{"id":"5054bd98-d8fc-4abe-9233-698e8a702cf7","arxiv_id":"2601.07937","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A transformer trained on short Lanczos-coefficient prefixes extrapolates coefficients and reconstructed observables more accurately than asymptotic fits, and transfers across system sizes in the two tested chaotic models.","lead":"Tests whether a transformer can forecast Lanczos coefficients—numbers that govern operator growth in quantum and classical dynamics—from a short prefix, reporting large error reductions over asymptotic fits on chaotic spin models. Also reports zero-shot transfer from 8-site to 12-site chains, though without code/data and with an unsupported integrable claim in the abstract.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation horizon equals training horizon (T=30 TFIM, T=100 XYZ); no test of prediction for n>T, so the 'deep Krylov space' extrapolation claim is unsupported.","rationale":"The reader's weakest_assumption concerns the accuracy of full-reorthogonalization Lanczos labels. That is a reasonable caution, but for these short sequences (T=30, T=100) full reorthogonalization is likely reliable; the more structurally significant gap is that the model is never evaluated at n greater than the training horizon. The paper's abstract and conclusion claim access to computationally prohibitive depths, yet every quantitative claim is within the horizon of the training sequences. This is not resolved by releasing code or adding error bars. The verdict remains CONDITIONAL, but the primary condition should be demonstrating genuine beyond-horizon prediction or explicitly retracting the depth-extrapolation framing.","tokens_in":16303,"tokens_out":7351,"duration_ms":76568,"concrete_test":"Generate 100 L=8 TFIM sequences to T=60 with full reorthogonalization; use the first 10 coefficients as prompt and the transformer trained on T=30 to autoregressively predict b_11..b_60; plot RMSE for n=31..60 against the same asymptotic fit. If the transformer error at n>30 is not consistently below baseline, the 'deep Krylov space' extrapolation claim is unsupported. If b_n cannot be computed reliably beyond 30 for L=8, state that the claim is untestable and remove the computational-prohibitive framing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central practical claim is that the transformer can probe operator dynamics 'deep in Krylov space' beyond where brute-force Lanczos is feasible. But the experiments never test prediction beyond the sequence lengths seen in training. In Sec. III, training sequences have fixed length T (T=100 for XYZ, T=30 for TFIM), and the loss in Eq. (24) trains next-step prediction for m from n_in to T−1. The model therefore observes complete coefficient sequences up to b_T during training. All reported test RMSEs (Figs. 3–6) and observable reconstructions (Fig. 5) are for n ≤ T, using the same T for test labels. Thus the 'extrapolation' is a continuation within the length distribution of the training data, not a forecast beyond the largest n ever observed. The L=12 transfer is a system-size extrapolation, not a depth extrapolation; the n-range is still 1..30 as in training. If the model fails for n>30, the practical-surrogate claim collapses even though the in-horizon RMSE improvement is real. This is the weakest load-bearing assumption in the central claim.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper frames the Lanczos coefficients {b_n} of operator growth as a causal time series and trains a masked decoder-only transformer to autoregressively predict future coefficient differences Δb_n from a short prefix. The model is benchmarked on the classical XYZ top (T=100) and on a chaotic transverse-field Ising chain (L=8, T=30, with an L=12 transfer), and it is compared against the standard asymptotic UOGH fit with an even-odd staggering term. Reported results include RMSE reductions in coefficient prediction, much smaller reconstruction errors in Krylov complexity K(t) and autocorrelation C(t), zero-shot transfer from L=8 to L=12, and attention-map/ablation analyses. The paper's central claim is that a transformer can serve as a practical surrogate for probing operator dynamics deep in Krylov space where brute-force Lanczos iteration is prohibitive.","tokens_in":16587,"tokens_out":10064,"duration_ms":113691,"significance":"If fully substantiated, the work would provide a useful ML-based tool for extending Lanczos sequences, and the zero-shot system-size transfer is a genuinely valuable feature. The paper is clearly written, the architecture is standard and reproducible from the hyperparameter table, and the attention ablation is a nice mechanistic addition. The main strengths are the direct comparison to the standard asymptotic baseline, the validation on two different chaotic models, and the transfer experiment. However, the support for the 'deep Krylov space' claim is incomplete: all evaluations stay within the sequence length T used in training, and the abstract claims an integrable-regime result that the body of the paper does not report. These issues are fixable with additional experiments or appropriately reworded claims, but they are load-bearing for the paper's central contribution as stated.","major_comments":[{"comment":"The evaluation horizon equals the training horizon. The loss in Eq. (24) trains next-token predictions for all m from n_in to T-1, so every target index m+1 <= T is present in the training sequences. All reported coefficient RMSEs (Figs. 3(b), 4(b), 6) and observable reconstructions (Fig. 5) are for n <= T. Thus the model is tested on continuation within the length distribution seen in training, not on forecasting coefficients for n > T. The L=12 transfer in Fig. 6 is a system-size extrapolation, not a depth extrapolation: the n range is still 1..30. Because the abstract and Secs. I and VI claim the model can probe 'deep in Krylov space' and act 'as a practical surrogate' where direct Lanczos is prohibitive, this is a central unsupported claim. I request an explicit test in which the training sequences are truncated at T_train < T_max and the model is evaluated for T_train < n <= T_max,","section":"Sec. III, Eq. (24); Sec. IV, Figs. 3-6"},{"comment":"The abstract states: 'The model also accurately extrapolates coefficients in integrable regimes, where no universal asymptotic fit exists.' The full text contains no integrable-regime experiment; Sec. VI lists testing the model on sub-linear asymptotics as a future direction ('Another interesting direction would be to test the model's ability to forecast Lanczos sequences that are not asymptotically linear'). This is an unsupported claim in the abstract and should be removed or backed by the missing experiments.","section":"Abstract; Sec. VI"},{"comment":"All ground-truth labels are generated by finite-precision Lanczos with full reorthogonalization, yet the introduction motivates the work by the numerical instability of the same procedure. No independent validation of the long-tail labels is reported (e.g., high-precision arithmetic, orthogonality residuals, or comparison with direct moment computations). If loss of orthogonality or finite-size saturation silently corrupts the labels for n near T (T=100 for XYZ, T=30 for TFIM), the RMSE improvements are measured against corrupted labels and the 'order-of-magnitude' claim would not reflect true coefficient extrapolation. Please add a validation of at least a subset of long sequences, reporting, for example, the norm of the reorthogonalization residual as a function of n.","section":"Sec. III, Eq. (4); Sec. IV"},{"comment":"The RMSE curves are point estimates averaged over N_T=100 test sequences without error bars, bootstrap confidence intervals, or per-sequence quantiles. The central quantitative claims ('order-of-magnitude reduction,' 'orders of magnitude smaller errors') are therefore not accompanied by any uncertainty estimate. Given the large reported separations this is not necessarily fatal, but for a quantitative ML benchmark in a physics journal the authors should report at least bootstrap CIs or the spread of per-sequence errors.","section":"Sec. IV, Eq. (27); Figs. 4-6"}],"minor_comments":[{"comment":"The text says 'Under UOGH, the raw coefficients grow without bound, whereas Δb_n approaches an n-independent constant in the asymptotic limit.' This is true for the d != 1 linear asymptote, but for the d=1 case in Eq. (9), b_n ~ α n / log n, so Δb_n ~ α / log n, which tends to 0 rather than a constant. Since the TFIM benchmark uses d=1, please qualify this statement.","section":"Sec. III, Eq. (13)"},{"comment":"The triangular support in the attention maps is a direct consequence of the hard causal mask M=-∞ in Eq. (20), not a learned property of the trained model. The meaningful learned patterns are the diagonal concentration, the checkerboard pattern, and the early-token attention. Please rephrase the sentence 'The observed triangular structure... provides a direct check that the causal constraint is indeed being enforced in the trained model' to avoid implying this is learned.","section":"Sec. V, Eq. (20), Fig. 7"},{"comment":"The abstract says the method achieves 'an order-of-magnitude reduction in error' for 'both classical and quantum chaotic systems.' In the classical XYZ case (Fig. 3(b)) the reported reduction is 'a few times,' not an order of magnitude. Please make the abstract's quantitative claim match the figures.","section":"Abstract"},{"comment":"No code or data availability statement is included. For reproducibility of a numerical ML benchmark, please indicate whether training/test data and code will be released.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"To the editor: I see no integrity concerns. The main issue is that the abstract and introduction overreach relative to the experiments: the integrable-regime claim is absent from the body, and the 'deep Krylov space' claim is not tested for n beyond the training horizon. Both are fixable with additional experiments or clearly stated limitations. The work is within the journal's scope and the core in-horizon comparison is likely sound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper trains a small transformer to autoregressively predict the next Lanczos coefficient from a prefix of ten, benchmarks it on the classical XYZ top and the transverse-field Ising model, and reports order-of-magnitude lower RMSE than the standard asymptotic fit. That in-horizon result is real and clean: the transformer beats the baseline on held-out sequences from two chaotic models, and the advantage grows with n. The zero-shot transfer from L=8 to L=12 is a genuine system-size extrapolation, and the attention/ablation analysis is a useful addition. These are legitimate new empirical results.\n\nThe central claim, though, is bigger than what is tested. Training sequences have fixed length T (100 for XYZ, 30 for TFIM), and the loss in Eq. (24) trains next-step prediction for every m from n_in to T−1. So the model has seen all coefficient indices up to T during training. At inference, every reported RMSE and observable reconstruction is for n ≤ T. The 'deep Krylov space' extrapolation is thus a continuation within the observed index range, not a forecast to depths beyond anything the model ever saw. The L=12 transfer is real, but it is system-size transfer within the same depth range. If the model fails for n>30 or n>100, the practical-surrogate story collapses. This is never tested, and it is the load-bearing assumption of the paper's motivation. The stress-test note has this right.\n\nThe abstract also contains a sentence claiming accurate extrapolation in integrable regimes, but the full text explicitly leaves that as future work and no integrable experiment appears anywhere. That sentence should be removed or supported. Minor but fixable issues: no code or data, no error bars on the RMSE curves, and only a single baseline.\n\nOverall, this is a solid in-horizon forecasting result with an overstated scope. The paper deserves peer review because the approach is promising and the in-horizon numbers are evidence, but I would expect major revision: add a true out-of-distribution depth test (e.g., train on T=30, evaluate at n=40–60 where ground truth is still computable), correct the abstract, and release artifacts. Without that, it stays a useful benchmark rather than a surrogate for beyond-Lanczos depth.","headline":"Solid in-horizon forecasting of Lanczos coefficients, but the 'deep Krylov space' claim is untested: the model never predicts beyond the largest index it saw in training.","tokens_in":17069,"tokens_out":3413,"would_cite":false,"duration_ms":32564,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A transformer trained on short Lanczos prefixes extrapolates operator-growth coefficients with roughly ten times lower error than asymptotic fits, and transfers without retraining from small to larger systems.","keywords":["Lanczos coefficients","Krylov complexity","operator growth","transformer","quantum many-body chaos","sequence extrapolation","universal operator growth hypothesis","autocorrelation function"],"falsifier":"Recompute the long-tail coefficients for the same L=8 Hamiltonians with arbitrary-precision arithmetic or via an independent moment-based recursion, then rerun the RMSE comparison; if the transformer's order-of-magnitude advantage over the asymptotic fit disappears or shrinks substantially, the central claim fails. Also check whether the L=12 'true' sequences used for zero-shot evaluation saturate before n=30—if they do, the benchmark is against truncated labels.","tokens_in":16163,"feed_emoji":"⚛️","tokens_out":6019,"duration_ms":60288,"temperature":0.7,"pith_summary":"The paper claims that Lanczos coefficients—the hopping amplitudes governing operator growth in quantum and classical dynamics—form a causal sequence with long-range, history-dependent structure that standard asymptotic fits miss. It trains a small transformer on short prefixes of these coefficients and uses it to predict the long tail autoregressively. Across chaotic spin models, the learned extrapolation is roughly an order of magnitude more accurate than the standard linear-asymptotic fit, and the same model trained on one system size transfers to a larger size without retraining. If correct, this gives a practical surrogate for probing operator dynamics deep in Krylov space, where direct Lanczos iteration is numerically unstable and memory-bound. The paper also identifies which parts of the coefficient history matter most by inspecting learned attention weights and ablating them.","feed_headline":"Transformer cuts Lanczos extrapolation error tenfold","feed_subtitle":"Trained on small chains, it forecasts coefficients on larger ones without retraining—opening deep Krylov space.","key_machinery":"The central object is a decoder-only transformer with masked multi-head self-attention—a sequence model that assigns learned weights to all past coefficients when predicting the next one—applied to the causal sequence of coefficient differences Δb_n = b_n − b_{n−1}. Predicting differences rather than raw b_n stabilizes the target scale, and the causal mask enforces that predictions depend only on earlier coefficients. Attention then captures long-range correlations that the iterative Lanczos recursion implicitly propagates.","core_discovery":"The paper claims that Lanczos coefficients are not just an asymptotic ramp; they carry subleading, history-dependent structure—even-odd staggering, slow deviations, and early-coefficient anchors—that standard fits miss. A small decoder-only transformer, trained on the increment sequence Δb_n = b_n − b_{n−1} from thousands of Hamiltonian instances, predicts coefficients for n>10 autoregressively. On unseen instances of the classical XYZ spin top and a chaotic transverse-field Ising chain with L=8, the transformer's coefficient RMSE is consistently an order of magnitude smaller than the universal-operator-growth-hypothesis asymptotic fit with staggering; reconstructed Krylov complexity and aut","pith_inferences":["One passage to weigh: the abstract credits the model with extrapolating integrable regimes, but the body's Sec. VI lists that as an open direction; a reader should treat the integrable-regime claim as stated but unsupported in the current text.","The success of the increment-sequence representation suggests that other spectral quantities derived from autocorrelation moments—not just Lanczos coefficients—may be forecastable by the same autoregressive scheme.","If attention weights on early coefficients act as global anchors, then an even shorter input prefix plus system-size metadata might suffice for transfer; this is a testable ablation.","A practical deployment would train on small chains and then predict coefficients for a Hamiltonian one cannot iterate; verifying against exactly solvable or sparse Hamiltonians at intermediate sizes would sharpen the zero-shot claim."],"forward_implications":["Coefficient extrapolation can be treated as sequence prediction: a model trained on short prefixes reproduces long Lanczos tails with order-of-magnitude lower RMSE than linear asymptotic fits on chaotic systems.","Zero-shot transfer from L=8 to L=12 means training data can be collected at accessible sizes and deployed to larger systems where direct Lanczos iteration is prohibitively expensive.","Observable-level reconstruction—Krylov complexity and autocorrelation—inherits the extrapolation accuracy, with errors orders of magnitude smaller at late times.","Attention ablations show that removing long-range context, parity information, or the first few coefficients degrades accuracy by an order of magnitude, so Lanczos sequences should be modeled as long-range causal sequences rather than fitted to a local asymptotic form."],"fun_headline_variants":["Transformer cuts Lanczos coefficient error tenfold","AI extrapolates Lanczos coefficients across system sizes","Transformer beats asymptotic fits in Krylov space","Lanczos coefficients learned by transformer from short prefixes"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The full-reorthogonalization Lanczos output used as ground truth is exact over the entire horizon (T=100 for XYZ, T=30 for TFIM); if loss of orthogonality or finite-size saturation corrupts the long-tail labels, the reported error reductions are measured against corrupted targets.","fun_headline_variants_meta":{"raw":{"variants":["Transformer cuts Lanczos coefficient error tenfold","AI extrapolates Lanczos coefficients across system sizes","Transformer beats asymptotic fits in Krylov space","Lanczos coefficients learned by transformer from short prefixes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000151,"raw_usage":{"total_tokens":1048,"prompt_tokens":763,"completion_tokens":285,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":226}},"tokens_in":507,"tokens_out":285,"duration_ms":3583,"temperature":1.0,"reasoning_tokens":226,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T10:57:56.934833+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the long-tail coefficients for the same L=8 Hamiltonians with arbitrary-precision arithmetic or via an independent moment-based recursion, then rerun the RMSE comparison; if the transformer's order-of-magnitude advantage over the asymptotic fit disappears or shrinks substantially, the central claim fails. Also check whether the L=12 'true' sequences used for zero-shot evaluation saturate before n=30—if they do, the benchmark is against truncated labels.","supporting_citations":[],"review_version":1}