{"id":"f2ac38fa-fda0-4766-ade7-1d6347123bda","arxiv_id":"2607.27145","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Training-free injection of YOLO-World pairwise geometric predicates plus depth captions raises MLLM spatial F1 (notably +46% relative on BLINK for Qwen2.5-VL) and recovers weak models on VSR.","lead":"ByDeWay-V2 adds explicit left/right/inside spatial predicates from an open-vocabulary detector to depth-layered prompts, improving MLLM spatial answers without training. It targets auditable, low-token spatial judgments for robotics and safety pipelines.","discovery_kind":"extension","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The headline 46% BLINK F1 gain rests on Table 3 — the only table whose F1 values are not the harmonic mean of its own reported precision/recall (Eq. 12), and whose accuracy actually drops for Qwen (0.748→0.727).","rationale":"The reader identified the right general soft spot — unverified reliability of detector-derived predicates (§3.2 thresholds, conf ≥ 0.20, query-seeded vocabulary, no detector-failure ablation) — and I agree that assumption is genuinely load-bearing; there is also a related circularity risk on BLINK, where the pipeline parses the two query nouns, computes left/right from box centers via Eq. 4 (a zero-margin center comparison), injects the predicate, and asks the MLLM to verify it, so the benchmark partially measures heuristic accuracy times model compliance rather than MLLM spatial reasoning (the Discussion's \"the model can exploit it reliably\" concedes as much). However, I judge the more immediate and more checkable concern to be one the reader did not flag: the internal arithmetic inconsistency of Table 3, the very table carrying the strongest claim. Because Tables 1 and 2 are fully consistent with Eq. 12, the Table 3 discrepancies are unlikely to be a rounding artifact and point either to an undisclosed metric protocol or to reporting errors; combined with the masked accuracy regression (0.748→0.727) and the VSR F1 degradation for the flagship model, the \"46% relative F1 gain\" is fragile in a way that detector-throbustness concerns alone do not capture. Credit where due: the paper is unusually transparent about the Qwen VSR precision–recall trade-off, reports accuracy alongside F1 (which is what made the accuracy drop visible), and acknowledges detector-error over-reliance in §4.7. Verdict: keep CONDITIONAL, but attach a specific condition — release of raw predictions/confusion matrices and a consistent recomputation of Table 3. If the recomputation fails to reproduce the reported F1s within a couple of points, the headline claim should be downgraded and the verdict moved toward REJECT of the central quantitative claim (the engineering direction itself could still stand as a systems contribution).","tokens_in":10271,"tokens_out":5642,"duration_ms":222354,"concrete_test":"Recompute Table 3 two ways. (1) From Eq. 12 using the reported P/R: verify Qwen LDP F1 = 0.529 (reported 0.496) and BLIP LDP F1 = 0.377 (reported 0.244). (2) From raw per-sample predictions/confusion matrices on the BLINK spatial subset under one stated protocol (binary true/false scoring, or macro over choice classes), recompute Acc/P/R/F1 for V1 and V2. If the recomputed Qwen V1→V2 relative F1 gain falls materially below 46%, or if Table 3 used macro-F1 while Tables 1–2 used binary F1 (invalidating cross-benchmark comparisons), the headline claim does not hold as reported. Report accuracy alongside F1 in all cases.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim's marquee number — BLINK spatial F1 0.496→0.727 for Qwen2.5-VL, the \"46% relative improvement\" — comes from Table 3, which is the only table in the paper whose F1 column does not satisfy the paper's own Eq. 12 given its precision and recall columns. Checks: Qwen LDP row: 2·0.550·0.510/(0.550+0.510)=0.529, but 0.496 is reported. Qwen V2: 0.737 vs reported 0.727. BLIP LDP: 2·0.505·0.301/0.806=0.377 vs reported 0.244 (a 0.13 gap). BLIP V2: 0.323 vs 0.254. ViLT LDP: 0.454 vs 0.418. In contrast, every row of Tables 1 and 2 satisfies Eq. 12 to three decimals. So either Table 3's F1 was computed under a different, undisclosed protocol (e.g., macro-averaged over answer classes while P/R are binary on a positive class), or the table contains errors. This is load-bearing for three reasons. (1) Table 3 is the sole quantitative support for the abstract's headline claim. (2) BLINK is a multiple-choice benchmark, so binary TP/FP metrics already require an unstated conversion; an inconsistent F1 deepens the protocol opacity. (3) The relative-F1 framing masks regressions: Qwen accuracy on BLINK drops 0.748→0.727 under V2, and on VSR (Table 1) the flagship model's F1 also degrades 0.7835→0.7179 — so for the strongest model, V2's benefit is confined to one benchmark under one metric whose provenance is unclear. Separately, the BLIP VSR \"recovery\" (0.053→0.525) starts from a degenerate operating point (recall 0.028, i.e., the model almost never answers \"true\"), so it largely reflects a shift in answer bias rather than recovered spatial reasoning. None of this requires new experiments to check — only the confusion matrices behind Table 3 and a stated scoring protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript presents ByDeWay-V2, a training-free extension of the authors' earlier Layered-Depth-Based Prompting (LDP) framework. A \"SpatialAnalyser\" module runs YOLO-World-L open-vocabulary detection, computes pairwise geometric relations from axis-aligned bounding boxes (above/below/left/right via centre comparisons; inside/touching via overlap ratios with thresholds θ_in=0.70, θ_touch=0.10; near/far via normalised centre distance with τ_near=0.30), serialises them as plain-English predicates, and concatenates them with LDP depth-layer captions into the MLLM prompt. Evaluation on VSR, the BLINK spatial subset, and a 150-sample curated POPE variant across Qwen2.5-VL-7B, BLIP-Base, and ViLT-B/32 reports: (i) a headline 46% relative F1 gain for Qwen2.5-VL on BLINK (0.496→0.727); (ii) recovery of BLIP-Base on VSR from a near-degenerate F1 of 0.053 to 0.525; (iii) modest POPE gains for all three models; and (iv) regressions for Qwen2.5-VL on VSR (F1 0.7835→0.7179) and for ViLT on VSR and BLINK. The method is simple, model-agnostic, and produces human-readable intermediate evidence, which is genuinely useful for auditability.","tokens_in":10837,"tokens_out":4539,"duration_ms":186668,"significance":"If the numbers hold, the contribution is a practical, reproducible-in-principle recipe: explicit detector-derived spatial predicates injected into prompts, with all thresholds and the full pipeline (Algorithm 1, Eqs. 3–8) disclosed, no training, and a demonstrated operating point under a 40-token budget for ViLT. The explicit LDP vs. LDP+Spatial ablation across three models and three benchmarks is a strength, as is the per-category VSR breakdown (Table 4), which localises gains to the relation types the predicates directly encode. The auditable-predicate framing is well matched to the decision-critical motivation. However, the significance is currently capped by a verification problem: the headline BLINK result (Table 3) is the only table whose F1 column is not the harmonic mean of its own precision/recall columns as defined in Eq. (12), so the paper's marquee number cannot presently be reproduced from the paper's own reported quantities. The honest reporting of the Qwen VSR regression is commendable, but the abstract's framing leans on the one benchmark/metric combination whose provenance is unclear.","major_comments":[{"comment":"Table 3 is internally inconsistent with the paper's own metric definition. Applying Eq. (12) to the reported precision and recall: Qwen LDP gives 2·0.550·0.510/(0.550+0.510)=0.529, not 0.496; Qwen V2 gives 0.737, not 0.727; BLIP LDP gives 0.377, not 0.244 (a 0.13 gap); BLIP V2 gives 0.323, not 0.254; ViLT LDP gives 0.454, not 0.418; ViLT V2 gives 0.340, not 0.318. Every row of Tables 1 and 2 satisfies Eq. (12) to three decimals, so Table 3 alone was computed under some other, undisclosed protocol (e.g., macro-averaged F1 over answer classes while P/R are binary on a positive class). This is load-bearing: Table 3 is the sole quantitative support for the abstract's headline '46% relative F1 improvement'. The authors must either correct the table or fully disclose the alternative computation, including how binary TP/FP/FN are defined on BLINK, which is a multiple-choice benchmark and does n","section":"§4.5, Table 3 (and Eq. 12)"},{"comment":"The text states 'BLIP-Base and ViLT also record meaningful gains in F1 and accuracy' on BLINK. The table contradicts this: ViLT F1 falls 0.418→0.318 (a substantial regression) and BLIP F1 moves only 0.244→0.254; BLIP accuracy 0.483→0.490 is within noise for a small subset. The sentence must be rewritten to match the data; as written it misrepresents two of the three model rows in the paper's headline table.","section":"§4.5, BLINK paragraph vs. Table 3"},{"comment":"For the flagship model (Qwen2.5-VL), V2 strictly regresses on VSR: F1 0.7835→0.7179, recall 0.7155→0.6130, accuracy 0.7872→0.7408. The paper's counterweight is a precision 'gain' of 0.8658→0.8660 (+0.0002), which is numerically indistinguishable from zero and cannot support the claim that the model is 'strictly more often correct'. Combined with the BLINK accuracy drop (0.748→0.727), the accurate summary for the strongest model is: V2 helps on one benchmark under one metric (whose computation is currently unexplained, see comment on Table 3) and hurts elsewhere. The abstract and §4.7 should state this trade-off plainly rather than framing a +0.0002 precision delta as a desirable property for high-stakes settings.","section":"§4.5, VSR paragraph / Table 1; Abstract"},{"comment":"The entire method rests on the correctness of the injected predicates, yet predicate quality is never measured. The heuristics are fragile in known ways: axis-aligned box centres give wrong left/right labels under perspective and for elongated or occluded objects; the priority order containment > contact > vertical/horizontal > proximity forces exactly one predicate per ordered pair even when several hold; and the detector vocabulary is seeded from the query nouns at conf≥0.20, so missed detections silently remove the relevant evidence. §4.7 acknowledges detector errors anecdotally ('earlier experiments') but reports no failure analysis. A small audit — e.g., predicate accuracy against VSR ground-truth relations on a few hundred samples, and performance stratified by whether the queried pair was actually detected — would directly test the paper's central assumption and is within the manu","section":"§3.2 and §4.1"}],"minor_comments":[{"comment":"POPE evaluation uses a 150-sample curated subset hosted by an author (ref. 14) rather than the standard POPE splits. With n=150, the reported accuracy differences (e.g., 86.0%→90.7%) have wide confidence intervals; please report CIs or evaluate on the full benchmark, and state the curation criteria for the subset.","section":"§4.2, POPE"},{"comment":"No variance or significance reporting anywhere; all numbers appear to be single runs. For the large VSR test set this is less concerning, but the BLINK spatial subset is small, and the Qwen accuracy drop there (0.748→0.727) may not be significant. Please give subset sizes and, where feasible, bootstrap CIs.","section":"§4.5 / §4.6"},{"comment":"§3.2 says all N(N−1) ordered-pair sentences are concatenated into C_Spatial, while §4.1 caps at 20 spatial relation pairs per image. Please specify the selection/pruning rule when N(N−1) exceeds the cap, since it affects which evidence the model sees.","section":"§3.2 / §4.1"},{"comment":"The 'Vision-First Advisory' instruction is credited with the Qwen VSR regression but is never shown verbatim; only Figure 3 paraphrases it ('Trust your visual analysis FIRST'). Include the exact prompt text (appendix is fine), since the ablation LDP vs. LDP+Spatial is confounded if the instruction wording also changed between V1 and V2 — please confirm the instruction was identical across both configurations.","section":"§4.5"},{"comment":"No code or prompt release is mentioned. Given that the contribution is a prompt-construction pipeline with several free thresholds (θ_in, θ_touch, τ_near, conf≥0.20, max 20 pairs, T1/T2 percentiles), releasing the serialisation code and exact prompts would substantially raise reproducibility.","section":"General"},{"comment":"The abstract's 'operates under a strict 40-token context budget on CPU' conflates two properties: the 40-token limit applies to ViLT (a 40-token text encoder limit, not a context budget in the usual LLM sense), and CPU inference is mentioned only for ViLT in §4.1. Please clarify which model/configuration supports the real-time CPU claim and report latency.","section":"§4.3 / Abstract"},{"comment":"All seven VSR categories show near-uniform ΔF1 (+0.39 to +0.57) including 'Unallocated', which by definition lacks a clear relation type. A brief comment on why unallocated relations benefit as much as projective ones would help rule out a generic answer-distribution shift (e.g., V2 simply making BLIP answer 'true' more often) rather than relation-specific grounding.","section":"§4.6, Table 4"}],"recommendation":"major_revision","confidential_remarks":"The Table 3 / Eq. (12) inconsistency is verifiable arithmetic, not a matter of interpretation, and it sits on the paper's single headline number; I would want it resolved before any further consideration. Secondary points for the editor's awareness: the POPE evaluation uses a 150-sample subset curated and hosted by one of the authors (ref. 14), and the paper is an incremental extension of the same group's ByDeWay V1 — both are disclosed in the text, but they reduce the independence of the hallucination evidence. The V1 self-citation pattern is otherwise unremarkable. Scope fit for a venue covering efficient/explainable multimodal inference is good if the evaluation issues are fixed."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a clear systems extension of ByDeWay/LDP. They run YOLO-World, turn axis-aligned boxes into English predicates (left/right/above/inside/touching/near), and concatenate those with the old depth-layer captions. Fully training-free, model-agnostic, and they keep a 40-token spatial-only path for ViLT/CPU. That is the real product.\n\nWhat is new is the hybrid serialisation and the three-model ablation on VSR, BLINK spatial, and a small POPE set—not the detector, not monocular depth, not geometric heuristics. Credit where due: the write-up is readable, the LDP vs LDP+Spatial contrast is clean, the predicates are human-inspectable (useful for the decision-support framing they advertise), and weak models move a lot. BLIP on VSR goes from near-dead (F1 0.05, recall ~0.03) to 0.53; POPE edges up; per-category BLIP gains concentrate on projective/proximity relations where box geometry should help.\n\nSoft spots in proportion. The abstract's 46% relative F1 on BLINK for Qwen rests on Table 3, and those F1 values are not the harmonic mean of the reported P and R under the paper's own Eq. 12, while Tables 1–2 check out to three decimals. BLINK is also multi-choice, so the binary conversion is underspecified. That is the one thing that must be fixed or explained before anyone cites the headline number. Separately, for Qwen the story is mixed: VSR F1 falls (0.78→0.72), BLINK accuracy falls (0.75→0.73) even as reported F1 rises; the Vision-First instruction is an honest choice but it means the flagship model is not uniformly helped. The BLIP “recovery” partly corrects a degenerate always-false bias. Thresholds (θ_in, θ_touch, τ_near, conf) are hand-set; no detector-miss/occlusion ablation; POPE is a 150-sample curated slice; no code. None of that sinks a systems paper, but they are real limits.\n\nWho it is for: people wiring prompt wrappers into robotics or safety MLLM pipelines who care about token budget and auditable intermediate evidence. Not a foundational spatial-reasoning result. Math is elementary geometry; citations are appropriate; data are public benchmarks plus one small author-hosted POPE slice.\n\nI would send it to peer review. Ask for a scoring protocol (or corrected Table 3), confusion matrices, and a short detector-error analysis. Worth a referee's time as scoped engineering; not worth treating the 46% figure as settled until the table is clean.","headline":"Sensible training-free hybrid of LDP depth captions and YOLO box predicates; useful engineering, but the marquee BLINK F1 numbers do not match the paper's own F1 formula and gains are mixed for the strongest model.","tokens_in":12139,"tokens_out":699,"would_cite":false,"duration_ms":26956,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Injecting detector-derived spatial predicates into prompts, with depth captions, fixes multimodal models’ fine-grained spatial reasoning without any training.","keywords":["Multimodal Large Language Models","Spatial Reasoning","Explainable AI","Decision Support Systems","Hallucination Reduction","Resource-Efficient Inference","Layered-Depth-Based Prompting","Open-Vocabulary Detection"],"falsifier":"Re-run BLINK spatial and VSR after systematically dropping, swapping, or corrupting the injected predicates (or replacing the detector with a much weaker one) and check whether the reported F1 gains over depth-only prompting vanish; if gains survive when the predicates are wrong, the geometric scaffold is not what is carrying the improvement.","tokens_in":11734,"feed_emoji":"📐","tokens_out":921,"duration_ms":36078,"temperature":0.7,"pith_summary":"Multimodal language models are moving into robotics, embodied agents, and safety monitoring, where a wrong left/right or inside judgment can matter and where black-box answers are hard to audit. Depth-layer prompting already helps with near-versus-far structure, but it does not tell the model how objects in the same depth band sit relative to one another. This paper shows that an open-vocabulary detector’s bounding boxes can be turned into plain-English pairwise predicates—left of, above, inside, touching, near—and injected beside the depth captions as secondary, human-readable evidence. Across spatial and hallucination benchmarks the hybrid prompt raises F1 sharply for both strong and compact models, and a predicate-only stream still fits a 40-token CPU budget. The result is a training-free scaffold that improves answers and leaves an inspectable trail for decision support.","feed_headline":"Spatial predicates lift MLLM F1 46% with no training","feed_subtitle":"Pairwise box relations plus depth captions fix left/right and inside queries under a 40-token budget.","key_machinery":"The SpatialAnalyser: YOLO-World detections supply axis-aligned boxes; pairwise heuristics (centre comparisons for left/right/above/below, overlap ratios for inside/touching, normalised centre distance for near/far) are serialised as readable predicates and concatenated with LDP depth-layer captions into one structured prompt.","core_discovery":"ByDeWay-V2 establishes that pairing layered monocular-depth captions with explicit pairwise geometric predicates computed from open-vocabulary bounding boxes, then serialising both into the MLLM prompt, recovers fine-grained spatial reasoning and reduces object hallucination in a fully training-free way. On the BLINK spatial subset the method delivers a 46 percent relative F1 gain for Qwen2.5-VL over depth-only prompting; on VSR it lifts BLIP-Base from near-random F1 0.05 to 0.53; POPE metrics also improve. The same light predicate stream remains usable under a strict 40-token context limit.","pith_inferences":["Confidence-weighted or ensemble predicate injection could ease the precision–recall trade-off that appears when a strong model overrides correct context.","The same geometric serialisation could ground multi-step embodied instructions (move left of X, then above Y) without retraining planners.","Heavy occlusion and non-axis-aligned objects remain an open stress test before the method is trusted in real safety pipelines."],"forward_implications":["Operators can treat the injected predicates as auditable intermediate evidence rather than accepting a black-box spatial verdict.","Compact multimodal models that previously failed spatial VQA become competitive without fine-tuning or extra parameters.","Token-constrained deployments can drop verbose depth captions and keep only the compact spatial predicates.","Object-existence hallucination rates fall when spatial grounding is added on top of depth context alone.","The pipeline is model-agnostic and can sit in front of any MLLM that accepts image plus text context."],"fun_headline_variants":["ByDeWay-V2 adds pairwise spatial predicates, lifts F1 46% training-free","Depth plus box relations fix MLLM left/right and inside under 40 tokens","Explicit geometric predicates recover BLIP spatial F1 from 0.05 to 0.53","Open-vocab box relations plus depth cut MLLM hallucination on VSR and BLINK","Training-free spatial predicates give Qwen2.5-VL 46% relative F1 gain"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The method assumes that 2D box comparisons from a query-seeded detector are accurate and complete enough to serve as trustworthy secondary evidence for the model’s spatial answers.","fun_headline_variants_meta":{"raw":{"variants":["ByDeWay-V2 adds pairwise spatial predicates, lifts F1 46% training-free","Depth plus box relations fix MLLM left/right and inside under 40 tokens","Explicit geometric predicates recover BLIP spatial F1 from 0.05 to 0.53","Open-vocab box relations plus depth cut MLLM hallucination on VSR and BLINK","Training-free spatial predicates give Qwen2.5-VL 46% relative F1 gain"]},"model":"grok-4.5","effort":"low","cost_usd":0.004094,"raw_usage":{"total_tokens":1356,"prompt_tokens":956,"num_sources_used":0,"completion_tokens":101,"cost_in_usd_ticks":40944000,"prompt_tokens_details":{"text_tokens":956,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":299,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":956,"tokens_out":101,"duration_ms":5558,"temperature":1.0,"reasoning_tokens":299,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T12:12:31.339810+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run BLINK spatial and VSR after systematically dropping, swapping, or corrupting the injected predicates (or replacing the detector with a much weaker one) and check whether the reported F1 gains over depth-only prompting vanish; if gains survive when the predicates are wrong, the geometric scaffold is not what is carrying the improvement.","supporting_citations":[],"review_version":1}