{"id":"b8f26617-578a-4118-a18a-05a9bf37b1e3","arxiv_id":"2601.05680","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"AGDC is an autoregressive model that combines categorical type prediction with diffusion-based continuous coordinate generation and length regularization, plus a new 334K-sample semiconductor-layout benchmark.","lead":"AGDC generates drawings, layouts, and chip designs by predicting each element's type from a menu and its coordinates with a diffusion model, avoiding the precision limits of tokenized coordinates. Its value is in testing whether continuous, variable-length autoregressive generation can replace discretization in high-precision design domains.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Self-reported appendix data contradict the central claim: LT at 8 bits beats AGDC on FID on PubLayNet and Rico, so 'higher-fidelity than discretization baselines' is unsupported as stated.","rationale":"Good-faith reading: the paper's contribution is a hybrid discrete/continuous autoregressive framework with an EOS logit adjustment and length regularization, plus a new ContLayNet benchmark. The theoretical precision analysis (Section 3.1) is coherent, and the ContLayNet DRC results are internally favorable. However, the central empirical claim is broader: the abstract asserts higher-fidelity generation than discretization-based and fixed-schema baselines across domains. The supporting main tables compare LayoutTransformer only at 18 bits and compare IconShop only at 8/9 bits while reporting AGDC's score as superior/comparable. The appendix's own precision sweep shows LT is not monotonic in precision: it peaks at 8 bits, and at 8 bits it outperforms AGDC on FID and alignment for both PubLayNet and Rico. That is not an outsider's competing result or a disagreement with consensus; it is a direct internal inconsistency between Table B and Table 3. The text-to-SVG table likewise shows IconShop's 8-bit FID is lower than AGDC's. So the unqualified headline claim is unsupported. The reader's weakest assumption is essentially the same concern: the main comparison selects the worst-performing feasible precision for LT instead of the best-performing precision. I agree with that diagnosis, with one minor correction: on Rico overlap, LT-8 is not better than AGDC (53.47 vs 33.11 against real 32.92), but the FID and alignment comparisons still carry the point. Given no code, no dataset, and no error bars, the REJECT verdict is justified for the stated claim; a narrowed claim restricted to ContLayNet DRC performance could be CONDITIONAL. I therefore recommend no change to the reader's verdict.","tokens_in":13758,"tokens_out":6186,"duration_ms":58095,"concrete_test":"Produce one comparison table for PubLayNet and Rico that includes all LT precisions already listed in Table B (4,8,12,16,18), DLT, and AGDC on identical splits/metrics, and identify the best value per cell. If any non-18-bit LT configuration beats AGDC on FID on either task, the unqualified 'higher-fidelity' claim in the abstract fails; if the authors instead intend 'higher-fidelity at high precision only,' they must restate the claim and re-run Table 4 with an 8-bit IconShop included in the headline comparison. Repeating all runs with at least 3 seeds would also determine whether the reported gaps exceed noise, since no error bars are provided.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 6.2's claim that 'AGDC outperforms both LT (18 bits) and DLT across all tasks' depends on omitting the 8-bit LT results reported in the paper's own Table B. On PubLayNet completion, Table B gives LT-8 FID 3.00, Overlap 2.78, Alignment 0.08; Table 3 gives AGDC FID 4.58, Overlap 4.59, Alignment 0.17. On Rico completion, LT-8 FID 5.20 and Alignment 0.21 vs AGDC 9.77 and 0.25 (AGDC has lower Overlap, 33.11 vs 53.47). Thus on the primary fidelity metric the discretized baseline beats AGDC in both layout datasets. The SVG experiment similarly shows IconShop at 8 bits with FID 37.35 vs AGDC 48.73 (Table 4); AGDC is 'comparable' but not higher-fidelity. The only place AGDC clearly beats the included baselines is ContLayNet under DRC metrics against LT-18/DLT. Therefore the abstract's unqualified 'higher-fidelity hybrid vector representations than discretization-based ... baselines' is not supported by the paper's own numbers. The claim could be salvaged by restricting to matched-high-precision settings, but that is not the claim as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AGDC, an autoregressive framework that jointly models discrete and continuous values in variable-length sequences by combining categorical prediction for discrete identifiers with diffusion-based modeling for continuous vectors. Two technical components are introduced: an MLP-based EOS logit adjustment and a length regularization loss. The authors also introduce ContLayNet, a 334K-sample semiconductor layout benchmark with Design Rule Check (DRC) metrics. Experiments cover ContLayNet, PubLayNet/Rico layouts, and FIGR-8-SVG, comparing against LayoutTransformer, DLT, and IconShop. The central claim is that AGDC achieves higher-fidelity hybrid vector representations than discretization-based and fixed-schema baselines, especially at high precision.","tokens_in":14077,"tokens_out":3661,"duration_ms":41615,"significance":"If the claims are substantiated, the work addresses a real limitation of tokenized autoregressive models in continuous-valued domains and proposes a novel architecture that integrates diffusion into autoregressive generation for variable-length hybrid sequences. The ContLayNet benchmark and DRC metrics are potentially valuable resources for the semiconductor layout community. However, the central outperformance claim is not supported by the paper's own data: the appendix reports that an 8-bit discretized baseline (LT) achieves better FID on both layout datasets, and the SVG experiment shows IconShop at 8 bits has lower FID than AGDC. The contribution is therefore more modest than advertised, resting primarily on ContLayNet under DRC metrics. The method is coherent and the ablations suggest the two proposed components help, but the evidence does not justify the unqualified 'higher-fidelity' claim.","major_comments":[{"comment":"The claim in Section 6.2 that 'AGDC outperforms both LT (18 bits) and DLT across all tasks' omits the 8-bit LT results reported in the paper's own Table B. On PubLayNet completion, LT at 8 bits has FID 3.00 and Overlap 2.78 versus AGDC's 4.58 and 4.59; on Rico completion, LT at 8 bits has FID 5.20 and Alignment 0.21 versus AGDC's 9.77 and 0.25. Since Table B is part of the same manuscript, this is an internal inconsistency. The main comparison should include the best-performing precision of the baseline, or the claim should be restricted to the high-precision setting.","section":"Section 6.2, Table 3 vs Appendix B.1, Table B"},{"comment":"Table 4 shows IconShop at 8 bits achieves FID 37.35 while AGDC achieves 48.73. The text states AGDC 'maintain quality comparable to IconShop's best outputs,' but the abstract and introduction claim 'higher-fidelity hybrid vector representations than discretization-based ... baselines.' This is not supported. The claim should be revised to state that AGDC is superior only at high precision and comparable or worse at the precision where the discretized baseline performs best.","section":"Section 6.3, Table 4"},{"comment":"No error bars, confidence intervals, or significance tests are reported for any of the headline numbers. Since several comparisons involve small margins (e.g., PubLayNet ablation rows) and the central claim depends on specific baseline configurations, the absence of variance estimates makes it difficult to assess whether AGDC's improvements are meaningful. I request at least three seeds with standard deviations, or a paired test for the key comparisons.","section":"Section 6.1, Table 2 and Section 6.2, Table 3"},{"comment":"The authors justify evaluating LT at 18-bit precision as 'the maximum precision achievable with a batch size of 2 on a single NVIDIA A6000 GPU.' Precision is a property of the data representation, not the hardware; presumably the constraint is the vocabulary size and memory footprint. This needs clarification, and more importantly, it should be acknowledged that Table B shows 8-bit is the best-performing LT configuration for the layout tasks. The choice of 18-bit as the main baseline appears to select a configuration that is known to perform poorly, rather than the one that best represents the discretization approach.","section":"Section 6.1, 'Baselines' paragraph"}],"minor_comments":[{"comment":"The paragraph contains a duplicated statement: 'with substantial structural complexity, each sample averaging 323 layers' appears effectively twice. Please remove the repetition.","section":"Section 5.1"},{"comment":"The sentence 'MLPCONT consists of three blocks with 1024 channels each' is unclear: is this the continuous branch of the transformer, the diffusion conditioning network, or the denoising network? Please define all components consistently.","section":"Section 6.1, Implementation details"},{"comment":"The expected-length formula is written without derivation, and the adjusted EOS probability is not explicitly defined. Please specify that p_i^EOS is computed from softmax over the adjusted logits in Equation (9).","section":"Equation (12)"},{"comment":"The caption lists 'LT (8 bits) LT (18 bits) DLT (     bits) AGDC (     bits)' with blank precision values for DLT and AGDC. In several figures (e.g., Figure 5, C, D) the precision annotations are also blank. Fill these in or remove the placeholder text.","section":"Figure 7"},{"comment":"The ContLayNet dataset is stated to be released 'upon acceptance,' but no code repository or data availability link is provided. For a benchmark paper, releasing the data and evaluation code is essential for the contribution to be verifiable.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The paper's own appendix (Table B) directly contradicts the abstract's 'higher-fidelity' claim, and the SVG result shows the discretized baseline winning on FID. This is not a matter of presentation alone — the central claim must be narrowed to 'superior at high precision' and the main tables should include the best-performing baseline configuration. The ContLayNet benchmark and DRC metrics are the strongest remaining contribution, but the authors should also release the data/code before acceptance. If the authors are unwilling to revise the claim, rejection would be appropriate; with an honest re-framing, the paper may be publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know before reading: the architecture is genuinely interesting, but the central empirical claim does not survive contact with the paper's own table in the appendix. The main text says AGDC outperforms discretization baselines across all tasks, but Table B shows LayoutTransformer at 8 bits getting FID 3.00 on PubLayNet completion versus AGDC's 4.58, and 5.20 versus 9.77 on Rico. The main tables only run LT at 18 bits, where it does poorly. That is a selective comparison, and the abstract's unqualified \"higher-fidelity than discretization-based baselines\" is not supported.\n\nWhat is actually new: combining a softmax discrete head with MAR-style diffusion in a variable-length autoregressive loop, plus the EOS logit adjustment and length regularization. The ablations show both components help, and the precision-scaling analysis in Section 3.1 is correct, if elementary. ContLayNet with DRC metrics is a reasonable attempt to measure functional correctness in a precision-sensitive domain, and AGDC's DRC numbers are much better than the baselines. That part holds up.\n\nSoft spots beyond the cherry-pick: no error bars anywhere, so we cannot tell if AGDC's margins over DLT are meaningful. The dataset and code are not released despite being central to the evaluation. The SVG experiment is also weak — AGDC's FID is 48.73 versus IconShop's 37.35 at 8 bits, so \"comparable\" is doing a lot of work. The ContLayNet DRC metrics are only evaluated on the authors' own data, with no external check on whether those rules capture realistic design constraints.\n\nI don't think the central method is broken. The idea is coherent, and the ContLayNet results are promising. But the paper as written overclaims, and the fix is straightforward: report LT at its best precision in the main tables, or explicitly restrict the claim to high-precision settings while acknowledging the trade-off. If they do that, the contribution stands on the architecture and the benchmark.\n\nThis is worth a serious referee. A good reviewer could push for honest comparisons, error bars, and dataset release. I would bring it to a reading group and I'd cite it for the hybrid AR-diffusion formulation and ContLayNet, but I would not take the empirical superiority claim at face value.","headline":"AGDC has a genuinely new hybrid autoregressive-diffusion recipe and a potentially useful benchmark, but the main tables hide the appendix's own LT-8 results that beat it on FID.","tokens_in":14580,"tokens_out":1893,"would_cite":true,"duration_ms":22004,"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":"AGDC jointly predicts discrete identifiers and diffuses continuous coordinates inside one autoregressive transformer, giving variable-length hybrid vectors effectively unbounded precision.","keywords":["autoregressive generation","continuous diffusion","hybrid discrete-continuous sequences","layout generation","vector graphics","precision scalability","semiconductor layouts","design rule checks"],"falsifier":"Re-run Table 3 and Table 2 with LayoutTransformer at every precision from 4 to 18 bits (and DLT with tuned length control), then compare against AGDC on the same metrics; if an 8-bit or 12-bit LT equals or beats AGDC on FID, overlap, and alignment, the claim that discretization fundamentally caps achievable fidelity is falsified for graphic layouts. On ContLayNet the decisive test is whether any discretized precision passes the DRC metrics at the same rate as AGDC's continuous outputs.","tokens_in":13628,"feed_emoji":"📐","tokens_out":5905,"duration_ms":60028,"temperature":0.7,"pith_summary":"AGDC argues that token-based discretization caps the precision of autoregressive generation, and that the fix is to stop quantizing continuous coordinates. It builds an autoregressive transformer whose discrete branch classifies element types while a small diffusion network, conditioned on the same latent vector, generates each continuous coordinate vector in its natural real-valued form. Two additions make sequence length controllable: an MLP that adjusts the end-of-sequence logit from context, and a differentiable length-regularization loss. The paper claims this beats discretization-based and fixed-schema baselines on semiconductor layouts (its new ContLayNet benchmark), graphic layouts, and text-to-SVG, and that precision errors, not just visual similarity, should be measured. If right, autoregressive models can scale to high-precision engineering domains without vocabulary explosion.","feed_headline":"Diffusion supplies the precision token quantization drops","feed_subtitle":"A hybrid model predicts discrete commands and diffuses continuous coordinates, keeping chip layouts and SVG icons at full precision.","key_machinery":"The load-bearing object is the atomic unit: a discrete identifier plus a real-valued vector. The architecture is a transformer decoder whose latent z_i feeds two parallel heads — an MLP+softmax for the discrete identifier and an AdaLN-conditioned denoising MLP that predicts the noise to remove from c_i (a MAR-style diffusion head). Sequence length is controlled by adding α·MLPEOS(z_i) to the EOS token logit and by Lℓ=(E[length]−L_target)^2, where E[length] is computed from adjusted EOS probabilities. This is what lets one model do categorical prediction and diffusion in the same autoregressive loop without fixing a schema or a token budget.","core_discovery":"The central claim is that a hybrid sequence of discrete identifiers and continuous vectors can be modeled autoregressively without lossy transformation: discrete values remain categorical predictions, and continuous values are sampled by a reverse diffusion process conditioned on the autoregressive latent z_i. The model represents each atomic unit u_i=[d_i,c_i] as e_i=concat(E(onehot(d_i)), W c_i), feeds previous embeddings to a transformer to obtain z_i, then predicts d_{i+1} with softmax and c_{i+1} with a denoising MLP. To terminate at the right length, the EOS logit is adjusted by MLPEOS(z_i), and training includes a squared penalty on expected length. The paper introduces ContLayNet, 33","pith_inferences":["Editorial: the main tables evaluate LayoutTransformer at 18 bits, but the supplement's Table B shows 8-bit LT achieves lower FID than AGDC on PubLayNet completion (3.00 vs 4.58) and Rico completion (5.20 vs 9.77); a head-to-head against each baseline's best-performing precision would be the more conservative test of the 'discretization fundamentally limits fidelity' claim.","Editorial: because the continuous branch samples with 10–100 diffusion steps per coordinate vector, inference is much slower per step than a discrete token; the reported 50-step Improved DDPM configuration trades quality for speed, and faster samplers or distillation would likely be needed for interactive design tools.","Editorial: the atomic-unit abstraction suggests immediate testable extensions — PCB routing, floor plans, and CAD sketches — and a natural stress test is adding curved or nested paths, where a fixed 8-dimensional coordinate vector may be insufficient.","Editorial: one could test the length-control mechanism in isolation across datasets by measuring calibration of P(EOS) under distribution shift; the paper only reports Gaussian-fitted length-error statistics on ContLayNet."],"forward_implications":["High-precision domains like semiconductor layout no longer need exponentially growing vocabularies; the continuous branch's diffusion output is effectively unbounded in resolution (15-bit or beyond) on a fixed model.","Variable-length generation becomes differentiable and controllable: the EOS adjustment and length loss reduce length error from a biased mean (~39) to near zero (~0.15) in the paper's ContLayNet measurement.","Autoregressive methods can be extended to any atomic-unit-style hybrid domain (typed boxes or path commands), not just layouts and SVG, by reusing the same embedding-branch structure.","Functional evaluation becomes feasible: DRC-based metrics catch failures that FID misses, so generative models for chips can be scored by whether the circuit would work.","In text-to-SVG, the continuous model keeps FID/CLIP quality comparable to the best 8-bit baseline while remaining functional at 10+ bits, where the discretized baseline fails entirely."],"fun_headline_variants":["Hybrid AR+diffusion: no lossy token quantization","Autoregressive discrete + continuous diffusion for precision","Chip layouts and icons: full precision generation","EOS adjustment and regularization: better length control","Discrete cats, continuous diffusion: vector modeling without loss"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The main comparison claims rest on evaluating LayoutTransformer at 18-bit precision; the paper's own Table B shows the 8-bit configuration achieves lower FID on PubLayNet and Rico completion (3.00 vs 4.58 and 5.20 vs 9.77), so if the right baseline is each method's best-performing precision, the central outperformance claim weakens substantially.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid AR+diffusion: no lossy token quantization","Autoregressive discrete + continuous diffusion for precision","Chip layouts and icons: full precision generation","EOS adjustment and regularization: better length control","Discrete cats, continuous diffusion: vector modeling without loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000143,"raw_usage":{"total_tokens":1012,"prompt_tokens":755,"completion_tokens":257,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":181}},"tokens_in":499,"tokens_out":257,"duration_ms":3958,"temperature":1.0,"reasoning_tokens":181,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T11:34:14.686999+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run Table 3 and Table 2 with LayoutTransformer at every precision from 4 to 18 bits (and DLT with tuned length control), then compare against AGDC on the same metrics; if an 8-bit or 12-bit LT equals or beats AGDC on FID, overlap, and alignment, the claim that discretization fundamentally caps achievable fidelity is falsified for graphic layouts. On ContLayNet the decisive test is whether any discretized precision passes the DRC metrics at the same rate as AGDC's continuous outputs.","supporting_citations":[],"review_version":1}