{"id":"1696aac3-c47b-4c35-bd38-c87b15f42cca","arxiv_id":"2607.06176","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":6,"one_line_summary":"A dual-query scene graph generation method unifies detector-based and query-based reasoning in a single decoder, achieving state-of-the-art results on Visual Genome, Open Images v6, and GQA-200.","lead":"This paper shows that detector-based and query-based scene graph generation methods fail in complementary ways, and combines both into a single model that outperforms prior work on three benchmarks. A smart generalist might read it to understand how to fuse different AI reasoning strategies without complex post-hoc calibration.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Joint training may not truly solve the inter-branch score calibration problem it identifies as motivation; NMS-based fusion could be masking residual scale mismatch between TD-Qs and BU-Qs predictions.","rationale":"The reader's weakest assumption about the external detector potentially being an artifact is a reasonable concern, but the paper addresses it with the Swin-B experiment (Table 5, bottom rows), where complementarity persists even with a stronger detector reducing UDet-T from 35% to 28%. This makes the reader's concern less load-bearing than it appears. My concern targets a different, more central issue: the paper's specific claim that joint training eliminates the need for post-hoc score calibration (Section 3, Eq. 1 motivation). The inference pipeline uses NMS, which is inherently robust to inter-branch scale differences, so the system could work well even if the branches remain miscalibrated. This means the paper's stated reason for why joint training works ('enabling joint optimization within a single framework' to avoid scale inconsistency) may not be the actual reason the system succeeds. That said, this concern does not undermine the empirical results — Dual-SGG achieves strong performance on three benchmarks with competitive speed. The architecture is well-motivated and the ablations (Tables 5-7) are thorough. The concern is about the precision of the causal claim (joint training solves calibration) rather than the validity of the overall approach. The verdict remains ACCEPT because the empirical contributions stand regardless of whether the calibration explanation is the precise mechanism. The code link being a placeholder ('Dual-SGG' rather than a URL) slightly limits reproducibility but does not affect the scientific assessment.","tokens_in":17319,"tokens_out":3780,"duration_ms":272852,"concrete_test":"On the VG test set, extract the raw triplet scores (s × o × p) for all true-positive predictions from TD-Qs and BU-Qs separately. Compute the score distributions for each branch (e.g., histogram, mean, median). Then, for triplets predicted by BOTH branches (overlapping detections), compare the relative ranking: in what fraction of cases does the TD-Qs score exceed the BU-Qs score? If one branch systematically dominates (e.g., >70% of overlapping predictions have higher scores from one branch), then the branches are not well-calibrated despite joint training, and the 'no post-hoc calibration' claim is only viable because NMS is robust to scale differences — not because joint training solved the calibration problem the paper identifies.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central architectural motivation (Section 3, Eq. 1 and surrounding text) is that directly applying max(S_d, S_q) to independent branches is problematic because 'score scale inconsistencies' would cause one branch to dominate. The proposed solution is joint training in a single decoder, which allegedly eliminates the need for post-hoc calibration. At inference (Sec. 3.5), predictions from TD-Qs and BU-Qs are concatenated, NMS is applied, and TopK is selected — effectively implementing the max operation from Eq. (1). However, NMS only compares scores for spatially overlapping triplets, so it is robust to global scale differences between branches: even if TD-Qs systematically produce higher scores than BU-Qs (or vice versa), NMS would still work because it only needs to rank within overlapping pairs, not across the full set. This means the inference pipeline could function well even if the two branches are NOT well-calibrated relative to each other, which would weaken the specific claim that joint training eliminates the calibration problem. Table 6 provides some evidence: normalization fusion (which explicitly calibrates) achieves 32.2/37.3 R@K vs Dual-SGG's 33.5/38.5, a gap of ~1.3 points. This gap could arise from joint training providing better feature sharing or gradient flow — benefits unrelated to score calibration per se. The paper does not analyze the actual score distributions of TD-Qs vs BU-Qs predictions to verify that joint training aligns their scales. If the branches remain miscalibrated but NMS masks this, the 'no post-hoc calibration needed' claim is technically true but for a different reason than stated.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This paper introduces a controlled experimental framework for analyzing the predictive behaviors of detector-based and query-based scene graph generation (SGG) methods from the perspective of 'detector-conditioned reachability.' Using an external object detector as a discriminator, the authors split ground-truth triplets into detected (Det-T) and undetected (UDet-T) subsets, revealing that detector-based methods excel on Det-T but collapse on UDet-T, while query-based methods show the complementary pattern. Motivated by this finding, the paper proposes Dual-SGG, which integrates both reasoning mechanisms within a single triplet decoder via a dual-query design: Top-Down Triplet Queries (TD-Qs) conditioned on detected entity pairs, and Bottom-Up Triplet Queries (BU-Qs) that globally explore triplets from center-biased initialization. The method is evaluated on Visual Genome, Open Images v6, and GQA-200, achieving state-of-the-art results while maintaining competitive inference speed (14.0 FPS).","tokens_in":17644,"tokens_out":1733,"duration_ms":310104,"significance":"The detector-conditioned reachability analysis (Tab. 1, Tab. 5) provides a useful diagnostic lens for understanding when and why detector-based versus query-based SGG methods fail, and the complementarity finding is well-motivated. The Dual-SGG architecture is a clean integration of both paradigms, and the ablation studies (Tab. 5–7) are thorough, including a strong-detector-condition analysis with Swin-B (Tab. 5, starred rows) and a complementarity comparison against post-hoc fusion baselines (Tab. 6). The method ships reproducible code and achieves SOTA on three datasets with competitive efficiency. The self-attention mask design to prevent TD-to-BU information leakage (Sec. 3.4, Tab. 7) is a thoughtful architectural detail. The paper provides falsifiable predictions: the micro-DR/micro-UDR decomposition quantitatively demonstrates that Dual-SGG maintains detector-based performance on Det-T while gaining on UDet-T.","major_comments":[{"comment":"Sec. 3, Eq. (1) and surrounding text: The paper motivates the unified decoder by arguing that directly applying max(S_d, S_q) to independent branches is problematic due to 'score scale inconsistencies' causing one branch to dominate. The proposed solution is joint training in a single decoder to eliminate the need for post-hoc calibration. However, at inference (Sec. 3.5), predictions from TD-Qs and BU-Qs are concatenated, NMS is applied, and TopK is selected. NMS only compares scores for spatially overlapping triplets, so it is inherently robust to global scale differences between branches — even if TD-Qs systematically produce higher scores than BU-Qs, NMS would still function correctly because it only needs to rank within overlapping pairs, not across the full set. This means the inference pipeline could work well even if the two branches are NOT well-calibrated relative to each other","section":null},{"comment":"Sec. 3, Eq. (1) continued: The gap between Dual-SGG (33.5/38.5 R@K) and normalization fusion (32.2/37.3 R@K) in Tab. 6 is approximately 1.3 points, but this gap could arise from joint feature sharing or gradient flow benefits unrelated to score calibration per se. The paper does not analyze the actual score distributions of TD-Qs vs BU-Qs predictions to verify that joint training aligns their scales. Without such analysis, the specific claim that joint training eliminates the calibration problem identified in the motivation remains unverified. The authors should either (a) provide score distribution analysis showing that joint training does align TD-Q and BU-Q score scales, or (b) soften the motivation to acknowledge that joint training provides multiple benefits (feature sharing, gradient flow, reduced architectural complexity) beyond score calibration.","section":null},{"comment":"Sec. 1 and Tab. 1: The detector-conditioned reachability split (Det-T vs UDet-T) is defined using an external object detector as a discriminator. The reader's weakest assumption concern is valid: if the external detector's recall characteristics differ significantly from the internal detectors used in the compared models (e.g., the detector-based model in Fig. 1a fine-tunes its own detector), the observed complementarity could partly reflect the specific detector choice rather than a fundamental property of the two reasoning paradigms. The paper does partially address this with the Swin-B experiment (Tab. 5, starred rows), showing robustness under a stronger detector. However, the external detector used for the Det-T/UDet-T split in Tab. 1 is the same Deformable-DETR that is then fine-tuned within the models, which creates a potential circularity: the discriminator is not fully external.","section":null},{"comment":"Sec. 1 and Tab. 1 continued: The authors should clarify whether the Det-T/UDet-T split is computed on the pre-trained detector (before SGG fine-tuning) or after fine-tuning, and discuss whether the complementarity finding is robust to the choice of discriminator detector. This is load-bearing because the entire architectural motivation rests on the complementarity claim.","section":null}],"minor_comments":[{"comment":"Sec. 3.2, Eq. (6): The EPS uses different scoring at training (sigmoid(S_sp) × sigmoid(S_ct)) vs inference (sigmoid(S_ct)). The rationale for dropping the spatial score at inference is stated briefly but could benefit from a more explicit justification.","section":null},{"comment":"Tab. 2: The #params for Dual-SGG is 84.7M, but the backbone (ResNet50) and detector parameters are not broken down. It would help readers to know how much of this is attributable to the triplet decoder vs the object detector.","section":null},{"comment":"Sec. 3.3: The BU-Qs content embeddings are 'randomly initialized with a set of learnable vectors.' It would be useful to specify the dimensionality and whether these are shared across images or image-conditioned.","section":null},{"comment":"Fig. 2: The EPS architecture diagram (a) is somewhat dense. Labeling the data flow more explicitly (e.g., marking which arrows correspond to S_sp vs S_ct) would improve readability.","section":null},{"comment":"Sec. 4.3, Tab. 5: The baseline model (Fig. 1a) reports R@50 of 31.6, while in Tab. 1 the same model reports micro-R of 32.5. The difference is presumably due to Top-100 vs Top-50 evaluation, but this should be clarified.","section":null},{"comment":"Sec. 3.4: The statement 'the predicate cost is excluded from the matching cost calculation, as its computational overhead outweighs its empirical benefit' could use a quantitative justification or reference.","section":null},{"comment":"Tab. 6: The 'Average' fusion strategy performs dramatically worse (17.6 R@50) than all other methods. A brief explanation of why averaging fails so catastrophically would help readers understand the score scale issue.","section":null},{"comment":"The paper references code availability ('Code is available at: Dual-SGG') but the actual URL is not provided in the text.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The stress-test concern about NMS robustness to score scale differences is technically valid and worth raising, but it does not undermine the paper's core contribution: Dual-SGG is a well-designed architecture that achieves SOTA results with thorough ablations. The calibration claim is a motivational framing issue rather than a load-bearing error — the method works regardless of whether the mechanism is specifically calibration or feature sharing. I recommend minor revision with the request to either provide score distribution analysis or adjust the motivational framing. The self-citation [34] (Salience-SGG, WACV'26) is from the same group and is recent; the authors should ensure this work is distinct from the current submission in terms of contribution scope."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for the careful reading and constructive feedback. The referee's comments center on two themes: (1) whether the motivation for joint training in a unified decoder—specifically the claim about eliminating score calibration problems—is fully supported by the evidence presented, and (2) whether the detector-conditioned reachability split (Det-T/UDet-T) is robust to the choice of external detector, given that the discriminator detector is the same architecture later fine-tuned within the models. We address both themes below. In brief, we agree that the score calibration claim should be softened and supplemented with empirical analysis, and we agree that the circularity concern regarding the discriminator detector warrants explicit clarification in the manuscript. We propose partial revisions addressing both points.","responses":[{"response":"The referee raises a valid and technically precise point. We agree that NMS, as applied in our inference pipeline, operates on spatially overlapping triplets and is therefore robust to global scale differences between the TD-Q and BU-Q branches. The referee is correct that the inference pipeline could function adequately even without perfect inter-branch calibration, because NMS only needs to rank within overlapping pairs rather than across the full candidate set. Our original motivation overstates the necessity of joint training for score calibration at inference time. We will revise the manuscript to acknowledge this explicitly. That said, we note that the calibration concern is not entirely vacuous: the final TopK selection step does rank across the full candidate pool (after NMS deduplication), so systematic scale differences could still influence which triplets survive into the final output. However, this is a weaker claim than what we originally stated, and we will adjust the text accordingly.","revision_made":"partial","referee_comment":"Sec. 3, Eq. (1) and surrounding text: The paper motivates the unified decoder by arguing that directly applying max(S_d, S_q) to independent branches is problematic due to 'score scale inconsistencies' causing one branch to dominate. The proposed solution is joint training in a single decoder to eliminate the need for post-hoc calibration. However, at inference (Sec. 3.5), predictions from TD-Qs and BU-Qs are concatenated, NMS is applied, and TopK is selected. NMS only compares scores for spatially overlapping triplets, so it is inherently robust to global scale differences between branches — even if TD-Qs systematically produce higher scores than BU-Qs, NMS would still function correctly because it only needs to rank within overlapping pairs, not across the full set. This means the inference pipeline could work well even if the two branches are NOT well-calibrated relative to each other"},{"response":"We accept this criticism. The 1.3-point gap between Dual-SGG and normalization fusion in Tab. 6 could indeed arise from multiple sources—joint feature sharing, gradient flow, shared decoder representations—rather than from score scale alignment alone. We did not provide score distribution analysis for TD-Qs vs. BU-Qs, and without that analysis, the specific claim that joint training eliminates the calibration problem is not empirically verified. We will adopt option (b): we will soften the motivation in Sec. 3 to acknowledge that joint training in the unified decoder provides multiple benefits beyond score calibration, including shared feature representations, gradient flow between the two reasoning pathways, and reduced architectural complexity. We will also add the requested score distribution analysis (histograms of TD-Q vs. BU-Q triplet scores on the VG validation set) in the supplementary material, so that readers can assess the degree to which joint training aligns the score scales. This will make the motivation more precise and the evidence more complete.","revision_made":"partial","referee_comment":"Sec. 3, Eq. (1) continued: The gap between Dual-SGG (33.5/38.5 R@K) and normalization fusion (32.2/37.3 R@K) in Tab. 6 is approximately 1.3 points, but this gap could arise from joint feature sharing or gradient flow benefits unrelated to score calibration per se. The paper does not analyze the actual score distributions of TD-Qs vs BU-Qs predictions to verify that joint training aligns their scales. Without such analysis, the specific claim that joint training eliminates the calibration problem identified in the motivation remains unverified. The authors should either (a) provide score distribution analysis showing that joint training does align TD-Q and BU-Q score scales, or (b) soften the motivation to acknowledge that joint training provides multiple benefits (feature sharing, gradient flow, reduced architectural complexity) beyond score calibration."},{"response":"We appreciate this careful observation. To clarify: the Det-T/UDet-T split in Tab. 1 is computed using the pre-trained Deformable-DETR before SGG fine-tuning—that is, the discriminator is the frozen, pre-trained detector, not the fine-tuned version. We will state this explicitly in the revised manuscript. That said, the referee's circularity concern is legitimate: even though the split is computed before fine-tuning, the discriminator is the same architecture (Deformable-DETR with ResNet-50) that is subsequently fine-tuned within the models, so the discriminator is not fully independent of the models being evaluated. The Swin-B experiment (Tab. 5, starred rows) partially addresses this by showing that the complementarity finding holds under a different, stronger detector configuration, but we agree this does not fully eliminate the concern since Swin-B is still a Deformable-DETR variant. We will add a discussion of this limitation in the revised manuscript, acknowledging that the complementarity finding is conditioned on the detector family used and that a fully external discriminator (e.g., a different detector architecture entirely) would provide stronger evidence. We believe the complementarity is a fundamental property of the two reasoning paradigms—detector-based methods are architecturally constrained to triplets whose entities are detected, while query-based methods are not—but we agree that the current experimental design does not fully rule out the influence of the specific detector choice.","revision_made":"partial","referee_comment":"Sec. 1 and Tab. 1: The detector-conditioned reachability split (Det-T vs UDet-T) is defined using an external object detector as a discriminator. The reader's weakest assumption concern is valid: if the external detector's recall characteristics differ significantly from the internal detectors used in the compared models (e.g., the detector-based model in Fig. 1a fine-tunes its own detector), the observed complementarity could partly reflect the specific detector choice rather than a fundamental property of the two reasoning paradigms. The paper does partially address this with the Swin-B experiment (Tab. 5, starred rows), showing robustness under a stronger detector. However, the external detector used for the Det-T/UDet-T split in Tab. 1 is the same Deformable-DETR that is then fine-tuned within the models, which creates a potential circularity: the discriminator is not fully external."},{"response":"As noted in our response above, the Det-T/UDet-T split is computed on the pre-trained detector before SGG fine-tuning. We will add this clarification to Sec. 1 of the revised manuscript. Regarding robustness to the choice of discriminator: the Swin-B experiment (Tab. 5, starred rows) provides partial evidence, as it uses a different backbone and more entity queries, yet the complementarity pattern persists—Dual-SGG w/o BU-Qs maintains high micro-DR but low micro-UDR, while Dual-SGG w/o TD-Qs shows the complementary pattern. However, we acknowledge that both the default and Swin-B configurations are Deformable-DETR variants, so the evidence is limited to one detector family. We will add an explicit discussion of this limitation and note that validating with a structurally different detector architecture (e.g., a Faster R-CNN-based or DINO-based discriminator) would strengthen the generality claim. We consider this an important direction for future work but note that the core architectural argument—that detector-based reasoning is constrained by detection reachability while query-based reasoning is not—holds by construction, independent of the specific detector used to demonstrate it.","revision_made":"partial","referee_comment":"Sec. 1 and Tab. 1 continued: The authors should clarify whether the Det-T/UDet-T split is computed on the pre-trained detector (before SGG fine-tuning) or after fine-tuning, and discuss whether the complementarity finding is robust to the choice of discriminator detector. This is load-bearing because the entire architectural motivation rests on the complementarity claim."}],"tokens_in":17634,"tokens_out":1778,"duration_ms":264824,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"Two things matter here: (1) the detector-conditioned reachability analysis is a genuinely useful diagnostic that explains why detector-based and query-based SGG methods fail differently, and (2) the dual-query architecture that combines both reasoning modes in a single decoder works well in practice, with SOTA results on three benchmarks at 14 FPS with 84.7M parameters — lighter and faster than most competitors that score lower. The pre-ablation in Table 1 is the conceptual core. Splitting ground-truth triplets into Det-T (both subject and object detected by an external detector) and UDet-T (at least one missed) cleanly shows that detector-based models collapse on UDet-T while query-based models recover some of those triplets but lose performance on Det-T. This is a clear, well-controlled experiment — same backbone, encoder, and training protocol across all compared models. The dual-query design is the right architectural response to this finding. TD-Qs condition on detected entity pairs; BU-Qs explore globally from center-biased anchors. The self-attention mask preventing TD-to-BU information leakage is a thoughtful detail, and the ablation in Table 7 confirms it matters. Table 6 is the key evidence that joint training beats post-hoc fusion: Dual-SGG at 33.5/38.5 R@K outperforms normalization fusion at 32.2/37.3, and the gap is consistent across micro-DR and micro-UDR splits. Now the soft spots. The stress-test concern about NMS masking score calibration is partially valid but overstated. The paper's claim is that joint training eliminates the need for post-hoc calibration, and Table 6 supports this — joint training beats all three fusion strategies. However, the paper does not analyze actual score distributions between TD-Qs and BU-Qs predictions, so the specific mechanism (calibration alignment vs. feature sharing vs. gradient flow) is unverified. The claim is empirically supported but mechanistically under-explained. This is a minor gap, not a load-bearing flaw. The reader's concern about the external detector choice is reasonable but addressed by the Swin-B experiment (Table 5), which shows the method remains effective when UDet-T shrinks from 35% to 28%. Code link is a placeholder, which slightly limits reproducibility confidence. Overall: the central argument holds, the experiments are well-controlled, and the results are strong. This is for SGG researchers and anyone building scene understanding pipelines. It deserves a serious referee.","headline":"Dual-SGG combines detector-based and query-based SGG reasoning in one decoder via dual queries; achieves SOTA on VG, OIv6, and GQA-200 with a useful diagnostic framework for understanding complementarity between the two paradigms.","tokens_in":18455,"tokens_out":611,"would_cite":true,"duration_ms":147029,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Two reasoning modes, one decoder: Dual-SGG unifies scene graph generation","keywords":["scene graph generation","detector-conditioned reachability","dual-query design","query-based reasoning","detector-based reasoning","triplet prediction","visual relationship detection","transformer decoder"],"falsifier":"Replace the external object detector used to define the Det-T/UDet-T split with a substantially different detector (different architecture, different miss patterns). If the complementarity between detector-based and query-based models disappears or reverses on the new split, the core empirical finding is detector-specific rather than paradigm-specific.","tokens_in":17471,"feed_emoji":"","tokens_out":1223,"duration_ms":198515,"temperature":0.7,"pith_summary":"Scene graph generation (SGG) — the task of extracting (subject, predicate, object) triplets from images — has been split between two paradigms. Detector-based methods first find objects, then reason over pairs, but cannot predict relationships involving objects the detector missed. Query-based methods learn global triplet queries that bypass the detector entirely, recovering some of those missed relationships, but at the cost of accuracy on the ones the detector does find. This paper introduces the concept of detector-conditioned reachability: a formal way to split ground-truth triplets into those whose subject and object are both detected by an external object detector (Det-T) and those where at least one is missed (UDet-T). Using this split, the authors show empirically that detector-based models dominate on Det-T but nearly collapse on UDet-T, while query-based models recover UDet-T triplets but lose ground on Det-T. This complementarity is the paper's central empirical finding. To exploit it, they propose Dual-SGG, which places two groups of triplet queries in a single transformer decoder: top-down triplet queries (TD-Qs) initialized from detected entity pairs to preserve detector-conditioned reasoning, and bottom-up triplet queries (BU-Qs) initialized from image-center anchors to explore globally. Because both query types are updated jointly in one decoder, their outputs share a common score scale and require no post-hoc calibration. NMS removes duplicates. The method achieves state-of-the-art results on Visual Genome, Open Images v6, and GQA-200 at 14 FPS.","feed_headline":"","feed_subtitle":"","key_machinery":"The detector-conditioned reachability split (Det-T vs UDet-T) using an external object detector as a discriminator; the dual-query design with top-down triplet queries (TD-Qs) initialized from selected entity pairs and bottom-up triplet queries (BU-Qs) initialized from center-biased anchors; a self-attention mask preventing TD-to-BU information leakage; an Entity Pair Selector (EPS) that ranks entity pairs by content and spatial compatibility to limit TD-Q count; and the unified triplet decoder that jointly updates both query types so their outputs share a score scale without post-hoc calibration.","core_discovery":"The paper's central discovery is that the predictive behaviors of detector-based and query-based SGG methods are complementary when analyzed through the lens of detector-conditioned reachability. Detector-based models achieve high recall on triplets whose entities are detected (Det-T) but near-zero recall on triplets with at least one undetected entity (UDet-T). Query-based models partially recover UDet-T triplets but sacrifice Det-T performance. By integrating both reasoning mechanisms in a single decoder via a dual-query design — TD-Qs conditioned on detected entity pairs and BU-Qs exploring globally from center-biased anchors — Dual-SGG preserves the strengths of both paradigms without a ","pith_inferences":["The complementarity may be partially an artifact of using a specific external detector to define the Det-T/UDet-T split. If a different detector with different miss patterns were used, the boundary between the two subsets would shift, potentially changing the magnitude of complementarity observed.","The paper does not report the overlap between what TD-Qs and BU-Qs predict when both succeed; a deeper analysis of whether the two query types converge on the same triplets or genuinely discover different ones would strengthen the complementarity claim.","The 14 FPS speed with 84.7M parameters is competitive, but the BU-Q count (800) is more than double the TD-Q count (300); the sensitivity of performance to this ratio is not explored, leaving open whether a more balanced allocation would change the trade-off."],"forward_implications":["If the complementarity finding generalizes, future SGG improvements should be evaluated on both Det-T and UDet-T subsets rather than aggregate metrics alone, since a model can improve aggregate recall while masking a collapse on one subset.","The dual-query principle — embedding two distinct reasoning priors in one decoder to avoid score-calibration problems — could transfer to other structured prediction tasks where DETR-style queries are used, such as human pose estimation or temporal action detection.","The center-biased initialization for BU-Qs suggests that spatial priors on query anchors matter for global exploration, which could inform anchor design in detection transformers more broadly.","As object detectors improve and UDet-T shrinks, the marginal benefit of BU-Qs should diminish; the paper's Swin-B experiment (UDet-T drops from 35% to 28%) confirms the benefit persists but is reduced, raising the question of when a detector-only approach would suffice."],"fun_headline_variants":["Dual-SGG Combines Detector-Based and Query-Based Scene Graph Generation","Combining Detector-Based and Query-Based Methods for Scene Graph Generation","Dual-Query Design Unifies Detector and Query-Based Scene Graph Generation","Scene Graph Generation Improved by Dual-Query Design","Dual-SGG Leverages Complementary Behaviors in Scene Graph Generation"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The paper assumes that using an external object detector to split triplets into Det-T and UDet-T cleanly isolates the fundamental difference between detector-based and query-based reasoning. If the external detector's miss patterns differ systematically from those of the detectors inside the compared models, the observed complementarity could be partly an artifact of the specific detector chosen rather than a property of the two reasoning paradigms themselves.","fun_headline_variants_meta":{"raw":{"variants":["Dual-SGG Combines Detector-Based and Query-Based Scene Graph Generation","Combining Detector-Based and Query-Based Methods for Scene Graph Generation","Dual-Query Design Unifies Detector and Query-Based Scene Graph Generation","Scene Graph Generation Improved by Dual-Query Design","Dual-SGG Leverages Complementary Behaviors in Scene Graph Generation","Unifying Detector and Query-Based Reasoning for Scene Graph Generation"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":893,"prompt_tokens":404,"completion_tokens":489,"prompt_tokens_details":null},"tokens_in":404,"tokens_out":489,"duration_ms":30983,"temperature":1.0,"reasoning_tokens":450,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T14:07:30.696563+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Replace the external object detector used to define the Det-T/UDet-T split with a substantially different detector (different architecture, different miss patterns). If the complementarity between detector-based and query-based models disappears or reverses on the new split, the core empirical finding is detector-specific rather than paradigm-specific.","supporting_citations":[],"review_version":1}