{"id":"7d0d7dd9-90c3-49b8-90a6-a87fa5941bbb","arxiv_id":"2607.03578","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Generative AI for MIT materials should target mechanism-informed switchable phase landscapes via modular physics adapters and staged verification under label scarcity.","lead":"This Perspective argues that generative AI for metal–insulator-transition materials should propose mechanism-informed phase-transition hypotheses, not just stable crystals. A modular physics-adapted generator plus staged verification could help under severe label scarcity.","discovery_kind":"review","skeptic_critique":{"model":"grok-4.5","headline":"The load-bearing premise is untested: proxy-trained adapters may not enrich true phase-pair hypotheses under intertwined MIT physics.","rationale":"The Reader correctly identifies the strongest claim as a design thesis for mechanism-informed phase-transition generation under label scarcity, and the weakest assumption as the untested claim that proxy-conditioned adapters will enrich survivors of the multiphysics ladder. That is the single load-bearing soft spot: everything else (scarcity of verified labels, multiphysics character of MITs, inadequacy of single-structure generation, staged verification) is well supported by the cited literature and the Georgescu-style dataset discussion. Because this is a Perspective with no model, code, or enrichment result, the concern does not introduce a new correctness failure; it confirms why CONDITIONAL is the right verdict. No stronger internal inconsistency appears. A small pilot enrichment test against the paper's own quantitative gates would settle whether the adapter premise is operationally useful or remains aspirational. Verdict stays CONDITIONAL; no upgrade or rejection is warranted from this stress pass.","tokens_in":26741,"tokens_out":671,"duration_ms":6215,"concrete_test":"Run a minimal pilot: freeze a pretrained crystal diffusion generator; train one LoRA domain adapter on TM oxides/chalcogenides and one mechanism-prior adapter on a single proxy (e.g., short metal–metal bonds / soft dimerization modes for Peierls-like bias). Generate N candidates; score the fraction that pass the paper's own gates (Ehull ≲ 50 meV/atom, phase-pair ΔE ≲ 25–50 meV/f.u., N(EF) contrast, soft mode linking phases) versus an unadapted baseline. If enrichment is not clearly above baseline (or collapses to known MIT-like chemistries), the adapter premise does not hold as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (Abstract; §3–4) is that a modular generator with mechanism-prior adapters trained on scalable proxies (U/W, metal–metal distances, bond disproportionation, soft modes, etc.; Table 2) over large oxide/chalcogenide pools—not the ~68 verified MIT positives—will produce a meaningfully enriched distribution of phase-pair hypotheses that survive the verification ladder (phase competition, electronic contrast, pathway, control). That premise is asserted, not shown. Because known MITs are multiphysics and contested (VO2, 1T-TaS2, Ta2NiSe5; Table 1, Fig. 3), proxies may bias toward familiar chemistries or single-structure motifs without raising the rate of accessible, electronically contrasting phase pairs. The manuscript itself flags collapse to training-like structures and proxy bias as failure modes (§3) but provides no enrichment metric, ablation, or pilot against the Georgescu set / Supplementary Table S1. Without that, the Perspective is a coherent design proposal whose practical claim remains open.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"This Perspective argues that generative AI for metal–insulator-transition (MIT) discovery should target mechanism-informed phase-transition hypotheses rather than stable single crystals alone. Under severe label scarcity (a curated set of ~343 MIT-related entries with only ~68 verified MIT positives), the authors propose a modular pipeline: a pretrained crystal diffusion generator, a domain adapter for transition-metal oxides/chalcogenides, and lightweight mechanism-prior adapters (correlation, Peierls/dimerization, charge/orbital order, excitonic/CDW, disorder) trained from scalable physics proxies rather than from the small verified MIT set. Generated candidates are treated as hypotheses and filtered by a staged verification ladder (chemical validity, thermodynamic accessibility, phase competition, electronic contrast, pathway plausibility, experimental/device accessibility), with explicit early-triage energy and contrast gates. The manuscript synthesizes the multiphysics character of known MITs (VO2, nickelates, 1T-TaS2, Ta2NiSe5, CuIr2S4, etc.), frames inverse design as search over switchable phase landscapes, and outlines a closed-loop outlook for datasets, phase-pair benchmarks, and standardized enrichment reporting.","tokens_in":27089,"tokens_out":1607,"duration_ms":21818,"significance":"The paper addresses a genuine and under-articulated gap: modern crystal generators and stability models optimize single-structure properties, whereas MIT function is a relation between states under a control parameter and device window. The scarcity argument is well grounded in the Georgescu/Rondinelli dataset, the multiphysics examples match the literature, and the verification-first stance (including honest failure modes such as proxy bias and collapse to familiar chemistries) is a useful corrective to overclaiming in generative materials AI. Strengths include a clear modular design philosophy, a concrete physics-prior table (Table 2), operational reporting gates, and a substantial supplementary compilation of the 68 MIT-positive compounds with temperatures and mechanism notes. If adopted, the framework could reorient evaluation of generative models from validity/novelty/stability toward enrichment of testable phase-pair hypotheses. As a Perspective without a trained model or pilot enrichment study, the contribution is conceptual and agenda-setting rather than empirical; its lasting value depends on whether the community can operationalize the proposed proxies and ladder.","major_comments":[{"comment":"The load-bearing premise of §§3–4 and Table 2—that mechanism-prior adapters trained or conditioned on scalable proxies (U/W, metal–metal distances, bond disproportionation, soft modes, nesting, etc.) over broad oxide/chalcogenide pools will meaningfully enrich accessible, electronically contrasting phase-pair hypotheses—is asserted but not demonstrated. The manuscript correctly reserves the ~68 verified positives for calibration and enrichment evaluation and flags proxy bias and training-set collapse as failure modes, yet it provides no enrichment metric definition, baseline comparison (generic vs domain-adapted vs mechanism-adapted generation), or even a small retrospective pilot against the Georgescu set / Supplementary Table S1. For a Perspective this need not be a full model, but the central claim remains open without a concrete, falsifiable success criterion (e.g., fraction of candi","section":null},{"comment":"Section 4 proposes operational triage gates (~50 meV atom−1 of the hull as strong metastability; ~25–50 meV per formula unit or active metal site for phase-pair separation; order-of-magnitude N(EF) suppression or gap opening for contrast). These are useful reporting conventions, but their calibration against known MIT energy scales is thin. Figure 3 and Table 1 emphasize that many canonical transitions sit at tens of meV or less (Fe3O4 ~10 meV, BaVS3 ~6 meV, V2O3 ~13 meV), while high-T systems such as NbO2 are much larger; free-energy, magnetic, and entropic contributions are acknowledged only qualitatively. Please either (i) justify the numerical windows more carefully against computed or experimental energy scales for a subset of the 68 positives, or (ii) reframe them more clearly as provisional reporting gates rather than physics-derived thresholds, and discuss how magnetic order, U c","section":null},{"comment":"The manuscript repeatedly states that the generated object should be a phase pair plus control parameter and operating constraint (Abstract; Fig. 4a; §3), yet the proposed implementation builds on single-structure crystal diffusion models with LoRA-style adapters. The architectural path from single-structure generation to joint sampling or sequential proposal of related metallic/insulating (or high-T/low-T) phases is underspecified. Without a clearer sketch—e.g., paired generation under a shared composition, distortion-coordinate conditioning, or post-generation competing-phase search—the claim that the generator itself proposes transition hypotheses (as opposed to single crystals later checked for competition) risks overstating what current modular adapters deliver. Please clarify the intended generation object and the minimal algorithmic steps that turn adapter-biased samples into phas","section":null}],"minor_comments":[{"comment":"Figure 2’s caption and body text describe a data-scarcity vs target-richness mismatch, but the figure itself (as rendered in the manuscript) is dominated by a schematic of what an MIT label encodes; consider aligning the visual more tightly with the scarcity hierarchy stated in the text (general structures → MIT-related labels → verified positives → mechanism-specific cases).","section":null},{"comment":"Table 1 is valuable but mixes control abbreviations, approximate kBT scales, and mechanism status unevenly across rows; a consistent column for approximate energy scale (meV) and a short note on whether the mechanism is contested would improve usability.","section":null},{"comment":"Supplementary Figure S1 and Table S1 are substantial assets; the main text should point more explicitly to them when discussing the 68 positives and near-room-temperature coverage (e.g., in §2 and Fig. 3).","section":null},{"comment":"A few repeated or near-duplicate references appear (e.g., Shao et al. on VO2 appears more than once; Verwey 1939 and Morin 1959 are cited in both main and supplementary lists). Consolidate where possible.","section":null},{"comment":"Terminology alternates between “mechanism-prior adapters,” “physics-prior adapters,” and “mechanism-aware” generation; pick one primary term and use it consistently after first definition.","section":null},{"comment":"The device-facing layer (operating window, hysteresis, thin-film compatibility, cycling) is invoked throughout but is thinner than the mechanism layer in Table 2 and Fig. 4b. A short paragraph or table row on how device constraints would enter as adapters or filters would balance the framework.","section":null}],"recommendation":"minor_revision","confidential_remarks":"This is a coherent, well-referenced Perspective rather than an empirical methods paper. The skeptic’s concern about untested proxy enrichment is real but is the natural open question for this genre; requiring a full trained model would be disproportionate. Fit is good for a materials/condensed-matter journal that publishes Perspectives on AI-for-science; less so if the venue expects new algorithms or benchmarks. I would not block on the absence of a pilot if the authors add a clear falsifiable evaluation protocol and tighten the phase-pair generation sketch. No integrity or citation-pattern concerns stood out beyond ordinary self-contained literature synthesis."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a Perspective, not a methods paper. The useful punchline is simple: for MITs under label scarcity, stop treating the generative target as a stable crystal or a small-gap score and treat it as a mechanism-informed phase-pair hypothesis that still has to clear phase competition, electronic contrast, pathway, and control checks.\n\nWhat is actually new is the packaging. They put a pretrained crystal diffusion prior, a TM-oxide/chalcogenide domain adapter, and LoRA-style mechanism-prior adapters (correlation, dimerization, charge/orbital, excitonic/CDW) next to a staged verification ladder with quantitative early gates (~50 meV/atom hull, ~25–50 meV phase-pair separation, N(EF)/gap contrast). They correctly refuse to train mechanism adapters by splitting the ~68 verified positives, and they reserve that set for calibration and enrichment. The multiphysics framing is solid: VO2, nickelates, 1T-TaS2, Ta2NiSe5, CuIr2S4 are used as intertwined cases, not clean labels. The Georgescu/Rondinelli 343/68 scarcity point is cited honestly, and the supplementary temperature/mechanism tables are careful about contested physics.\n\nThe soft spot is exactly the load-bearing premise: that proxy-trained adapters over large oxide/chalcogenide pools will enrich true switchable phase pairs rather than familiar chemistries or single-structure motifs. The paper names collapse and proxy bias as failure modes but shows no pilot enrichment, ablation, or ladder pass rates against the known set. As a Perspective that is not fatal; it is just the open claim. No new model, code, or compound is delivered, so do not over-read it as a validated pipeline.\n\nWho it is for: people building inverse-design or autonomous-lab loops for correlated oxides/chalcogenides, neuromorphic materials, or phase-change photonics who need a better objective than “stable + small gap.” Citation pattern looks current and appropriate (CDVAE/DiffCSP-style generators, MatterGen-scale models, LoRA, Georgescu dataset, classic MIT literature). Math is light and not load-bearing; the argument is conceptual.\n\nI would send it to peer review as a Perspective. Engage if you care about how generative models should be scored for switchable quantum materials; skip if you only want trained models or new candidates.","headline":"Clear Perspective that reframes MIT generation as phase-pair hypotheses plus a verification ladder; the modular adapter claim is coherent but untested.","tokens_in":27655,"tokens_out":591,"would_cite":true,"duration_ms":5988,"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":"Generative AI for metal–insulator materials should invent switchable phase pairs, not just stable crystals, under severe label scarcity.","keywords":["metal–insulator transition","generative AI","physics-adapted materials discovery","quantum materials","crystal diffusion models","low-rank adaptation","transition-metal oxides and chalcogenides","phase-transition hypotheses"],"falsifier":"Run the modular generator with and without mechanism-prior adapters on the same chemical space, then measure enrichment: if adapters do not raise the fraction of candidates that pass quantitative gates on phase-pair energy separation, electronic contrast (gap or N(EF) change), soft-mode or pathway connection, and accessible control relative to a generic generator baseline, the central claim fails.","tokens_in":27627,"feed_emoji":"⚡","tokens_out":678,"duration_ms":5224,"temperature":0.7,"pith_summary":"This Perspective argues that discovering metal–insulator-transition (MIT) materials is a different problem from ordinary crystal generation. An MIT is not a property of one structure; it is a relation between at least two electronically distinct phases linked by an accessible control parameter such as temperature, strain, or field. Verified examples are scarce—on the order of dozens of clear positives in curated sets—so models cannot be trained directly on the MIT class. The authors therefore propose treating generated structures as mechanism-informed transition hypotheses and steering a pretrained crystal generator with lightweight domain and physics-prior adapters, then filtering candidates through a staged verification ladder of thermodynamic access, phase competition, electronic contrast, pathway plausibility, and device-relevant control. If that program works, generative models become proposal engines for rare switchable electronic matter rather than black-box structure inventers.","feed_headline":"AI should invent switchable phase pairs, not just crystals","feed_subtitle":"Under scarce MIT labels, generators need mechanism priors and a verification ladder before any discovery claim.","key_machinery":"A modular physics-adapted generative stack: a pretrained crystal diffusion model supplies broad inorganic structure priors; domain adapters bias toward transition-metal oxides and chalcogenides; mechanism-prior adapters (implemented via parameter-efficient methods such as low-rank adaptation) steer sampling using scalable physics proxies for correlation, dimerization, charge/orbital order, excitonic or density-wave routes; and a verification ladder tests phase competition, electronic contrast, pathway plausibility, and control accessibility before any discovery claim.","core_discovery":"The paper’s central claim is that generative AI for MIT discovery should target mechanism-informed phase-transition hypotheses—candidate crystals plus competing phases, electronic contrast, a switching coordinate, and a realistic control route—rather than valid or stable crystals alone, and that a modular physics-adapted generator plus staged verification can do this under severe label scarcity.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["AI should target MIT phase pairs with mechanisms not crystals alone","Physics-adapted generators need verification under MIT label scarcity","Invent switchable phases via priors and staged checks not structures","Mechanism-informed hypotheses beat crystal search for MIT AI discovery","Modular framework guides credible MIT candidates when labels are scarce"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The load-bearing premise is that physics-proxy adapters trained on larger oxide and chalcogenide pools—not on the tiny set of verified MIT compounds—will actually enrich the fraction of generated candidates that survive phase-pair and switching tests.","fun_headline_variants_meta":{"raw":{"variants":["AI should target MIT phase pairs with mechanisms not crystals alone","Physics-adapted generators need verification under MIT label scarcity","Invent switchable phases via priors and staged checks not structures","Mechanism-informed hypotheses beat crystal search for MIT AI discovery","Modular framework guides credible MIT candidates when labels are scarce"]},"model":"grok-4.5","effort":"low","cost_usd":0.004652,"raw_usage":{"total_tokens":1252,"prompt_tokens":619,"num_sources_used":0,"completion_tokens":62,"cost_in_usd_ticks":46520000,"prompt_tokens_details":{"text_tokens":619,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":571,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":619,"tokens_out":62,"duration_ms":4779,"temperature":1.0,"reasoning_tokens":571,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T01:24:12.532791+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the modular generator with and without mechanism-prior adapters on the same chemical space, then measure enrichment: if adapters do not raise the fraction of candidates that pass quantitative gates on phase-pair energy separation, electronic contrast (gap or N(EF) change), soft-mode or pathway connection, and accessible control relative to a generic generator baseline, the central claim fails.","supporting_citations":[],"review_version":1}