{"id":"d42f6026-bdb9-4884-881c-09b4d9dbd480","arxiv_id":"2608.10981","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"ThinkAfford decomposes 3D affordance grounding into high-recall proposal generation and GRPO-trained vision-language selection, reporting state-of-the-art AP25 on SceneFun3D.","lead":"ThinkAfford is a new pipeline for task-driven 3D affordance grounding that first generates many fine-grained interaction proposals from 2D heatmaps and then uses a GRPO-trained vision-language model to pick the proposal that matches the instruction. On SceneFun3D it reports the best AP25 (25.46) and AP50 (10.69) among the compared baselines, though the gain at strict overlap thresholds is small.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim hinges on unreleased self-reproductions of Fun3DU, TASA, AffordBot, and Mask3D-F; without a faithful baseline run the headline margin is unverified.","rationale":"The reader's weakest assumption targeted the lifted-IoU reward and the rho=0.25 threshold. That is a real concern for the internal diagnostics, but the end-to-end official AP is an independent check: if the lifting were badly biased, the final 3D mask would be penalized by the official evaluator. The more load-bearing point is the external validity of the SOTA comparison itself. The paper's strongest claim is comparative ('outperforming comparable baselines'), and four of six baselines are self-reproductions with no artifacts released. The omission of OpenMask3D-F, the benchmark's functional baseline, compounds this. This does not require assuming bad faith: reproduction of a complex VLM pipeline (view pruning, point prompting, mask lifting) is easy to get subtly wrong, and the paper provides no way to check. The internal ablations (Tables 2 and 4) are consistent and support the value of APG and GRPO within their own pipeline, so the correct verdict remains CONDITIONAL rather than REJECT: the design may well be effective, but the SOTA claim should not be accepted until the baseline comparison is independently reproducible. I therefore keep the reader's verdict unchanged while shifting attention from the reward-lifting assumption to the baseline validity assumption.","tokens_in":11302,"tokens_out":10584,"duration_ms":102382,"concrete_test":"Require the authors to release the exact baseline-reproduction artifacts (checkpoints, prompts, view-selection settings) and rerun the four self-reproduced baselines and OpenMask3D-F on the SceneFun3D validation split using the official releases and the same 25-view protocol and official evaluator; if any properly run baseline matches or exceeds ThinkAfford's 25.46 AP25 / 10.69 AP50, the SOTA claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is state-of-the-art fine-grained 3D grounding on SceneFun3D, supported by Table 1. In that table, four of the six competitors (Mask3D-F, TASA, AffordBot, Fun3DU) are described as 'reproduced by us and converted to the official format'; OpenMask3D and LERF are taken from Fun3DU's split0. No reproduction code, prompt templates, per-baseline view budgets, or VLM choices are released, and the paper does not state which underlying VLM/segmentor each baseline was run with. Since ThinkAfford uses Qwen-VL for parsing and VPAR, a reproduction that feeds a weaker VLM or fewer views to Fun3DU/AffordBot would inflate the reported margins (+12.90 AP25 over Fun3DU, +4.32 AP50 over AffordBot). The related work cites OpenMask3D-F, the SceneFun3D benchmark's own functional 3D baseline, but Table 1 omits it. This is a correctness risk, not a matter of taste: if the baseline runs are not faithful, the headline 'outperforming comparable baselines' is unsupported even though the internal ablations and decomposition may be sound.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"ThinkAfford proposes a two-module decomposition for task-driven 3D affordance grounding on SceneFun3D: the Affordance Proposal Generation (APG) module produces variable-length, interaction-level 2D proposals from affordance-conditioned heatmaps, trained with visibility-aware projections of 3D ground-truth masks, and the Visual-Prompted Affordance Reasoning (VPAR) module, a VLM fine-tuned with GRPO using proposal-level rewards computed from lifted 3D overlap, selects proposal identifiers in a think-then-answer format; the selected masks are lifted and fused by weighted voting into a final 3D mask. On the official SceneFun3D validation evaluator the method reports 25.46 AP25 and 10.69 AP50, above the compared baselines, with module-level diagnostics of 77.5% R@25 proposal coverage and 72.1% selection accuracy on APG-covered queries. The paper additionally reports zero-shot transfer to four self-annotated cross-dataset sets and a qualitative real-robot deployment.","tokens_in":11649,"tokens_out":9538,"duration_ms":82938,"significance":"If the results hold, the paper's contribution is a clean and separately diagnosable separation of proposal coverage from instruction-grounded selection, plus a GRPO training scheme whose reward is anchored in lifted 3D overlap rather than in parsed object names. The ablations are internally consistent, with the largest drops when removing GRPO and APG, and the main evaluation uses the official SceneFun3D evaluator; I also credit the authors' candor, since Sec. 4.3 explicitly calls the qualitative examples illustrative rather than evidence of solved boundary errors, Sec. 4.4 states that coverage and selection numbers are not numerically composable, and Sec. 4.6 disclaims the deployment as not a manipulation success-rate study. The main risks are verification gaps rather than circularity: four baselines in Table 1 are unreleased self-reproductions, the GRPO reward shares its positive threshold with the definition of the APG-covered subset, and no variance estimates accompany small ablation deltas. These issues are fixable within the manuscript's scope, which leads me to major revision rather than rejection.","major_comments":[{"comment":"The headline claim of state-of-the-art results rests on four self-reproductions (Mask3D-F, TASA, AffordBot, Fun3DU), but the manuscript does not state for any of these baselines which VLM or segmentor was used, how many views were processed, or which prompt templates were employed, so the reported margins (+12.90 AP25 over Fun3DU, +4.32 AP50 over AffordBot) cannot be checked by readers, and the possibility that weaker components were used in those runs is a concrete correctness risk for the comparative claim. Please release the full reproduction configuration (model versions, view budgets, prompts, conversion code) and, as a minimal check, re-run Fun3DU and AffordBot with a VLM of capability comparable to the one used in ThinkAfford's VPAR. In addition, the related work cites OpenMask3D-F, the SceneFun3D benchmark's own functional-3D baseline, but Table 1 omits it; include it in the table or give an explicit reason for its exclusion.","section":"Sec. 4.1, Table 1"},{"comment":"The positive threshold ρ=0.25 plays a double role: it defines the GRPO reward target set C⋆v (Sec. 3.3: 'Proposals with ov,j ≥ ρ form the target set') and it defines the APG-covered subset on which Table 4 measures selection accuracy, since samples whose best proposal fails to reach ρ are excluded from VPAR training and from the selection-accuracy denominator. This coupling means that the 72.1% selection accuracy and the GRPO-vs-SFT gains are established only under a coarse IoU criterion where proposals that would fail AP50 still count as positive; the paper's own observation that the averaged-AP margin over Fun3DU is only +0.52 (Table 1) is consistent with weak high-IoU alignment, so the claim that GRPO aligns VPAR selection with final 3D grounding needs support from a sensitivity analysis at ρ=0.5 or with a graded IoU reward, together with end-to-end AP reported separately on the covered and non-covered subsets.","section":"Sec. 3.3, Eq. (2); Table 4"},{"comment":"All reported numbers come from a single training run of APG and a single GRPO run of VPAR; several component claims rest on small deltas (GRPO without format reward −0.56 AP25, uniform view weights −1.26 AP25, APG without visibility-aware projection −1.66 AP25), and the text describes these as 'consistent gains' without any variance estimate, so signal and noise cannot be separated at these magnitudes, even though the large main-effect deltas (removing GRPO, −8.66 AP25; replacing APG with parsed-object Grounded-SAM, −7.16 AP25) are likely robust. Please run at least three seeds for the full model and for the two largest ablation variants, and report mean ± std or a paired test for the Table 2 deltas.","section":"Sec. 4.4, Tables 2 and 4"}],"minor_comments":[{"comment":"The heatmap threshold γ and the DBSCAN eps and min_samples values used in Proposal Extraction are never reported, which prevents reproduction of the variable-length proposal sets that underlie all coverage and selection numbers.","section":"Sec. 3.2"},{"comment":"λfmt is described only as 'small'; please report its exact value and the exact format-reward definition, since Eq. (2) and the corresponding ablation in Table 2 depend on it.","section":"Sec. 3.3, Eq. (2)"},{"comment":"It is unclear whether K=25 is the number of views retained after the visibility and relation pruning or the uniform sampling interval; clarify the relationship between the pruning stage and K.","section":"Sec. 3.1 vs. Sec. 4.1"},{"comment":"The 20-scene, 15-instruction cross-dataset set has no description of the annotation protocol, no inter-annotator agreement statistics, and no release plan; without these, the transfer claim is not independently checkable.","section":"Sec. 4.5, Table 5"},{"comment":"Please state why the AP column is '–' for OpenMask3D and LERF and clarify that AP is the 0.50:0.95 average, so that the table is self-contained for readers.","section":"Table 1 caption"},{"comment":"APG is trained with the ground-truth affordance type a but used at inference with the VLM-parsed type a⋆, and the manuscript never measures how often a⋆ is correct; report this parsing accuracy, since parsing errors propagate directly into the APG heatmaps.","section":"Sec. 3.2"},{"comment":"The figure contains a stray text fragment ('画一段video') that appears to be a leftover annotation; remove it or translate it into the paper's language.","section":"Figure 1"},{"comment":"The sentence about overlapping relation subsets appears immediately after Table 1, but Table 1 contains no relation slices; move this sentence to the discussion of Tables 3 and 4.","section":"Sec. 4.2"},{"comment":"Several typographical errors remain: 'Yeshwanthetal.' is missing a space, the Radford reference has a broken journal rendering ('PmL R'), and 'laser_scan_5mmsupport' appears in the Sec. 4.4 prose.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is honestly written and the internal evidence is consistent; the circularity concern does not land, because the rewards come from external 3D ground-truth masks on the training split and the official evaluation is on the held-out validation split. For the editor: the decisive issue is verification of the comparative claim — Table 1's self-reproductions need released configurations and comparable component choices, and the omitted OpenMask3D-F row should be addressed. I would resist rejection because all major concerns are addressable within the manuscript's scope (a ρ-sensitivity experiment, multiple seeds for the key variants, baseline release details, and annotation details for the cross-dataset set). No scope mismatch for this venue; the qualitative robot section is appropriately hedged and should not be over-weighted in the decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things upfront. The paper's internal logic is sound, and its ablations are among the clearer ones I've seen in this area. The external comparison, however, is the soft spot: the headline gains over Fun3DU and AffordBot come from self-reproductions that are not released.\n\nThe genuinely new piece is the explicit split between heatmap-based affordance proposal generation (no object-name prompts) and a GRPO-trained VLM that selects among labeled overlays. The module diagnostics support the split: APG reaches 77.5% R@25, and GRPO lifts selection accuracy from 63.4% to 72.1% over SFT. Those numbers line up with the end-to-end ablations, where removing GRPO or swapping APG for parsed-object Grounded-SAM costs the most AP25. The authors also use the official SceneFun3D evaluator and are honest about the strict-IoU ceiling.\n\nThe stress-test concern lands: Table 1 lists four baselines as 'reproduced by us' with no prompt templates, view budgets, or VLM choices, and no code. If those runs were not faithful, the +12.9 AP25 over Fun3DU could shrink. The paper's own ablations are less affected because they vary one component under the same pipeline, so the decomposition story likely holds. Still, the lack of error bars matters for the small AP-averaged margin (+0.52). Several hyperparameters (gamma, DBSCAN eps, lambda_fmt, beta) are unspecified, and the module-level numbers are computed on the APG-covered subset—fine as diagnostics, but not directly comparable to end-to-end scores. Also, OpenMask3D-F is cited but missing from Table 1; that omission should be fixed.\n\nWho this is for: anyone working on functional grounding or embodied instruction following. The decomposition is worth discussing even if the exact numbers don't fully replicate. It deserves a serious referee: the claim is clear, the ablations are controlled, and the mechanism is plausible. I would not desk reject it. I would ask for code, exact baseline configurations, and error bars on the main table before trusting the SOTA margin.","headline":"A sensible two-stage decomposition with believable internal ablations, but the SOTA claim rests on unreleased baseline reproductions—worth refereeing, not desk rejecting.","tokens_in":12127,"tokens_out":1759,"would_cite":true,"duration_ms":17630,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ThinkAfford claims that fine-grained 3D affordance grounding improves when proposal generation and instruction reasoning are decoupled, reaching 25.46 AP25 and 10.69 AP50 on the official SceneFun3D evaluator.","keywords":["3D affordance grounding","task-driven grounding","proposal generation","visual reasoning","group relative policy optimization","vision-language models","fine-grained 3D grounding","cluttered scenes"],"falsifier":"Take a random stratified sample of queries per affordance type, manually inspect the lifted APG proposals against the ground-truth contact regions, and rerun the official evaluator with the positive IoU threshold raised from 0.25 to 0.4; a sharp coverage drop or frequent visible misalignment of thin parts would weaken the central claim that APG generates high-recall interaction-level proposals.","tokens_in":11148,"feed_emoji":"🎯","tokens_out":9388,"duration_ms":77680,"temperature":0.7,"pith_summary":"The paper is trying to show that fine-grained 3D affordance grounding fails for two separable reasons, and that addressing them separately is better than addressing either alone. The target in this task is a small interaction region, such as a drawer handle or a socket, so a proposal that returns the whole object is wrong even if it is semantically correct. ThinkAfford splits the pipeline: the Affordance Proposal Generation module deliberately over-produces fine-grained candidates from interaction-conditioned heatmaps, and the Visual-Prompted Affordance Reasoning module selects identifiers from a labeled overlay using the full instruction. A Group Relative Policy Optimization (GRPO) objective rewards selections by their lifted 3D overlap with ground truth, teaching the reasoner to handle spatial, ordinal, and part-whole relations. On the official SceneFun3D evaluation the full system reaches 25.46 AP25 and 10.69 AP50, and module-level diagnostics report 77.5% proposal coverage at IoU 0.25 and 72.1% selection accuracy on covered queries.","feed_headline":"Search-then-reason split lifts 3D affordance grounding by 12.9 points","feed_subtitle":"Decoupling proposal search from language reasoning cuts misses and misselections, beating prior methods by 12.9 points.","key_machinery":"The load-bearing mechanism is the labeled proposal overlay together with a discrete selection action. APG turns each affordance type into a heatmap, thresholds it, and clusters connected components into a variable-length set of candidates, each tagged with a letter identifier, so the reasoning module operates on a finite list of physically distinct regions rather than generating open-ended masks. The second carrying mechanism is the GRPO reward: every sampled think-then-answer response is parsed into identifiers, each proposal is lifted to 3D with a visibility-aware z-buffer and local-plane disambiguation, and point-level IoU at threshold $\\rho=0.25$ defines which proposals count as covering the target; group-relative advantages then let the policy learn relational distinctions from the differences among similar candidates.","core_discovery":"ThinkAfford's central claim is that task-driven 3D grounding should be a two-stage process with distinct measures: first generate a high-recall set of interaction-level proposals, then reason over those proposals to select the correct ones. The paper argues that this decomposition isolates the two failure modes that plague existing systems, missing the interaction region and confusing visually similar alternatives, and that each stage can then be optimized with an appropriate supervision. APG uses learnable affordance prompts with multi-level visual features to predict heatmaps and extract a variable number of small candidates without needing parsed object or part names; VPAR consumes labeled overlays and returns identifiers in a think-then-answer format, trained with GRPO where proposal-level rewards come from lifting each candidate to 3D and comparing point-level IoU with ground truth. The reported numbers support the claim that both stages matter: replacing APG with parsed-object segmentation costs 7.16 AP25, and removing GRPO costs 8.66 AP25.","pith_inferences":["Editorial inference: the same search-then-select decomposition could generalize to other fine-grained localization tasks with many small repeated parts, such as part-level picking in assembly or identifying interaction points in surgical scenes, wherever a high-recall candidate stage and a language-guided selection stage can be separated.","Editorial inference: a stress test with the positive threshold raised from $\\rho=0.25$ to, say, 0.4 would reveal how much of the reported advantage is robust target identification rather than loose-overlap tolerance; the paper does not report this sweep.","Editorial inference: the qualitative robot demonstrations set up a measurable downstream claim, that higher selection accuracy should raise first-attempt manipulation success when the localized contact point is fed to a controller, but the paper does not evaluate that success rate."],"forward_implications":["Systems should track proposal coverage and selection accuracy separately; an end-to-end score alone cannot show whether errors come from missed candidates or wrong choices.","Interaction-conditioned heatmaps can replace parsed object or part names as the prompt for segmenting functional regions, avoiding the granularity mismatch that produces whole-object masks.","GRPO with lifted 3D overlap rewards is a viable way to teach a vision-language model relational instruction following without hand-built negatives.","The smaller gain at stricter IoU thresholds indicates that once the right region is found, boundary tightness is the remaining bottleneck.","The complementary residual errors, 22.5% proposals missed at IoU 0.25 and 27.9% misselection on covered queries, imply that further progress needs both better proposal tightness and better relation-aware selection."],"supporting_citations":[{"why":"Supplies the base OOAL architecture that APG adapts with affordance prompts and multi-level feature fusion.","marker":"Li et al. 2024a"},{"why":"CoOp-style learnable prompts make the frozen encoders affordance-specific in APG.","marker":"Zhou et al. 2022"},{"why":"Defines GRPO, the group-relative policy optimization used to train VPAR.","marker":"Shao et al. 2024"},{"why":"The Fun3DU baseline and the point-to-mask 2D-to-3D route that ThinkAfford is compared against and improves on.","marker":"Corsetti et al. 2025"},{"why":"Provides the SceneFun3D task, annotations, and official evaluator used for all main results.","marker":"Delitzas et al. 2024"},{"why":"DINOv2 frozen visual encoder is the APG image feature backbone.","marker":"Oquab et al. 2023"},{"why":"CLIP frozen text encoder provides the affordance-language conditioning in APG.","marker":"Radford et al. 2021"},{"why":"Scene-R1 demonstrates GRPO for visual grounding, motivating its use for discrete proposal selection.","marker":"Yuan et al. 2025"},{"why":"AffordBot is the strongest VLM-based 2D-to-3D AP50 comparison behind ThinkAfford.","marker":"Wang et al. 2026"}],"fun_headline_variants":["ThinkAfford: two-stage reasoning for fine-grained 3D grounding","Decouple proposal search from language reasoning, gain 12.9 points","Search-then-reason approach lifts 3D grounding by 12.9 points","Two-stage affordance reasoning gains 12.9 points in 3D grounding","ThinkAfford: decouple search and reasoning for sharper 3D grounding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything depends on the assumption that lifting 2D proposal masks into the 3D point cloud faithfully matches the true visible interaction region; if the z-buffer and local-plane disambiguation bias that mapping, the training rewards and all reported coverage and selection numbers shift.","fun_headline_variants_meta":{"raw":{"variants":["ThinkAfford: two-stage reasoning for fine-grained 3D grounding","Decouple proposal search from language reasoning, gain 12.9 points","Search-then-reason approach lifts 3D grounding by 12.9 points","Two-stage affordance reasoning gains 12.9 points in 3D grounding","ThinkAfford: decouple search and reasoning for sharper 3D grounding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000794,"raw_usage":{"total_tokens":3568,"prompt_tokens":1085,"completion_tokens":2483,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":701,"completion_tokens_details":{"reasoning_tokens":2381}},"tokens_in":701,"tokens_out":2483,"duration_ms":17868,"temperature":1.0,"reasoning_tokens":2381,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:54:01.040431+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random stratified sample of queries per affordance type, manually inspect the lifted APG proposals against the ground-truth contact regions, and rerun the official evaluator with the positive IoU threshold raised from 0.25 to 0.4; a sharp coverage drop or frequent visible misalignment of thin parts would weaken the central claim that APG generates high-recall interaction-level proposals.","supporting_citations":[],"review_version":1}