toward a smoothed ground-truth affordance-box mask (Eq. 7). Stage 2, Spatial-Aware GRPO, refines coordinate prediction with rewards based on IoU, L1 distance, and format validity. The reported experiments show large improvements over general-purpose and embodied VLMs, e.g., 74.92 gIoU / 82.50 P@50 on ReasonAff versus 67.41 / 74.50 for Affordance-R1, and a 4B model outperforming several 7B+ baselines. The paper also reports an OOD evaluation on AGD20K and an ablation study indicating that SAA and GRPO contribute complementary gains.","tokens_in":12651,"tokens_out":3389,"duration_ms":42708,"significance":"If the central causal claim holds—that explicitly aligning attention to the affordance region before policy optimization is what drives the grounding improvements—the work is a useful contribution to compact-VLM training for embodied affordance reasoning. The method is simple, the two-stage decomposition is well motivated, and the empirical gains over strong 7B+ baselines are substantial. The ablation in Table IV shows a clear component synergy, and the AGD20K OOD evaluation is a positive feature. The paper does not ship code or trained models, and the verification of the proposed 'where-to-look' mechanism is incomplete, so the significance is conditional on the missing causal controls being supplied.","major_comments":[{"comment":"The paper's central mechanistic claim is that the spatial location of the attention target matters, but SAA changes both the location and the peakedness/entropy of the attention distribution. The presented ablations (Table IV) and α-sweep (Table V) do not isolate the spatial prior. A control experiment using a shifted, random, or uniform attention target is necessary to show that the improvement comes from aligning attention to the affordance region rather than from a generic attention-regularization effect. Without this, the 'where to look' interpretation is not supported.","section":"§III-B, Eq. (7); §IV-D, Table IV"},{"comment":"The attention-alignment metrics in Table III are partly circular as evidence for SAA, since the SAA loss directly minimizes KL(PS||PM) on the training data. More importantly, the OOD results do not consistently show improved alignment: SpatialAfford's KLD (9.35) is slightly worse than the Qwen3-VL-4B baseline (9.29), even though SIM and NSS improve. The paper should reconcile this contradiction and avoid presenting Table III as unambiguous confirmation of better attention alignment.","section":"§IV-C, Table III"},{"comment":"The diagnostic supporting Eq. (2), the claimed attention–output mismatch, is deferred to an appendix that appears as 'Appendix??' and is not included. Likewise, training details (data mixture, hyperparameters, GRPO setup, metric definitions for cIoU and P@50–95) are referenced only as 'Appendix??'. This is a load-bearing omission: Eq. (2) motivates the entire method, and the missing implementation details prevent reproducibility and independent verification.","section":"§III-A and §IV-A; unreferenced appendices"},{"comment":"No error bars, standard deviations, or significance tests are reported for any table. The headline gains are large, but some comparisons are close, e.g., ShareRobot-Bench gIoU: SpatialAfford 47.80 vs RoboBrain2.5-8B 48.07. Without multiple seeds or statistical testing, the reader cannot assess the reliability of these comparisons; at minimum a statement that the runs are single-seed with deterministic decoding should be made.","section":"Tables I–IV"},{"comment":"The reward equations are inconsistent with the manuscript's own statement that training data contain single-bbox annotations. Eq. (11) defines RIoU using an optimal assignment M* between sets of predicted and ground-truth boxes, which is never used or explained for the single-box case; Eq. (12) then reverts to a single-box L1 distance. Please simplify the reward description to match the actual single-bbox setting and clarify how the 'optimal assignment' would apply.","section":"§III-C, Eqs. (11)–(14)"}],"minor_comments":[{"comment":"Multiple 'Appendix??' placeholders remain unresolved (e.g., in §§III-A, IV-A, IV-B). These must be filled before resubmission.","section":"Throughout"},{"comment":"The phrase 'norm1k' (also in Eq. 12) is undefined. Presumably '1k' denotes a 1/1000 normalized coordinate system; please define explicitly.","section":"§III-B, Eq. (3)"},{"comment":"Typo: 'red box is predicated results' should be 'predicted results'. Also standardize capitalization of 'RynnBrain' / 'Rynnbrain' across text, tables, and figures.","section":"Fig. 3 caption"},{"comment":"cIoU and P@50–95 are used in tables but never defined in the main text. Please define these metrics at first use or in the experiments section.","section":"§IV-A, metrics"},{"comment":"The coverage-loss comparison is interesting, but the paper should clarify why coverage loss degrades grounding despite improving attention alignment. A brief explanation would strengthen the choice of KL divergence.","section":"§IV-D, Table V"}],"recommendation":"major_revision","confidential_remarks":"The paper has a clear, potentially valuable core idea and strong raw results, but the central mechanistic claim is not yet convincingly disentangled from generic attention regularization. The missing shifted-target control is the key experiment needed; without it, the contribution could be viewed as an auxiliary-loss trick rather than the proposed 'where to look' mechanism. The unresolved appendix references and lack of error bars add to the risk. I am not recommending rejection because the empirical results are strong and the needed experiments are well-defined and feasible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a within-subfield training recipe that might be useful, but the paper oversells its mechanism and leaves too many loose ends. What's new is the specific combination: a box-derived attention alignment loss (SAA) followed by GRPO with an IoU/L1/format reward. The ablation shows a real boost from SAA on top of SFT, and a further boost from GRPO on top of SAA. If the numbers hold up, this is a practical way to squeeze affordance grounding out of a 4B VLM. The writing is clear and the benchmarks are reasonable, including an OOD set. The soft spots are in proportion. The main one: the causal claim that aligning attention to the affordance region is what helps is not supported. SAA directly optimizes KLD/SIM/NSS, so those metrics are partly by construction. There is no control with a shifted or incorrect attention target, so the improvement could come from any peaky regularizer rather than from the spatial location. The paper's own OOD table is mixed: KLD is slightly worse than the base model, so 'attention becomes more aligned' isn't uniformly true. The attention-output mismatch diagnostic in Eq. 2 and the NSS numbers are relegated to a missing appendix, which is a load-bearing gap. No error bars or significance tests appear anywhere, and some baseline numbers are inconsistent between Table I and Table IV. The abstract claims 'outperforming stronger 7B+ baselines,' but on ShareRobot gIoU it trails RoboBrain2.5-8B, and it trails on OOD gIoU/cIoU too. That should be reworded. None of this kills the empirical contribution. The ablation is internally consistent, the gains are large, and the recipe is concrete enough for others to try. The paper is for researchers working on compact VLM training for embodied grounding; it deserves a serious referee, provided the authors complete the appendix, report variance, and either add the shifted-target control or soften the mechanistic language. I would send it to review with a request for major revision.","headline":"A plausible two-stage training recipe for compact VLM affordance grounding, but the 'where to look' mechanism is not isolated and the manuscript needs completion before the causal story is credible.","tokens_in":634,"tokens_out":2237,"would_cite":true,"duration_ms":55538,"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":"SpatialAfford claims that teaching a compact VLM where to look—by aligning its cross-modal attention to the ground-truth affordance region—before optimizing where to ground lets a 4B model outperform stronger 7B+ baselines.","keywords":["affordance grounding","compact vision-language models","cross-modal attention alignment","Group Relative Policy Optimization","spatial reasoning","embodied AI","attention-output mismatch","reinforcement learning"],"falsifier":"Train SpatialAfford with SAA applied to a surrogate mask that is not the true affordance box—e.g., a random box or the object's center—while keeping Stage 2 fixed. If gIoU/P@50 gains on ReasonAff and PartAfford persist, the specific affordance semantics of the alignment target are not what drives the improvement. As a second check, after SAA, freeze the model and perturb the extracted attention map at inference; if coordinate outputs do not move, cross-modal attention is not mediating the prediction.","tokens_in":12217,"feed_emoji":"🎯","tokens_out":10616,"duration_ms":117388,"temperature":0.7,"pith_summary":"SpatialAfford claims that compact vision-language models fail at affordance grounding not because they cannot output coordinates, but because their cross-modal attention is not anchored to the functional region before decoding. The paper proposes a two-stage fix: first train the model to concentrate attention inside the ground-truth affordance box (Spatial Attention Alignment), then refine the coordinates with a spatial-reward reinforcement step (Spatial-Aware GRPO). On three affordance benchmarks this ordering lifts a 4B model above several 7B+ baselines, with the clearest win on reasoning-heavy ReasonAff. The claim matters for embodied AI because it suggests small models can be made reliable by supervising internal visual focus, not just final predictions.","feed_headline":"Attention-first training puts a 4B model ahead of 7B+ rivals","feed_subtitle":"Aligning visual focus to the functional region before refining boxes beats bigger models on three benchmarks.","key_machinery":"The load-bearing object is the cross-modal attention matrix from text instruction tokens to visual patch tokens. Spatial Attention Alignment converts the ground-truth affordance box into a spatial target distribution P_M by rasterizing the box on the vision encoder's grid, smoothing it with an adaptive kernel sized to the box, and normalizing; it then penalizes KL(P_S || P_M) between the model's normalized attention map and that target, with attention extracted only from instruction tokens to avoid teacher-forcing leakage. This turns 'where to look' into a supervised spatial-reasoning objective. Stage 2 then runs GRPO with a composite reward (mean IoU, mean L1 coordinate distance, format val","core_discovery":"SpatialAfford's central claim is that the attention–output mismatch in compact autoregressive VLMs is the bottleneck for affordance grounding, and that it is fixable by directly supervising the intermediate attention distribution before policy optimization. The method distills the ground-truth affordance box into a smoothed probability map over visual tokens and minimizes the KL divergence between that map and the model's text-to-image attention over instruction tokens; only after this alignment does it apply GRPO with an IoU-plus-L1-plus-format reward. Reported results: on ReasonAff, SpatialAfford (4B) reaches 74.92 gIoU and 82.50 P@50, beating Affordance-R1 (7B) by +7.51 and +8.00; on Shar","pith_inferences":["An editorial control that the paper does not run: replace the ground-truth affordance mask in SAA with a random or object-centered mask; if the gains on ReasonAff and PartAfford persist, the specific affordance semantics of the target are not what drives the improvement.","An editorial transfer prediction: the same 'align attention with a spatial target, then run spatial-reward RL' ordering should help other fine-grained spatial tasks, such as part-level detection, referring segmentation, and dense grounding, where output-level supervised fine-tuning is shortcut-prone.","An editorial flag on evidence: the paper's motivating diagnosis—attention density inside the ground-truth region is suppressed with near-zero or negative NSS—is cited to an appendix (Appendix??) that is not present in this version, so the attention–output mismatch is asserted rather than directly inspectable; the benchmark gains themselves do not depend on that figure.","An editorial scalability note: because SAA needs only a box annotation, the same loss could be trained on pseudo-labels, such as saliency maps or heatmaps from a larger model, in settings where true affordance boxes are expensive."],"forward_implications":["A compact 4B model can beat several 7B–8B systems on ShareRobot-Bench, ReasonAff, and PartAfford by adding attention alignment before RL, narrowing the parameter gap in embodied affordance grounding.","The ablation shows the order is the mechanism: SAA alone reaches 42.59 gIoU and 40.50 P@50, GRPO alone reaches 42.45 and 45.55, and only SAA followed by GRPO reaches 47.80 and 55.00, so neither stage substitutes for the other.","OOD results on AGD20K give P@50 37.44 and NSS 1.05, suggesting the learned visual focus transfers to unseen egocentric scenes rather than overfitting training boxes.","The KL form of the attention loss is load-bearing: a coverage loss improves attention alignment but degrades coordinate prediction (P@50 32.15 vs 40.50 at alpha=0.1)."],"supporting_citations":[{"why":"Supplies the Qwen3-VL-4B backbone whose attention–output mismatch is diagnosed and whose baseline scores are compared against.","marker":"Bai et al., 2025"},{"why":"Supplies the Affordance-R1 baseline and the ReasonAff-Bbox training/evaluation source that SpatialAfford must beat.","marker":"Wang et al., 2026"},{"why":"Contributes the GRPO update used in Stage 2 and the group-relative advantage formulation.","marker":"Shao et al., 2024"},{"why":"Contributes ShareRobot-Affordance training data and the ShareRobot-Bench baseline used in the main table.","marker":"Ji et al., 2025"},{"why":"Supplies PartAfford part-level functional annotations used for training and evaluation.","marker":"Xu et al., 2022"},{"why":"Provides AGD20K, the out-of-distribution pixel-level benchmark used for attention-alignment diagnostics.","marker":"Luo et al., 2021"},{"why":"Shows cross-modal attention heatmaps carry affordance cues, supporting the choice of attention as the supervision target.","marker":"Lee et al., 2026"},{"why":"Demonstrates that supervising visual-token attention improves grounding, the prior result SAA builds on.","marker":"Esmaeilkhani and Latecki, 2026"}],"fun_headline_variants":["Teach attention before grounding: 4B VLM beats 7B+","Align visual focus first, then refine boxes: 4B > 7B+","Attention supervision gives 4B model edge over 7B+ baselines","Look where it matters: 4B VLM outgrounds 7B+ models","SpatialAfford: attention-first training beats bigger VLMs"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the model's internal 'where it looks' map is a genuine intermediate representation of visual focus, so steering that map toward the ground-truth region actually improves grounding; if that map is just a byproduct of language priors or token-position quirks, the attention-alignment stage would be an uninterpretable regularizer and the reported gains would not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Teach attention before grounding: 4B VLM beats 7B+","Align visual focus first, then refine boxes: 4B > 7B+","Attention supervision gives 4B model edge over 7B+ baselines","Look where it matters: 4B VLM outgrounds 7B+ models","SpatialAfford: attention-first training beats bigger VLMs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000396,"raw_usage":{"total_tokens":1928,"prompt_tokens":774,"completion_tokens":1154,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":1050}},"tokens_in":518,"tokens_out":1154,"duration_ms":9966,"temperature":1.0,"reasoning_tokens":1050,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T00:49:04.130518+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train SpatialAfford with SAA applied to a surrogate mask that is not the true affordance box—e.g., a random box or the object's center—while keeping Stage 2 fixed. If gIoU/P@50 gains on ReasonAff and PartAfford persist, the specific affordance semantics of the alignment target are not what drives the improvement. As a second check, after SAA, freeze the model and perturb the extracted attention map at inference; if coordinate outputs do not move, cross-modal attention is not mediating the prediction.","supporting_citations":[],"review_version":1}