{"id":"4963918c-222f-4a1e-9197-c2196a0f9ca5","arxiv_id":"2607.18536","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A multi-agent vision-language framework for chip macro placement improves post-route timing (WNS/TNS) over commercial, human, and Hier-RTLMP baselines while trading higher wirelength, and introduces four human-likeness metrics.","lead":"MAGE is a multi-agent AI system that rearranges large circuit blocks on a chip using expert rules and visual checks, and it reports better timing than commercial tools and a human expert on nine designs. It matters because chip design still spends days to weeks on manual macro placement, and this is one of the first agentic pipelines aimed directly at that bottleneck.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Tournament selection uses eGR rWL as a proxy for timing, but the asserted proxy correlation (Sec. IV-G) is never shown; if weak, the reported PPA gains are not attributable to the claimed mechanism.","rationale":"The reader identified the eGR-rWL proxy as the weakest assumption; I agree. It is the only place where the manuscript explicitly asserts an empirical relationship without providing supporting data. The claim appears in Section IV-G and is used to justify the tournament's selection rule, which directly produces the headline P* reported in Tables II–III. Without a correlation analysis, the tournament's 9% rWL improvement (Fig. 12) cannot be tied to the WNS/TNS improvements; the causal story 'human-like structure → timing closure' is then supported only by the base-pipeline ablations and post-hoc path analysis, not by the tournament. This is a significant but addressable gap.\n\nOther concerns—no repeated seeds, circular human-likeness metrics, missing code artifacts—are real but secondary. Lack of variance affects statistical confidence but is mitigated by consistency across nine designs; metric circularity affects the human-likeness sub-claim, not the PPA headline; missing code is a reproducibility matter. The eGR proxy gap directly undermines the mechanism stated as the paper's contribution.\n\nA single correlation check on tournament candidates would settle it. If the proxy holds, the concern is closed. If not, the paper should be revised to either select on post-route timing at lower tournament cost or present the base-pipeline results separately. This does not change the reader's CONDITIONAL verdict; it sharpens the condition that must be met before ACCEPT.","tokens_in":24215,"tokens_out":6437,"duration_ms":74370,"concrete_test":"Take one NG45 design (e.g., Ariane) and one GF12 design (e.g., Tabla01). Rerun Algorithm 3 for all 6 rounds while recording, for every candidate placement in the evaluation history H, both the eGR rWL µ_v (already computed) and the full post-route WNS/TNS from Innovus after place-and-route. Compute Spearman and Pearson correlations between µ_v and WNS/TNS (e.g., ~36 points per design). If |ρ| < 0.5 or the sign is wrong, the tournament's selection criterion is not a valid timing proxy; the paper should then report PPA both with and without tournament selection to determine whether the claimed timing gains survive. If |ρ| > 0.7 with correct sign, the proxy concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central causal claim—that human-like macro organization yields post-route timing gains—relies on Algorithm 3 selecting P* by minimum eGR routed wirelength. Section IV-G states that eGR rWL is 'empirically a strong proxy for downstream PPA metrics (Table II)', but Table II reports only post-route rWL/Power/WNS/TNS; no eGR rWL values and no eGR-versus-post-route correlation analysis are presented anywhere. Since eGR rWL is explicitly distinguished from post-route rWL in Section VI, the proxy claim is not tautological and is load-bearing: all reported WNS/TNS improvements are for P* = argmin over µ_v (Alg. 3, lines 9-10, 23). If eGR rWL correlates weakly or negatively with WNS/TNS, the tournament is optimizing the wrong objective, and the 9% rWL improvement in Fig. 12 may not translate to timing. The final PPA gains could then come from the base six-phase pipeline or from selection among 36 stochastic variants, not from the tournament mechanism the paper emphasizes. The ablation tables (V, VI) show that K and visual feedback affect timing, so the concern is specifically about the tournament component; but because the headline results are for the full system with tournament enabled, the missing proxy validation leaves the central attribution unproven.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MAGE, a multimodal multi-agent framework for macro placement refinement. Ten agents execute a six-phase pipeline (parsing, hierarchy-aware grouping, group placement, feasibility, reference placement/transform-and-merge, independent validation) and optionally a Go-With-The-Winners tournament that selects among variants by Innovus early-global-routing (eGR) routed wirelength. Expert floorplanning knowledge is encoded as natural-language directives, visual checks, and validation criteria rather than learned from labeled placements. Four human-likeness metrics are introduced. Experiments on nine designs across NanGate45 and GF12 report geometric-mean WNS/TNS improvements of 11.1-19.3%/70-74% over two commercial macro placers, 18.3%/72.5% over a human expert, and 47.0%/80.4% over Hier-RTLMP, at the cost of higher wirelength and modestly higher power versus the commercial placers. Ablations show the knowledge corpus and visual feedback affect timing; additional case studies address anonymized netlists, unseen designs, rectilinear floorplans, and high utilization. The paper's central claim is that human-like macro organization provides a useful structural prior for downstream timing closure.","tokens_in":24594,"tokens_out":6202,"duration_ms":76066,"significance":"The result is significant if it holds: it would demonstrate that a multimodal, multi-agent LLM system can encode expert physical-design judgment and improve post-route timing over commercial and human baselines without any training on labeled placements. The evaluation has real strengths: it uses a standard Innovus post-route flow, reports full PPA across two enablements, includes a human-expert baseline, presents ablations and cost/runtime data, and points to a public repository. However, two caveats are load-bearing: the tournament's selection metric (eGR routed wirelength) is asserted to be a strong proxy for downstream PPA but the correlation is never shown, and all results are single stochastic runs with no error bars. These caveats currently prevent full confidence in the causal attribution claimed in the abstract and conclusion.","major_comments":[{"comment":"The tournament selects P* by minimum eGR routed wirelength, and Section IV-G states that eGR rWL is 'empirically a strong proxy for downstream PPA metrics (Table II)'. Table II, however, contains only post-route rWL, power, WNS, and TNS; no eGR rWL values and no eGR-versus-post-route correlation are provided anywhere. Since Section VI-A explicitly distinguishes post-route rWL from eGR rWL, the proxy claim is not tautological. All headline WNS/TNS improvements are for the tournament-selected P*, so if eGR rWL correlates weakly or negatively with timing, the reported gains could come from the base six-phase pipeline or from best-of-N selection rather than from the mechanism the paper emphasizes. Please supply the eGR-vs-post-route correlation per benchmark, or an ablation that selects on post-route timing.","section":"§IV-G, Algorithm 3 (lines 9-10, 23); §VI-A"},{"comment":"The 'Human' baseline is not described. The reader is told only that human-expert baselines are available on the three NG45 designs, with no information about who produced the placements, under what instructions, whether one or multiple experts were involved, or whether the placements come from a published benchmark. Since the abstract and Section VI-A claim an 18.3% WNS / 72.5% TNS improvement over the human expert, this baseline must be defined and, ideally, verified or cited precisely.","section":"§VI experimental setup; Table II"},{"comment":"All reported results appear to be single runs. The pipeline is stochastic (LLM sampling, tournament variants, validation iterations), yet no standard deviations, number of seeds, or repetition protocol are reported. The per-run cost is acknowledged, but without variance information the geometric-mean improvements cannot be distinguished from selection effects among the sampled variants. At minimum, seeded reruns for one NG45 and one GF12 design, and confidence intervals for the headline metrics, would be needed to support the precision of the claimed improvements.","section":"Tables II, V-VIII, X; Figures 8 and 12"},{"comment":"The four human-likeness metrics are direct operationalizations of the six principles that are also encoded as prompts, validation checks, and visual criteria in the MAGE pipeline. Consequently, Table IV and Figure 8 partly measure the method against its own rule system. The paper's 'Metric design and limitations' paragraph is candid about this, but the central premise that 'human-like macro organization provides a useful structural prior' would be materially stronger with an external validation, e.g., independent human judgments of placement quality or a comparison against human placements not used to define the metrics.","section":"§III, §III-A, §IV"}],"minor_comments":[{"comment":"The anonymization study is confounded: anonymizing names also disables the knowledge corpus K, so the observed timing degradation could be due to K removal rather than to loss of semantic names. The text claims the study isolates what names contribute beyond structure, but K removal is a second independent change. Please ablate name anonymization with K retained, or explicitly acknowledge the confound.","section":"§VI-D.1, Table VII"},{"comment":"The Pearson correlation is reported as r=+0.38, p=0.053, which is not statistically significant at the 0.05 level; the text should not rely on it without reporting the sample size and should give more weight to the Spearman result or describe the relationship as suggestive.","section":"Figure 8"},{"comment":"The notation for effective column widths and row heights (\\hat{w}_i, \\hat{h}_j) and channel spacings (c_x, c_y) is introduced only in prose; a one-sentence definition in the equation caption or immediately before would improve readability.","section":"§IV-D, Eq. (1)"},{"comment":"For the Mixed rectilinear variant, MAGE does not always improve TNS (Ariane-Mixed and BP-Mixed show roughly equal or worse TNS). The text says only that WNS improves on all four variants; the TNS tradeoff should be stated explicitly for these cases.","section":"§VI-D.3, Table IX"},{"comment":"The caption says 'Normalized Routed Wirelength', which could be confused with the post-route rWL reported in Table II. Since the tournament uses eGR rWL, the caption and axis label should say 'eGR routed wirelength'.","section":"Figure 12 caption"}],"recommendation":"major_revision","confidential_remarks":"The reader's strongest concern is on point: the eGR proxy assertion in Section IV-G is not backed by data in Table II, and it is load-bearing for the tournament attribution. I would ask the authors for the eGR-versus-post-route correlation or a timing-based selection ablation before accepting the central causal claim. The paper is otherwise a serious, well-executed empirical study, and the missing items are addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a serious referee, but the headline claim is not yet load-bearing. What's new is that MAGE makes macro placement decisions directly through a multi-agent VLM pipeline rather than just proposing anchors or cost hints for a base placer, and the evaluation is more transparent than most: a standard Innovus post-route flow, human-expert and Hier-RTLMP baselines, and an explicit statement that better timing comes at the cost of 29.7% higher routed wirelength and about 4% higher power. That tradeoff deserves credit. The four human-likeness metrics are well-specified and reusable.\n\nThe biggest soft spot is the tournament mechanism. Algorithm 3 selects P* by minimum early-global-routing (eGR) routed wirelength, and Section IV-G justifies that by saying eGR rWL is 'empirically a strong proxy for downstream PPA metrics (Table II)'—but Table II contains post-route results only, and no eGR-versus-post-route correlation appears anywhere in the manuscript. That is a load-bearing missing analysis. If eGR rWL does not track WNS/TNS, the tournament is optimizing the wrong objective, and the reported PPA gains cannot be attributed to the mechanism the paper emphasizes. The ablation tables suggest the base pipeline and visual feedback carry much of the effect, but the headline numbers are for the full system with tournament enabled.\n\nOther soft spots are more routine: the LLM pipeline is stochastic but every result is a single run with no seeds or variance; the ablations are on two designs only; and the human-likeness metrics are partly circular, since the principles encoded in the prompts and visual checks are the same ones being scored. The external PPA results stand on their own, but the 'human-like' claim needs independent expert validation. Code and data are referenced but not verifiable from the manuscript.\n\nThese are addressable, not disqualifying. A revision should show the eGR-to-post-route correlation (or change the tournament objective), report variance across seeds, validate the metrics against external expert ratings, and make the repository verifiable. Send it to peer review; the missing analyses can be required of the authors.","headline":"A serious, honest integration of agentic LLM reasoning into macro placement with an unproven tournament mechanism; send to reviewers, but the eGR proxy and single-run results need to be addressed.","tokens_in":25075,"tokens_out":4116,"would_cite":true,"duration_ms":46036,"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":"MAGE claims that encoding expert floorplanning heuristics as natural-language rules and visual checks, in a ten-agent pipeline, produces human-like macro placements that improve post-route timing by 11–19% in WNS and 70–74% in TNS over comm","keywords":["macro placement","chip floorplanning","multi-agent systems","vision-language model","multimodal reasoning","timing closure","routed wirelength","human-likeness metrics"],"falsifier":"Run one tournament round on a design and record, for every variant, both the pre-routing wirelength that selects survivors and the final post-route WNS/TNS. If the variant with the best proxy is not generally the best-timing variant, or if the rank correlation between the proxy and timing is near zero or negative, then the tournament's contribution to the reported timing improvements is unsupported.","tokens_in":24091,"feed_emoji":"🧠","tokens_out":6411,"duration_ms":70818,"temperature":0.7,"pith_summary":"MAGE tackles macro placement, the task of placing large pre-designed blocks in a chip floorplan. The paper's premise is that expert engineers arrange macros according to structural principles—boundary placement, regular stacking, IO keepouts, avoiding unusable pockets, and preserving contiguous whitespace—and that these principles, not raw wirelength optimization, are what make a floorplan easy to route and close timing. Rather than training a model on labeled placements, MAGE encodes those principles as natural-language directives and visual validation gates in a ten-agent pipeline, then uses a tournament to refine candidate placements. Across nine designs and two technology nodes, it reports large timing improvements over commercial and hierarchy-aware baselines, at the cost of higher routed wirelength in most comparisons. The decisive assertion, supported by case studies on unseen, rectilinear, and high-utilization designs, is that human-like macro organization is a useful structural prior for downstream routability and timing closure.","feed_headline":"Human-like chip layout beats commercial placers on timing","feed_subtitle":"A rule-and-image-driven pipeline improves worst-slack and total-slack by 11–74% without training on labeled placements.","key_machinery":"The framework is MAGE, a multi-agent system of ten specialized agents organized into six phases: data parsing, hierarchy-aware grouping, group placement, feasibility checking, macro reference placement, and independent validation. Two mechanisms carry the argument: mandatory visual gates, in which a vision-language model inspects rendered flyline diagrams, whitespace heatmaps, and boundary-coverage maps and must report visual-only facts; and a go-with-the-winners tournament that runs six placement variants in parallel, ranks them by a fast pre-routing wirelength estimate, keeps the top three survivors, and propagates analysis and reflection feedback. The four human-likeness metrics—notch, wh","core_discovery":"The paper's central discovery is that human-like macro organization is a useful structural prior for downstream timing closure. It shows that a purely prompt-and-vision-driven system, with no training on labeled placement data, can produce macro placements that align with six expert principles and, across nine benchmarks, improve geometric-mean worst negative slack by 11.1%–19.3% and total negative slack by 70.0%–74.0% over commercial macro placers. On the three designs with human-expert baselines, it improves WNS by 18.3% and TNS by 72.5% while keeping wirelength and power comparable. The paper attributes the gains to lower clock insertion delay, reduced clock skew, and lower congestion aro","pith_inferences":["A direct test of the paper's mechanism would be to run the tournament with WNS or TNS (or a fast timing estimate) as the in-loop objective; if the reported timing gains persist, the structural-prior story is confirmed, and if they vanish, the gains likely come from selection rather than the human-like rules.","The frequent tradeoff between higher total wirelength and better timing suggests that total wirelength is the wrong proxy for the final objective; a natural extension is to optimize congestion-aware or critical-path-aware wirelength inside the tournament.","Because the human-likeness metrics are computable from a floorplan alone, they could be inserted as regularizers into learned placers or used to build a fast surrogate for routing outcomes—something the paper does not propose.","The finding that anonymizing design names degrades timing hints that semantic labels encode regular arrays and symmetry that connectivity alone does not; an inference worth testing is whether graph-isomorphism-style symmetry detection could recover this without human-readable names."],"forward_implications":["Macro placement can be framed as a knowledge-encoding problem: expert floorplanning rules written in natural language, plus visual checks, can substitute for labeled training data.","Structural regularity—boundary alignment, regular stacks, and contiguous whitespace—is worth pursuing even when it increases total routed wirelength, because the timing gains come from reduced clock skew and congestion.","The tournament's roughly 9% wirelength improvement from analysis and reflection feedback suggests that iterative, feedback-driven refinement is where much of the placement-quality gain lives.","Zero-shot transfer to unseen, rectilinear, and high-utilization designs implies that the expert principles, rather than memorized design-specific layouts, are doing the work.","The four human-likeness metrics give designers a vocabulary for floorplan quality beyond PPA and could be used as design-time checks for structural regularity."],"fun_headline_variants":["No-training multimodal agents beat commercial placers on timing","Human-like chip layout improves slack up to 74% without training data","Agentic multimodal framework improves on human expert placement","No-training multimodal agent improves chip timing 11-74% over commercial","Human-like macro placement: 74% TNS gain without any training data"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the fast pre-routing wirelength estimate used to rank candidate placements in the tournament faithfully predicts the final post-route timing metrics (WNS and TNS) that the headline gains are measured on; the paper asserts this correlation but does not show it.","fun_headline_variants_meta":{"raw":{"variants":["No-training multimodal agents beat commercial placers on timing","Human-like chip layout improves slack up to 74% without training data","Agentic multimodal framework improves on human expert placement","No-training multimodal agent improves chip timing 11-74% over commercial","Human-like macro placement: 74% TNS gain without any training data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001015,"raw_usage":{"total_tokens":4183,"prompt_tokens":863,"completion_tokens":3320,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":3242}},"tokens_in":607,"tokens_out":3320,"duration_ms":26863,"temperature":1.0,"reasoning_tokens":3242,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T15:07:04.895042+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run one tournament round on a design and record, for every variant, both the pre-routing wirelength that selects survivors and the final post-route WNS/TNS. If the variant with the best proxy is not generally the best-timing variant, or if the rank correlation between the proxy and timing is near zero or negative, then the tournament's contribution to the reported timing improvements is unsupported.","supporting_citations":[],"review_version":1}