{"id":"e352e45f-6247-4193-8b31-1d3c3ae991dc","arxiv_id":"2608.07051","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A constraint-planning framework places PEFT adapters on YOLO detectors, beating full fine-tuning on YOLO11s/YOLO12s and refusing RT-DETR-L before training.","lead":"YOLO-PEFT treats adapter placement on real-time detectors as a constraint-planning problem, filtering unsafe modules and returning a budgeted plan or a refusal before training. On VOC07+12, planner-placed LoRA matches or beats full fine-tuning on YOLO11s and YOLO12s while refusing RT-DETR-L as catastrophic.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Refuse decision is calibrated and evaluated on the same seven RT-DETR-L collapses, so it does not yet demonstrate pre-training prediction for a new architecture-adapter combination.","rationale":"The most defensible reading of the paper is a scoped systems claim: within the evaluated families, planner-selected placements improve over Full-SFT on YOLO11s/YOLO12s under a shared protocol, and the Refuse rule summarizes observed RT-DETR-L instability rather than predicting unseen-family safety. The paper is unusually explicit about this boundary in Sec. 4.2, the Eq. (5) discussion, and the conclusion. I considered the VOC2007 test-set checkpoint-selection protocol as an alternative concern; it affects absolute numbers but is shared across all methods and is disclosed, so it is secondary. I also considered the HRA-versus-RS-LoRA discrepancy; it weakens the phrase 'planner-selected RS-LoRA' but does not threaten the planner-placement comparison because HRA is evaluated under the same planner placement. The single load-bearing gap is the absence of any held-out detector family for the Refuse calibration, and the reader's weakest_assumption identifies exactly this issue. Since the reader already conditions the verdict on this gap, the CONDITIONAL verdict remains appropriate; the proposed leave-one-family-out test would determine whether the concern actually lands.","tokens_in":21431,"tokens_out":6669,"duration_ms":63275,"concrete_test":"Hold out RT-DETR-L entirely when fitting Eq. (5): fit beta_0..beta_4 and xi_p on the diagnostic entries from YOLOv8n, YOLO11n, YOLO12n, and YOLO-World-s only; then evaluate the fitted rule on the seven RT-DETR-L LoRA-family configurations and report whether all seven are refused and what the false-refusal/false-acceptance counts are. Repeat this for each family as the held-out family to build a leave-one-family-out confusion matrix. If the rule refuses all held-out catastrophic configurations and accepts held-out safe ones, the Refuse mechanism is predictive; if it fails on a held-out family, the current Refuse evidence is descriptive rather than pre-training predictive.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition for the 'Refuse before training' contribution is that the Eq. (5) calibration transfers beyond the exact matrix on which it was fit. The paper fits beta_0..beta_4 and variant coefficients xi_p on the diagnostic sweep, manually sets the catastrophe threshold at delta = -0.05, and then reports that all seven RT-DETR-L LoRA-family configurations fall below this threshold. Because RT-DETR-L is one of the families used to fit the calibration and because the seven evaluated configurations are the same data points that define the collapse pattern, the Refuse decision is currently a post-hoc classification of those seven runs rather than a before-training prediction for an unseen family. The paper states this limitation explicitly ('refusal on unseen detector architectures remains an open validation problem'), but the abstract's 'calibrated Refuse-to-Full-SFT decision within the evaluated coverage' and the reader's strongest claim ('can detect high-risk architecture-adapter combinations before training') rely on the untested assumption that the linear model extrapolates at least to new variants on the same architecture. This assumption is load-bearing for the safety/refusal contribution, although not for the YOLO11s/YOLO12s placement gains, which rest on a shared protocol and disclosed test-set checkpoint selection.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes YOLO-PEFT, a structure-aware constraint-planning framework for adapter placement on real-time detectors. Given a detector graph, a PEFT request, and a resource budget, the framework parses operator and semantic roles, applies hard constraints (operator validity, detection-head semantics, graph-interface, deployment), allocates ranks under a budget, and either emits a placement plan or returns a Refuse decision with a Full-SFT fallback. The main empirical claim is that under the official VOC07+12 trainval-to-VOC07 test protocol, planner-selected RS-LoRA reaches 0.7138 and 0.7307 mAP50-95 on YOLO11s and YOLO12s, versus 0.6428 and 0.6662 for Full-SFT, while RT-DETR-L returns Refuse because all seven evaluated LoRA-family configurations cross a predefined catastrophe threshold of Δ = −0.05. The paper also reports a controlled efficiency audit (LoRA reduces peak VRAM by 43.9% but training takes 1.72× longer), a merge-equivalence proof for a fallback grouped-convolution LoRA backend, and an scoped MoE stress test. The authors repeatedly and explicitly state that the Refuse rule is calibrated only within the evaluated families and that refusal on unseen detector architectures remains an open validation problem.","tokens_in":21648,"tokens_out":7815,"duration_ms":68669,"significance":"If the results hold, the paper makes a useful contribution by recasting PEFT placement as an auditable constraint-satisfaction problem rather than module-name matching, and by demonstrating on two YOLO variants that graph-aware placement can substantially beat full fine-tuning on a mid-scale detection benchmark. The paper is unusually transparent about its limitations: it discloses test-set involvement in checkpoint selection, reports a controlled memory/time trade-off rather than a universal speedup, and scopes the Refuse mechanism to calibrated coverage. It also ships a deployment contract with ONNX/TensorRT checks and a merge-equivalence proof for the fallback backend. The main weakness is that the Refuse-before-training capability is not yet validated as a predictive tool for unseen architectures; current evidence is a within-sample classification of observed collapses on the evaluated families.","major_comments":[{"comment":"The reliability calibration in Eq. (5) is fitted and assessed on the same seven RT-DETR-L runs, and the paper itself states that this 'is not a held-out-architecture test.' The variant-level LOVO (86.7% accuracy, F1 = 0.850) holds out variants, not detector families, so it does not demonstrate that the Refuse rule can predict a collapse for an unseen architecture-adapter combination. Since the framework's advertised contribution is 'Refuse before training,' the authors should either add a held-out-family validation (e.g., calibrate on four of the five families and test the refusal decision on the fifth) or explicitly reframe the contribution as a within-coverage post-hoc risk assessment and remove any implication of pre-training prediction for new architectures.","section":"Sec. 4.2, Eq. (5)"},{"comment":"The primary mAP comparisons are obtained by selecting the best checkpoint on the VOC2007 test set for every method. The paper discloses this and calls the results 'controlled relative comparisons,' but model selection on the test set can inflate absolute scores and potentially alter rankings in ways not captured by the shared protocol. Because the central claim is that planner-selected PEFT outperforms Full-SFT, the authors should report at least one validation-based selection (for example, on a held-out subset of the training set or on VOC2012 val) to confirm that the superiority holds when the test set is not used for checkpoint selection.","section":"Sec. 4.1 and Appendix H"},{"comment":"In the YOLO-Master-EsMoE-S stress test, the Full-SFT anchor is taken from a run whose W&B state is logged as 'crashed.' Using metrics from a crashed run as a baseline anchor is not reliable, even if the row is secondary and marked with a dagger. This row should be re-run to completion or omitted from the stress-test table; as presented, the 'HRA beats Full-SFT' observation in that table rests on a possibly corrupted anchor.","section":"Sec. 4.3, Table 4"},{"comment":"The catastrophe threshold Δ = −0.05 is described as 'predefined' but no justification or sensitivity analysis is provided. Since the RT-DETR-L Refuse decision is driven entirely by this threshold, a short analysis showing that the decision is robust to reasonable variations (for example, Δ = −0.03 or −0.08) would materially strengthen the claim that the rule is not an artifact of a single manually chosen constant.","section":"Sec. 3.2, Eq. (5)"}],"minor_comments":[{"comment":"The utility function u(i, π(i); p, ψ) is not fully specified in the main text; the rule-only decomposition u = u_op + u_sem + u_range + u_rank − λc_p is given only in words. Please provide explicit definitions or a more detailed supplementary reference in the main text.","section":"Sec. 3.2, Eq. (4)"},{"comment":"The heatmap in Fig. 2 clips colors below −0.30, but several entries are −0.600 (the clipped failure value). The caption and table should state this clipping explicitly so readers do not interpret the clipped cells as equal to −0.30.","section":"Supplementary Tab. S2"},{"comment":"The explanation that the W&B field 'val2007' is an alias for the VOC2007 test set is important and easily overlooked; consider moving it from Appendix H into the main experimental setup for greater prominence.","section":"Sec. 4.1"},{"comment":"In the top pipeline, the 'Feasible? No' branch leads to 'REFUSE' and then to 'Full-SFT fallback'; consider labeling this fallback as a valid, auditable outcome in the figure, matching the paper's framing that Refuse is a first-class planning result.","section":"Fig. 1"},{"comment":"The LOVO accuracy of 86.7% and F1 of 0.850 are reported without a description of the exact folds (e.g., number of variants, class balance, or variance across folds). A one-sentence fold description would improve reproducibility.","section":"Sec. 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is methodologically more honest than most: it discloses test-set checkpoint selection, scopes the Refuse claim to calibrated coverage, and provides a controlled efficiency audit rather than a universal speedup claim. The main risk is that the Refuse contribution is advertised as 'before training' while the evidence is a within-sample classification of the evaluated families; the authors should either add a held-out-family experiment or soften the claim. The MoE stress test's crashed Full-SFT anchor is a secondary but concrete data-quality issue that should be fixed. If the authors address these points, the paper could be a solid systems-oriented contribution to the PEFT-for-detection literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuine systems contribution, not another adapter variant. The constraint-planning formulation with role-typed graphs, reason codes, and Refuse-as-outcome is new in the detector-PEFT literature, and the paper is unusually honest about what it does not show. The main thing to know: the Refuse mechanism is calibrated on the same seven RT-DETR-L collapses it then evaluates, so it is a post-hoc description, not a before-training prediction for a new architecture. The body text says this explicitly, but the abstract's phrasing will mislead casual readers.\n\nWhat it does well: the empirical matrix of 14 variants across five detector families is the broadest I have seen in this space, the VOC07+12-to-test protocol is shared and the split audit shows zero train/eval ID overlap, the merge-equivalence proof for the fallback conv-LoRA backend is formal and checkable, and the controlled VRAM/time audit cleanly separates memory reduction from wall-clock cost. The train-save-merge-export contract closes a real gap that most PEFT papers ignore.\n\nSoft spots, in proportion. The Refuse calibration is the main one: Eq. (5) coefficients and the delta = -0.05 threshold come from the same diagnostic matrix used to report the refusal, and no detector family is held out. That is a genuine limitation for the safety claim, and the authors disclose it. I would ask them to either hold out one family during calibration or to reframe 'Refuse before training' as 'Refuse within calibrated coverage' in the abstract and contribution claims. As written, a reader can walk away thinking the planner predicts failure for a new detector; it cannot yet.\n\nThe VOC2007 test set participates in best-checkpoint selection; the authors disclose this and correctly label the results as controlled relative comparisons. Acceptable for their claim, but the absolute mAP numbers are not blind evaluations. The seed-0 main matrix with selected reruns (SD <= 0.006, rankings unchanged) is also disclosed; a full three-seed matrix would be better but this is minor.\n\nOne wording issue: the abstract says 'planner-selected RS-LoRA', but HRA beats RS-LoRA on both YOLO11s and YOLO12s in the variant sweep. If the planner can choose variants, it should have chosen HRA; if RS-LoRA is the default, say so. Cosmetic, but confusing for readers.\n\nWho this is for: engineers fine-tuning YOLO-family detectors and researchers working on adapter placement. It deserves a serious referee. I would accept it for review with a request to fix the Refuse validation framing and clarify variant selection. The core placement results are plausible and the reporting is more transparent than most papers in this subfield.","headline":"A real engineering contribution with an honest limitation: the Refuse mechanism is a calibrated description of seven collapses, not yet a prediction for unseen architectures; the placement results are plausible and worth referee time.","tokens_in":22239,"tokens_out":3249,"would_cite":true,"duration_ms":28349,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T45","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that adapter placement for PEFT on real-time detectors should be solved by constraint planning on the detector graph, and that the resulting plans match or exceed full fine-tuning on YOLO11s and YOLO12s while returning…","keywords":["parameter-efficient fine-tuning","YOLO detection","adapter placement","constraint planning","low-rank adaptation","VOC fine-tuning","Refuse decision","object detection"],"falsifier":"Run a full VOC fine-tuning on a held-out detector family (for example a mixture-of-transformers detector) with planner-selected RS-LoRA: if an accepted plan produces $\\Delta\\mathrm{mAP}$ below $-0.05$, or a refused configuration would have matched or exceeded Full-SFT, then the calibration's predictive claim fails. The paper's own absence of a held-out family makes this test the decisive one.","tokens_in":2022,"feed_emoji":"🎯","tokens_out":3509,"duration_ms":76807,"temperature":0.7,"pith_summary":"Parameter-efficient fine-tuning (PEFT) methods transferred from language models select adapter targets by module name or type and can silently degrade real-time object detectors, whose graphs mix grouped and depthwise convolutions, loss-coupled DFL projections, attention blocks, text fusion, and MoE routing. The paper proposes YOLO-PEFT, which treats adapter placement as a constraint-planning problem: it parses the detector graph into operator and semantic roles, evaluates explicit operator-validity, detection-semantic, graph-interface, deployment, and budget predicates with reason codes, and either emits a budgeted target-module plan or returns Refuse before training. Under the official VOC07+12 trainval-to-VOC07 test protocol, planner-selected RS-LoRA reaches 0.7138 and 0.7307 mAP50-95 on YOLO11s and YOLO12s, versus 0.6428 and 0.6662 for Full-SFT; on RT-DETR-L, all seven evaluated LoRA-family configurations cross the catastrophic threshold $\\Delta = -0.05$, so the planner refuses and falls back to Full-SFT. A controlled YOLO11 audit shows LoRA reduces peak training VRAM by 43.9 percent but takes 1.72 times longer to train.","feed_headline":"Planner-selected LoRA beats full fine-tuning on YOLO11/12","feed_subtitle":"Structure-aware placement gains +7.1 and +6.5 mAP on VOC and refuses RT-DETR before training.","key_machinery":"The load-bearing object is the role-aware typed graph $G=(V,E)$ of the detector: each module $i$ carries operator metadata $o_i$, semantic role $s_i$, graph-interface metadata $q_i$, and deployment metadata $h_i$. The planner evaluates predicates such as dense conv supported by the backend, grouped conv requires $G \\mid r$, depthwise and normalization excluded, DFL bins and MoE routers frozen, and shape, arity, and residual-preserving substitution, then solves budget-aware rank assignment $\\max_{\\pi} \\sum_i u(i,\\pi(i))$ subject to $\\sum_i c_p(i,\\pi(i)) \\le B$. Reliability is estimated by a linear calibration $\\Delta \\mathrm{mAP} \\approx \\beta_0 + \\beta_1 \\varphi_{\\mathrm{attn}} + \\beta_2 \\varphi_{\\mathrm{text}} + \\beta_3 \\varphi_{\\mathrm{dw}} + \\beta_4 \\xi_p$, where the fingerprint includes attention, text-fusion, MoE, depthwise, and dense, grouped, or convolution fractions; a prediction below $\\Delta = -0.05$ triggers Refuse. The runtime contract preserves train-save-merge-export invariants, with merge equivalence for the fallback grouped-convolution LoRA backend proven as $\\mathrm{conv}(W_0,x) + s\\, g((A_g,B_g)_g,x) = \\mathrm{conv}(W_0 + s\\Delta W, x)$ up to floating-point tolerance.","core_discovery":"The central claim, stated for a fair reader, is that the reason PEFT often fails on real-time detectors is structural misplacement, not the low-rank parameterization itself. YOLO-PEFT formulates placement as an auditable multi-constraint decision problem on a directed acyclic graph of the detector, where every module carries operator metadata, a semantic role, graph-interface metadata, and deployment metadata. The planner applies explicit constraint classes in a fixed order: operator validity, detection-semantic safety, graph-interface compatibility, architecture-conditioned policies, budget feasibility, deployment compatibility, and reliability calibration, and returns either an accepted plan with a rank assignment or Refuse with a terminal reason code. Within the evaluated VOC protocol and detector families, planner-selected RS-LoRA matches or exceeds full fine-tuning on YOLO11s and YOLO12s, while RT-DETR-L produces a calibrated Refuse because every one of the seven swept LoRA-family configurations falls below $\\Delta = -0.05$; the paper explicitly scopes these results to the evaluated families and calibration coverage.","pith_inferences":["If the planner is applied to a detector family outside the five calibrated ones, the Refuse decision should be re-derived from that family's own collapses; the paper's own admission that held-out-architecture validation remains open suggests the core value is the auditable decision log rather than the refusal threshold itself.","The same graph-role filters could block unsafe placements for cross-modal and mixture-of-transformers detectors before any training, because the predicates are structural rather than task-specific.","A decisive test of the calibration's predictive power would hold out an entire detector family at calibration time and measure false-refusal and false-acceptance rates on that held-out family."],"forward_implications":["Planner-selected PEFT can replace full fine-tuning for YOLO11s and YOLO12s on the VOC protocol, gaining +7.1 and +6.5 mAP50-95 over Full-SFT while freezing most of the detector weights.","Refuse is a legitimate planner outcome: on RT-DETR-L it avoids seven training runs that would each collapse below $\\Delta = -0.05$, and Full-SFT becomes the fallback rather than a failure.","Adapter training trades memory for time: on the controlled YOLO11 audit, LoRA cuts peak VRAM by 43.9 percent but runs 1.72 times longer, so efficiency claims must separate VRAM from wall-clock.","Accepted plans survive the full deployment contract: adapters can be saved, reloaded, merged, and exported through ONNX and TensorRT, with merged models recovering base operator cost."],"supporting_citations":[{"why":"Supplies the low-rank adapter parameterization that the planner places.","marker":"[13]"},{"why":"Defines the official PASCAL VOC dataset and the trainval-to-test protocol used for the head-to-head numbers.","marker":"[5]"},{"why":"The generic module-name and module-type PEFT interface that YOLO-PEFT replaces with constraint checks.","marker":"[26]"},{"why":"One of the heterogeneous detector families studied, contributing text-fusion branches that need separate placement policies.","marker":"[4]"},{"why":"The Transformer-decoder detector whose seven LoRA-family configurations all collapse and trigger Refuse.","marker":"[25]"},{"why":"The attention-centric YOLOv12 family whose Area-Attention blocks motivate architecture-conditioned guard rules.","marker":"[33]"},{"why":"The mixture-of-experts detector used for the scoped stress test with planner and expert targeting disabled.","marker":"[22]"},{"why":"A detector-specific PEFT method that shows selective adaptation but lacks a common abstraction across detector families.","marker":"[47]"},{"why":"Defines Distribution Focal Loss whose fixed bin projections the planner freezes as a safety constraint.","marker":"[20]"}],"fun_headline_variants":["Planner-selected LoRA adds +7.1 mAP on YOLO11, +6.5 on YOLO12","Structure-aware PEFT: YOLO11/12 gains, RT-DETR refused","Why PEFT fails on detectors? Misplacement, not low-rank","YOLO-PEFT: auditable adapter planning beats full fine-tune"],"cache_read_input_tokens":24320,"weakest_assumption_plain":"The load-bearing premise is that the reliability calibration, with coefficients fitted on the same five detector families used for evaluation, can make pre-training Refuse decisions meaningful within claimed calibrated coverage; because no family is held out, the RT-DETR-L refusal is validated on data that already determined that refusal.","fun_headline_variants_meta":{"raw":{"variants":["Planner-selected LoRA adds +7.1 mAP on YOLO11, +6.5 on YOLO12","Structure-aware PEFT: YOLO11/12 gains, RT-DETR refused","Why PEFT fails on detectors? Misplacement, not low-rank","YOLO-PEFT: auditable adapter planning beats full fine-tune"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00031,"raw_usage":{"total_tokens":1844,"prompt_tokens":1094,"completion_tokens":750,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":710,"completion_tokens_details":{"reasoning_tokens":653}},"tokens_in":710,"tokens_out":750,"duration_ms":6095,"temperature":1.0,"reasoning_tokens":653,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:42:13.639490+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a full VOC fine-tuning on a held-out detector family (for example a mixture-of-transformers detector) with planner-selected RS-LoRA: if an accepted plan produces $\\Delta\\mathrm{mAP}$ below $-0.05$, or a refused configuration would have matched or exceeded Full-SFT, then the calibration's predictive claim fails. The paper's own absence of a held-out family makes this test the decisive one.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the official PASCAL VOC dataset and the trainval-to-test protocol used for the head-to-head numbers."},{"cited_title":"PEFT: State-of-the-art parameter-efficient fine-tuning methods.https://github.com/huggingface/peft, 2022","cited_arxiv_id":null,"evidence_quote":"The generic module-name and module-type PEFT interface that YOLO-PEFT replaces with constraint checks."},{"cited_title":"YOLO-World: Real-time open-vocabulary object detection","cited_arxiv_id":null,"evidence_quote":"One of the heterogeneous detector families studied, contributing text-fusion branches that need separate placement policies."},{"cited_title":"RT-DETR: DETRs beat YOLOs on real-time object detection","cited_arxiv_id":null,"evidence_quote":"The Transformer-decoder detector whose seven LoRA-family configurations all collapse and trigger Refuse."},{"cited_title":"YOLOv12: Attention-centric real-time object detectors","cited_arxiv_id":null,"evidence_quote":"The attention-centric YOLOv12 family whose Area-Attention blocks motivate architecture-conditioned guard rules."},{"cited_title":"YOLO-Master: MOE-accelerated real-time detection,","cited_arxiv_id":null,"evidence_quote":"The mixture-of-experts detector used for the scoped stress test with planner and expert targeting disabled."},{"cited_title":"YOLO- IOD: Towards real time incremental object detection","cited_arxiv_id":null,"evidence_quote":"A detector-specific PEFT method that shows selective adaptation but lacks a common abstraction across detector families."},{"cited_title":"Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection","cited_arxiv_id":null,"evidence_quote":"Defines Distribution Focal Loss whose fixed bin projections the planner freezes as a safety constraint."}],"review_version":1}