{"id":"8ca0906f-be77-423e-9b97-589dbf023fcb","arxiv_id":"2506.16680","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"ASCC and its partially linearized variant PLASCC produce nearly starting-point-independent excitation energies, converting charge transfer errors of multiple eV from CIS and TD-DFT into errors around 0.1 eV.","lead":"Aufbau suppressed coupled cluster (ASCC), normally started from an expensive state-specific orbital optimization, gives nearly the same excited state energies when started from cheap linear response calculations like CIS or TD-DFT. This makes accurate charge transfer excitation energies potentially available at much lower cost than current state-specific methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Selection bias from excluding non-converged starting points makes the quantitative 'starting-point insensitivity' claim optimistic; the charge-transfer headline rests on seven states.","rationale":"The reader's weakest_assumption identifies the same selection-bias concern: the exclusion rule in Section III.a makes the reported MUEs conditional on the converged subset, and the charge-transfer conclusion relies on a very small sample. My reading confirms this is the most load-bearing issue because the paper's headline is explicitly about starting-point insensitivity, and the evidence for that claim is strongest when all starting points converge. The paper deserves credit for publishing the raw data table and the explicit list of excluded states, which makes the proposed re-analysis straightforward. I do not see an internal inconsistency in the ASCC equations or an error that would invalidate the method; the concern is about the strength of the empirical claim relative to the data. The existing CONDITIONAL verdict is appropriate: the method appears promising and well supported for valence and Rydberg states, but the strong '20x improvement' and 'highly resilient' phrasing should be tempered until the selection bias is addressed. I would not change the reader's verdict; if anything, the concreteness of the SI data makes it easy for the authors to run the proposed robustness checks and upgrade the claim.","tokens_in":25371,"tokens_out":5470,"duration_ms":59571,"concrete_test":"Recompute PLASCC-from-TDDFT and PLASCC-from-ESMF MUEs on a fixed test set without excluding non-converged cases, using the raw data in the SI table. For each non-converged state, assign either the starting method's own error (conservative) or a large-error penalty (e.g., 2x the starting error). If the TD-DFT-based PLASCC MUE rises above about 0.2 eV, or if the ESMF advantage exceeds about 0.1 eV, the starting-point-insensitivity claim is conditional on convergence. Separately, run PLASCC from TD-DFT on an expanded charge-transfer benchmark (e.g., the full Szalay set if it contains more than seven states, or a larger CT benchmark such as the one in Mester and Kállay, JCTC 2022) and report MUE over at least 20 states, again including convergence failures in the denominator. If the MUE ratio to TD-DFT remains near 20x on the larger set, the CT claim would be substantiated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ASCC/PLASCC is quite insensitive to its starting point, with the strongest quantitative statement being the 20-fold reduction in TD-DFT charge-transfer MUE. The most load-bearing weakness is the test-set construction in Section III.a: states are excluded when two or more starting points fail to converge or have more than two large CSFs. This means the MUEs reported in Figures 2 and 3 are not computed on a fixed set of states. For example, in the SI (Section a), PLASCC from ωB97X-V omits five states that failed to converge (formaldehyde 1 1B2, formamide 3 1A′, isobutene 1 1B1, thioacetone 2 1A1, cyclopropenone 3 1B2), while PLASCC from ESMF converges for those states. If convergence failure correlates with precisely the states where the starting point is poor—e.g., because the starting point lands on the wrong root or has strong multi-reference character—then the reported 'near independence' of starting point is partly an artifact of discarding the failures. The charge-transfer conclusion is even more fragile: it depends on seven states, and two CIS starting points fail to converge, so the CIS-based MUE is over only five states. The 20x improvement from TD-DFT is a ratio of small-sample MUEs and could be driven by one or two large TD-DFT errors. The paper discloses these exclusions, but the central claim—particularly the abstract's 'highly resilient'—goes beyond what the conditional data support.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates whether Aufbau suppressed coupled cluster (ASCC) and its partially linearized variant (PLASCC) can serve as post-linear-response corrections for electronically excited states, starting from CIS, TD-DFT/ωB97X-V, EOM-CCSD, or ESMF wave functions. The authors construct truncated starting points via singular value decompositions of transition density matrices or natural-orbital-based procedures, then solve ASCC/PLASCC equations. They report that on a reduced set of 138 valence and Rydberg excitations from the QUEST databases plus seven charge-transfer states, final ASCC/PLASCC excitation-energy errors are nearly independent of the starting point. In particular, PLASCC from TD-DFT reduces the charge-transfer MUE by a factor of about 20, from multi-electron-volt errors to ~0.1 eV. The paper concludes that ASCC's singles operator provides its own orbital relaxation, so the more expensive ESMF orbital optimization may not be necessary when ASCC is used.","tokens_in":25689,"tokens_out":6483,"duration_ms":63708,"significance":"If the central claim holds, the paper describes a practically valuable protocol: a high-accuracy, state-specific coupled-cluster correction that can be initialized from cheap linear-response methods. The study is built on well-established external benchmarks (QUEST and the Szalay charge-transfer set), uses a fixed SVD truncation threshold rather than parameters fitted to the target energies, and provides a large raw-data table in the supplementary information. The starting-point-insensitivity claim is falsifiable and is supported by a broad set of states. The main weakness is that the headline quantitative conclusions rely on a reduced test set with nonrandom exclusions and a very small charge-transfer sample, so the reported resilience is partly conditional on convergence outcomes.","major_comments":[{"comment":"The central starting-point-insensitivity claim is measured on a reduced test set that excludes states where two or more starting points fail to converge or have more than two large CSFs. Because poor starting points fail to converge more often, the reported MUEs are conditional on the converged subset and are not computed on a common fixed set of states. This is not merely a hypothetical concern: the SI lists five ωB97X-V-based PLASCC failures (formaldehyde 1 1B2, formamide 3 1A′, isobutene 1 1B1, thioacetone 2 1A1, cyclopropenone 3 1B2) for which ESMF-based PLASCC converges with errors near 0.0–0.1 eV. Dropping those states from the ωB97X-V statistics can artificially lower its MUE and make it look closer to ESMF than it actually is. I request a sensitivity analysis: for each starting point, report the MUE on the maximal set where at least one starting point converges, with missing entries handled by a stated rule (for example, using the ESMF-based result as a proxy, or excluding the state from all methods so that all MUEs are on an identical subset). Without such an analysis, the headline 'insensitivity' claim is not fully supported by the data as presented.","section":"III.a, III.c, SI a"},{"comment":"The charge-transfer conclusion rests on only seven states, and two of the CIS-based PLASCC calculations fail to converge (ammonia-oxygendifluoride 4 1A′ and 3,5-difluoro-penta-2,4-dienamine 1 1A″). The CIS-based MUE is therefore computed over five states, and the '20 times smaller' MUE ratio for TD-DFT is a ratio of small-sample statistics that could be dominated by one or two large starting-point errors. The paper does disclose these exclusions, but the abstract's characterization of the charge-transfer improvement as 'especially stark' goes beyond what seven states can robustly support. Please provide a per-state error table for the charge-transfer set, explicitly identify which states contribute to each MUE, and include a stability analysis (for example, leave-one-out or bootstrap) to show that the reported improvement factors are not artifacts of a single outlier.","section":"III.d, Figure 5"}],"minor_comments":[{"comment":"The sentence comparing starting-point MUEs from Figure 1 (which includes the seven charge-transfer states) to the post-ASCC/PLASCC MUEs from Figure 2 (which excludes them) is an apples-to-oranges comparison. Since TD-DFT's large charge-transfer errors inflate its Figure 1 MUE, this sentence overstates the improvement. The fair comparison is already provided in Figure 4, and the text should either refer to Figure 4 or recompute the starting-point MUEs on the same 138-state subset.","section":"III.c"},{"comment":"The EOM-CCSD starting-point construction (interleaving singular vectors with natural orbitals followed by Gram–Schmidt) is more ad hoc than the CIS/TD-DFT construction, and the paper does not provide the same numerical validation as it does for the TD-DFT case (where the construction agrees with the CIS-like approach to within 0.01 eV). A brief test or justification of the EOM-CCSD construction would increase confidence that this starting-point choice does not introduce an uncontrolled bias.","section":"II.d"},{"comment":"The raw-data table is extremely wide and difficult to parse in printed form. I recommend also providing the data as a machine-readable file (CSV or similar) to allow readers to recompute the statistics and test alternative inclusion rules.","section":"SI b"},{"comment":"The figure caption should state explicitly that the MUE insets are computed over the converged states only, and that the numbers of states differ across starting points (the insets do show this, but the caption does not draw attention to it).","section":"III.d, Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is clearly within scope and the underlying ASCC/PLASCC methodology is novel and potentially useful. The main concern is not circularity or parameter fitting—the method uses a fixed truncation threshold and external benchmarks—but rather selection bias in the test-set construction and the small charge-transfer sample. The authors have been transparent about the exclusions, which is commendable, but the quantitative claims need to be made robust to those exclusions before publication. I would not reject the paper; the required revisions are analytical rather than conceptual."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a useful, believable paper. The new thing is not the ASCC ansatz itself but the demonstration that it can be initialized from CIS, TD-DFT, or EOM-CCSD and produce excited-state energies nearly as good as from ESMF. The practical takeaway—that ESMF's orbital-optimized start may not be worth the trouble—is worth having.\n\nThe paper does several things well. The benchmarks are external (QUEST plus the Szalay CT set). The SVD truncation threshold is a fixed methodological choice, not a fitted parameter. The raw data are in the SI, so the results are checkable. The EOM-CCSD starting-point construction, using the TDM plus natural orbitals, is a genuine technique. For valence and Rydberg states, the pattern is consistent: PLASCC converts starting points with MUEs of 0.87, 0.72, 0.36, and 0.11 eV into results with MUEs at or below 0.12 eV. That is a strong result.\n\nThe soft spot is the selection bias the reader identified, and it is real. The reduced test set excludes states where two or more starting points fail to converge or have too many CSFs. PLASCC from ωB97X-V fails on five states that converge from ESMF, so the reported near-independence is conditional on the states each starting point can handle. The authors disclose this, and it does not sink the valence/Rydberg claim: the ESMF MUE is computed over all 138 states and still lands near 0.1 eV. Even if the five ωB97X-V failures had larger errors, the conclusion that PLASCC is much better than TD-DFT would survive. The charge-transfer evidence is thinner. Seven states, two CIS failures, so the CIS-based MUE is over five states. The 20x TD-DFT improvement uses the full seven-state set and is less suspect, but it is a small-sample headline. The abstract's 'highly resilient' goes a bit further than the conditional data support.\n\nWho is this for? Quantum chemists who use TD-DFT or CIS for excited states and want a low-cost path to better CT energies. It deserves a serious referee. I would send it out with a request for a robustness table on the intersection of states where all starting points converge, and ideally a larger CT set with more CIS failures included. As it stands, the main qualitative conclusions are sound.","headline":"A solid, incremental method paper showing ASCC/PLASCC is mostly starting-point insensitive; the headline is slightly ahead of the conditional data.","tokens_in":26207,"tokens_out":4013,"would_cite":true,"duration_ms":41911,"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":"The paper claims that Aufbau suppressed coupled cluster (ASCC) is nearly insensitive to which linear-response method seeds it, so a cheap CIS or TD-DFT calculation can be refined into energies close to those from the best excited-state…","keywords":["Aufbau suppressed coupled cluster","post-linear-response refinement","excited-state coupled cluster","charge transfer states","orbital relaxation","linear response","TD-DFT","EOM-CCSD"],"falsifier":"Recompute the ASCC and PLASCC errors on the full set of states without dropping non-converged cases, treating non-convergence as a failure; a specific check is the two charge-transfer CIS states where PLASCC did not converge (ammonia-oxygendifluoride 4 1A' and 3,5-difluoro-penta-2,4-dienamine 1 1A\"). If including those states raises the TD-DFT-seeded PLASCC MUE well above 0.1 eV, or widens the spread across starting points, the central resilience claim is contradicted. If a continuation method converges those states and they land near the benchmark, the claim is confirmed.","tokens_in":25182,"feed_emoji":"⚛️","tokens_out":8117,"duration_ms":77546,"temperature":0.7,"pith_summary":"This paper sets out to establish that Aufbau suppressed coupled cluster (ASCC), when seeded with a starting state from any of the common linear-response methods, corrects that starting state so well that the final accuracy barely depends on which method produced the seed. The central evidence is a 138-state valence/Rydberg set plus seven charge-transfer states: PLASCC, the partially linearized form of ASCC, takes CIS, TD-DFT, EOM-CCSD, and ESMF starting points with mean unsigned errors of 0.87, 0.72, 0.36, and 0.11 eV and returns results with MUEs of 0.12 eV or less. In charge-transfer states the effect is largest: PLASCC from TD-DFT cuts the MUE by about a factor of 20, from well over 2 eV to roughly 0.1 eV. A sympathetic reader would care because it suggests that expensive state-by-state orbital optimizations such as ESMF may not be worth their cost when ASCC is available, and that simple linear-response calculations can serve as entry points to high-accuracy excited-state energetics.","feed_headline":"PLASCC turns multi-eV TD-DFT errors into ~0.1 eV","feed_subtitle":"Cheap linear-response starting points end up nearly as accurate as expensive orbital-relaxed ones for excited states.","key_machinery":"The load-bearing object is the ASCC wave function $|\\Psi_{\\mathrm{ASCC}}\\rangle = e^{-\\hat S^\\dagger} e^{\\hat T}|\\phi_0\\rangle$, in which $\\hat T$ contains excitation operators and $\\hat S^\\dagger$ contains de-excitations. Applied to the Aufbau determinant $|\\phi_0\\rangle$, the product $\\hat S^\\dagger \\hat T$ creates a second copy of the ground-state determinant with a controlled sign, so the Aufbau determinant can be suppressed or cancelled rather than simply occupied. PLASCC uses the same equations but drops a selected set of nonlinear diagrams, which prior perturbative analysis showed to be the balancing act that yields the best accuracy. The starting-point protocol is the second piece: for each linear-response method, the transition density matrix is rotated and truncated by singular value decomposition to one or two dominant configuration state functions, giving a compact reference for $\\hat T$. The exponentiated singles operator is what carries the argument, because it can build in orbital relaxations that the starting method lacks, which explains why CIS and TD-DFT seeds, despite missing most relaxation physics, end up almost as accurate as ESMF seeds.","core_discovery":"The paper's central claim, stated on its own terms, is that ASCC performs its own orbital relaxations through its singles operator, so the quality of the starting point matters much less than in typical post-linear-response schemes. The evidence is in the convergence of error distributions: regardless of whether ASCC or PLASCC is initialized from CIS, TD-DFT with a range-separated hybrid functional, EOM-CCSD, or ESMF, the final MUEs cluster near 0.1 eV for valence and Rydberg states, and the small remaining differences track the starting point's ability to help convergence more than its raw energy accuracy. For the seven charge-transfer states, TD-DFT starting points with errors larger than 2 eV are transformed into PLASCC results with an MUE about 20 times smaller, while the already-accurate EOM-CCSD starting points improve by more than a factor of three. The paper concludes that ASCC is best understood as a high-accuracy post-linear-response refinement method, with ESMF retaining only a very small accuracy edge that is probably not worth its added cost.","pith_inferences":["A natural, untested extension is to seed PLASCC from cheaper-than-EOM-CCSD methods such as CC2 or selectively expanded CIS; if the insensitivity holds, sub-0.1 eV accuracy might be reachable at reduced scaling.","The same orbital-relaxation argument suggests PLASCC could improve excited-state potential energy surfaces, where partial relaxation errors distort photochemistry; testing this along nuclear coordinates would show whether the insensitivity persists beyond vertical energies.","The winnowing step means the reported numbers condition on convergence; extending the code to two-CSF and three-CSF starting points, or counting non-convergence as an outcome, would show whether the resilience claim holds for the hardest states rather than only for the tractable ones.","If ASCC truly erases starting-point dependence, the practical distinction between state-specific and response-based excited-state methods blurs: expensive orbital optimization could become optional, with convergence control replacing accuracy as the main design problem."],"forward_implications":["PLASCC used after TD-DFT or CIS should give valence and Rydberg excitation energies with typical errors of a few tenths of an eV, close to the results obtained from EOM-CCSD or ESMF starting points.","For charge-transfer states, post-linear-response PLASCC reduces TD-DFT's mean unsigned error by roughly a factor of 20, and improves even EOM-CCSD's charge-transfer errors by more than a factor of three.","Because starting-point accuracy matters so little, users can choose the cheapest convenient linear-response seed without expecting a large accuracy penalty, provided the ASCC equations converge.","The practical case for ESMF's nonlinear orbital optimization weakens: its roughly 0.02 eV MUE advantage over linear-response starts may not justify its state-by-state cost.","ASCC's own orbital relaxation means that post-linear-response use should be treated as a general mode of ASCC, not a specialized correction."],"supporting_citations":[{"why":"Introduces Aufbau suppressed coupled cluster and the idea that its singles operator supplies state-specific orbital relaxation.","marker":"[17]"},{"why":"Defines the partially linearized PLASCC working equations and the prior filtering and benchmark protocol this paper adapts.","marker":"[18]"},{"why":"Supplies the small-molecule QUEST reference excitation energies used as benchmarks for valence and Rydberg states.","marker":"[90]"},{"why":"Supplies the medium-molecule QUEST reference excitation energies used as benchmarks for valence and Rydberg states.","marker":"[91]"},{"why":"Supplies the charge-transfer benchmark set whose EOM-CCSDT reference energies anchor the CT comparisons.","marker":"[41]"},{"why":"Provides the EOM-CCSD transition density matrix and natural-orbital machinery used to build the EOM-CCSD starting point.","marker":"[26]"},{"why":"Supplies the range-separated hybrid functional whose TD-DFT starting points and charge-transfer errors are the paper's central test case.","marker":"[96]"}],"fun_headline_variants":["ASCC self-relaxes, making starting point almost irrelevant","From multi-eV errors to 0.1 eV via ASCC orbital relaxation","Charge-transfer errors shrink 20-fold in PLASCC","ASCC rescues poor linear-response starting points"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim of starting-point insensitivity is measured only on the subset of states where at least two starting points converged; if the states that are hardest for a starting method are also the ones most likely to fail to converge, the reported insensitivity is optimistic.","fun_headline_variants_meta":{"raw":{"variants":["ASCC self-relaxes, making starting point almost irrelevant","From multi-eV errors to 0.1 eV via ASCC orbital relaxation","Charge-transfer errors shrink 20-fold in PLASCC","ASCC rescues poor linear-response starting points"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000589,"raw_usage":{"total_tokens":2735,"prompt_tokens":889,"completion_tokens":1846,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":1776}},"tokens_in":505,"tokens_out":1846,"duration_ms":13372,"temperature":1.0,"reasoning_tokens":1776,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:20:29.499725+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the ASCC and PLASCC errors on the full set of states without dropping non-converged cases, treating non-convergence as a failure; a specific check is the two charge-transfer CIS states where PLASCC did not converge (ammonia-oxygendifluoride 4 1A' and 3,5-difluoro-penta-2,4-dienamine 1 1A\"). If including those states raises the TD-DFT-seeded PLASCC MUE well above 0.1 eV, or widens the spread across starting points, the central resilience claim is contradicted. If a continuation method converges those states and they land near the benchmark, the claim is confirmed.","supporting_citations":[],"review_version":2}