{"id":"e519335d-7d3c-42be-8e65-847280be8c65","arxiv_id":"2608.04593","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A new pruning criterion for Echo State Networks, based on neuron participation in the dominant modes of a trajectory-averaged Jacobian Gramian, removes 20% of reservoir neurons with no loss in forecasting accuracy on five benchmarks.","lead":"This paper introduces Dynamical Mode Pruning, a one-shot method that removes Echo State Network neurons with the least influence on the reservoir's dominant input-driven transition modes. The authors show the pruned reservoirs match or beat the original networks on five forecasting benchmarks while cutting recurrent parameters by 36%.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"One-shot Gramian validity is the load-bearing fragility: DMP's scores are computed from the full reservoir before pruning, and the paper itself concedes the Gramian no longer describes the reduced system.","rationale":"The CONDITIONAL verdict is appropriate. The paper's controlled matched comparison, its honest reporting of runtime and p-value limitations, and its explicit limitation section are creditworthy. The one-shot Gramian issue is not by itself a rejection: the empirical results are consistent with DMP working at 20% pruning, and the authors flag the approximation. However, the central claim's mechanism—removing low-impact units identified by dynamical influence—depends on the full-reservoir Gramian ranking remaining informative for the reduced system. The proposed re-estimation check would directly settle whether the approximation is the source of the reported gains or a tolerated error. If re-estimation makes no material difference, the one-shot claim is validated; if it improves results, the paper should either iterate the pruning procedure or qualify the one-shot framing. The reader's weakest assumption identifies the same load-bearing concern, and the concrete test targets exactly that assumption.","tokens_in":16622,"tokens_out":9066,"duration_ms":119619,"concrete_test":"Run the matched N=1000→800 protocol on the same ten paired seeds but add a one-round re-estimation arm: after pruning, drive the reduced ESN with the same washout and training input, recompute G_I = (1/T)Σ J_I(t)ᵀJ_I(t) and the resulting scores η_I, then (i) measure the subspace distance between the top-r eigenspace of the original G and the corresponding projection of the retained coordinates, and (ii) retrain the readout using the re-ranked reservoir. If re-estimation changes the retained set or reduces mean NRMSE by more than the standard error on any dataset, the one-shot assumption is falsified at the tested operating point; if the re-estimated and one-shot results are indistinguishable, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"DMP's central claim requires that the neuron scores computed from the original reservoir remain informative after pruning. In Algorithm 1, scores η_i are derived from the trajectory-averaged Gramian G = (1/T)Σ J(t)ᵀJ(t) of the full N=1000 reservoir, and the reduced reservoir keeps the principal submatrix W[I,I]. But the reduced Jacobian J_I(t) = (1−a)I + a D_I(t)W[I,I] is not the principal submatrix of J(t): after removing neurons, the pre-activations of retained neurons lose the contributions W[I,I^c]x_{I^c}(t), so both D_I(t) and the retained trajectory differ. The paper explicitly acknowledges this in Section 5.1: 'after neurons are removed, the Gramian computed from the original reservoir no longer exactly describes the reduced system,' and the 30% pruning results degrade. At the operating point of 20% pruning (N=1000→800), the claim that DMP removes low-impact units requires that the dominant modes of the original Gramian remain dominant for the pruned reservoir. No direct evidence is provided for this stability; readout refitting could absorb miscalibration without validating the dynamical ranking. This is the load-bearing fragility of the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Dynamical Mode Pruning (DMP), a one-shot pruning method for Echo State Networks that ranks reservoir neurons by their participation in dominant modes of a trajectory-averaged Jacobian Gramian. The method computes J(t) = (1-a)I + aD(t)W along a teacher-forced trajectory, forms G = (1/T)∑J(t)^T J(t), retains the top-r eigenmodes up to an energy threshold tau, and assigns each neuron a score from the weighted squared eigenvector components. The lowest-scoring neurons are removed, and only the readout is refitted; an optional spectral-radius rescaling is treated as an ablation. The paper reports benchmark comparisons on Mackey–Glass and four real-world time series, including a controlled matched evaluation at N=1000→800 over ten paired seeds, where DMP has the lowest mean NRMSE among the compared methods and yields a 35.7% model-size reduction with roughly 1.9–2.0x inference speedup. The limitations section acknowledges the offline cost and the fact that the original Gramian no longer exactly describes the reduced reservoir after pruning.","tokens_in":16826,"tokens_out":5333,"duration_ms":61682,"significance":"If the central claim holds, DMP would offer a principled, data-dependent alternative to static connectivity or activation-based pruning for ESNs, with a one-shot criterion and negligible inference-time overhead. The paper has notable strengths: the Jacobian derivation in Eq. (4) is correct for the leaky ESN, the controlled matched evaluation pairs DMP and random pruning on the same reservoir and data split, the code is publicly available, and the limitations section is unusually candid about the one-shot approximation and the 30%-pruning degradation. However, the empirical evidence that the dynamical score beats matched random pruning or simple static criteria is currently weak on several real-world datasets, and the load-bearing one-shot Gramian assumption is acknowledged but not directly tested. The significance of the paper therefore depends on additional validation rather than on the theoretical construction alone.","major_comments":[{"comment":"The controlled matched evaluation does not support a robust advantage of DMP over matched random pruning on four of the five datasets. Table S2 reports mean DMP-minus-random NRMSE differences of -0.0026 (Electricity, 6/10 wins), -0.0102 (Temperature, 6/10), -0.0001 (Solar, 5/10), and -0.0001 (Wind, 5/10); only Mackey–Glass shows a clear separation with 9/10 wins. Because the paper's stated contribution is that trajectory-dependent dynamical influence is more informative than an arbitrary pruning mask, this evidence is load-bearing. Please report paired confidence intervals and win/loss statistics for the DMP-versus-random comparison, and either soften the claim to 'preserves' for Solar and Wind or provide additional evidence for improvement on those datasets.","section":"Table S2 / Section 4"},{"comment":"The one-shot Gramian assumption is acknowledged in Section 5.1 but is not directly validated. After pruning, the reduced Jacobian is J_I(t) = (1-a)I + aD_I(t)W[I,I], which is not the principal submatrix of the original J(t) because the retained pre-activations lose the term W[I,I^c]x_{I^c}(t). The paper states that the original Gramian 'no longer exactly describes the reduced system' but provides no evidence that the dominant modes of the original Gramian remain dominant for the reduced reservoir. Please add an experiment that measures, at each pruning ratio, the subspace distance between the top-r eigenspace of G and the Gramian of the reduced reservoir, or the rank correlation between neuron scores computed before and after pruning. This would directly test the central assumption and would also help interpret the degradation observed at 30% pruning.","section":"Section 5.1 / Algorithm 1"},{"comment":"The expanded comparison with static and simple baselines weakens the claim that DMP outperforms other pruning criteria. Under the uniformly rescaled protocol in Table S1, DMP+ρ has the lowest mean NRMSE only on Temperature; activation-variance+ρ is best on Mackey–Glass, magnitude+ρ is best on Electricity and Wind, and Random+ρ is best on Solar. Because the main unscaled DMP configuration is not included in this table, the comparison is not a complete matched test of DMP against static criteria. Please either include the unscaled DMP (and unscaled baseline variants) in the matched protocol with paired statistics, or explicitly restrict the paper's conclusion to 'DMP improves or preserves accuracy relative to the full ESN' rather than claiming superiority over static structural importance measures.","section":"Table S1 / Section A.1"}],"minor_comments":[{"comment":"The notation 'eλk' is not defined and is difficult to read; please use a standard symbol such as λ̂_k and state explicitly that it is the eigenvalue normalized by the trace of G.","section":"Equations (11)–(13)"},{"comment":"For Wind, the Base ESN mean is 19.80 with a standard deviation of 42.69 and a confidence interval reported as [0.0000, 72.8170], so the headline 89.8% mean reduction is driven by a small number of seeds. Consider reporting geometric means, log-scaled confidence intervals, or per-seed values to make the variability transparent.","section":"Table 2"},{"comment":"The caption of Table 1 explains that DMP runtime and memory include the dense Jacobian–Gramian pipeline and are not comparable to Table 4, but the main text should repeat this caveat at the first mention of the 'Runtime' and 'Peak Offline Memory' columns so that readers do not interpret DMP's 1836 MB peak usage as the cost of the final method.","section":"Table 1 / Section 5"},{"comment":"The sentence suggesting that a 'large change in spectral radius after pruning' or a 'strong shift in the Gramian eigenvalue spectrum' can indicate one-shot unreliability is not operationalized; please specify a concrete diagnostic procedure or threshold.","section":"Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is positioned as a general advance in reservoir pruning, but the controlled evidence mainly supports the weaker claim that DMP is a safe 20% pruning heuristic rather than that the dynamical criterion substantially outperforms random or simple static criteria on real-world data. The authors should be encouraged to add the proposed Gramian-stability experiment and paired statistical summaries; if those results confirm the current weak separation, the claims should be narrowed accordingly. I also note several future-dated references (e.g., 2026) and a self-citation; these should be checked for editorial compliance before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. DMP is a genuinely new construction for pruning fixed ESN reservoirs—trajectory-averaged Jacobian Gramian plus eigenvector-participation scores—and the evaluation is more honest than most pruning papers: they compare against matched random pruning, a fresh smaller ESN, spectral rescaling ablations, and they report seed-level win counts. At 20% pruning it delivers what the abstract claims: lower mean NRMSE than the full ESN on all five datasets, ~1.9x inference speedup, and 35.7% size reduction. The Jacobian derivation (Eq. 4) is correct for the leaky ESN, the Gramian construction is clean, and using participation in dominant modes is a sensible way to score neurons.\n\nThe soft spots are real but not fatal. The one-shot Gramian assumption is the load-bearing fragility, and the paper says so itself (Section 5.1): after pruning, the original Gramian no longer describes the reduced system. The reduced Jacobian is not the principal submatrix of the original because the retained neurons lose the cross terms W[I,I^c]x_{I^c}; the scores are computed once and never re-estimated. The 30% pruning degradation is consistent with this mismatch. The paper flags this and suggests diagnostics, but it doesn't provide direct evidence at the 20% operating point that the dominant modes survive. Readout refitting can absorb some miscalibration, which may explain why the forecasting numbers still hold up.\n\nThe second soft spot is the timing inconsistency. Table 1 reports DMP runtime of 370–570 seconds and peak memory in GB; Table 4 says offline pruning time is 0.13 seconds. The paper says these measure different scopes, but the discrepancy is two orders of magnitude and will puzzle a careful reader. They should reconcile it or give the streamlined implementation explicitly.\n\nThird, the claimed advantage over static importance is dataset-dependent. In the supplementary matched comparison, DMP+rho is best on only one of five datasets, and paired wins over random pruning on Solar and Wind are 5/10—chance. The authors acknowledge this. The clean separation on Mackey-Glass plus smaller margins on Electricity and Temperature is real evidence, but it's not a sweeping win.\n\nWho gets value? Reservoir computing practitioners and anyone who wants to see a pruning paper that actually runs the right baselines. It is a solid incremental contribution with honest reporting and a code link. A serious referee should engage with it. My revision list: reconcile the timing numbers, add a stability check of the Gramian modes after pruning (even a small experiment comparing rankings before/after one-step re-estimation), and temper the abstract's claim about dynamical influence beyond static importance. I'd send to peer review rather than desk reject.","headline":"DMP's Gramian-based pruning is a new and mostly honest incremental contribution, but the one-shot assumption and the over-promoted dynamical edge need scrutiny before publication.","tokens_in":17454,"tokens_out":3570,"would_cite":true,"duration_ms":39912,"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":"The paper proposes a one-shot, data-driven criterion that ranks echo-state reservoir neurons by their contribution to dominant trajectory-dependent transition modes, and shows that removing the lowest-ranked 20% preserves or improves…","keywords":["echo state networks","reservoir pruning","Jacobian Gramian","dynamical modes","time-series forecasting","structured pruning","chaotic systems","model compression"],"falsifier":"A direct check: after pruning at 30%, recompute the trajectory-averaged Jacobian Gramian for the pruned reservoir and compare its dominant subspace with the original; if the principal-angle overlap is small, and re-ranking with the new Gramian changes the retained set and improves NRMSE, then the one-shot assumption fails.","tokens_in":16334,"feed_emoji":"⚙️","tokens_out":7360,"duration_ms":73956,"temperature":0.7,"pith_summary":"The paper proposes Dynamical Mode Pruning (DMP), a one-shot method for compressing Echo State Networks by removing reservoir neurons that contribute least to the input-driven transition structure. DMP builds a trajectory-averaged Jacobian Gramian from the driven reservoir, keeps the dominant eigendirections, and scores each neuron by how strongly its coordinate participates in those directions. Pruned networks retrain only the linear readout, leaving recurrent weights fixed. In tests on chaotic Mackey-Glass and four real-world time series, pruning from 1,000 to 800 neurons lowers mean NRMSE relative to the full ESN on every dataset, cuts storage by 35.7%, and speeds up inference by 1.90 to 1.97 times. The paper's claim is that dynamical influence is a useful and principled criterion for reservoir refinement beyond static connectivity or activation statistics.","feed_headline":"Prune 20% of reservoir neurons, keep accuracy, gain 1.9x speed","feed_subtitle":"Neurons are ranked by their contribution to dominant input-driven transitions, not static connectivity.","key_machinery":"The object that carries the argument is the trajectory-averaged Jacobian Gramian $G = \\frac{1}{T}\\sum_{t=1}^{T} J(t)^{\\top}J(t)$, with $J(t) = (1-a)I + a\\,\\mathrm{diag}(f'(z(t)))W$ for the leaky ESN transition. Its eigendecomposition $G = V\\Lambda V^{\\top}$ gives directions in state space along which input-driven local perturbations grow most strongly on average; after normalizing eigenvalues and keeping the smallest number of modes whose cumulative energy reaches $\\tau = 0.9$, each neuron's score is $\\eta_i = \\sum_{k=1}^{r}\\tilde{\\lambda}_k V_{ik}^{2}$, the mode-energy-weighted participation of coordinate $i$ in the dominant subspace. Low-$\\eta_i$ neurons are removed from $W$, $W^{\\mathrm{in}}$, and $W^{\\mathrm{fb}}$, and only the readout is retrained by ridge regression. This turns pruning into a data-dependent dynamical-subspace calculation rather than a static graph or activation heuristic.","core_discovery":"On the paper's own terms, DMP establishes that a random untrained ESN reservoir contains dynamical redundancy that can be identified from the trajectory it actually follows. The importance of a neuron is defined as its weighted participation in the leading modes of the trajectory-averaged Jacobian Gramian, truncated to modes carrying at least 90% of the Gramian energy. Removing the lowest-scoring neurons in one shot and refitting only the readout improves or preserves mean NRMSE compared with the full reservoir, with the clearest gains on Mackey-Glass and Electricity, and remains competitive with newly initialized smaller ESNs and random pruning under a matched ten-seed protocol. Spectral-radius rescaling after pruning is not needed; the unscaled pruned reservoir is the recommended configuration.","pith_inferences":["A natural extension the paper does not test is iterative DMP: recompute the Gramian on the pruned reservoir and prune again; this would directly test the one-shot approximation and may recover the 30% pruning regime.","The same Gramian-energy criterion could also select reservoir capacity per task, since the cumulative-energy elbow provides a data-driven bound on how many neurons the dynamics actually require.","Because random pruning already captures much of the gain in real-world datasets, a practical deployment might treat DMP as a tie-breaker among several cheap criteria, with its largest added value on dynamics-dominated tasks such as chaotic forecasting."],"forward_implications":["At 20% pruning (1000 to 800 neurons), DMP yields lower mean NRMSE than the full ESN on all five datasets in the matched ten-seed protocol, with the largest separation on Mackey-Glass.","The pruned model uses 640,000 instead of 1,000,000 recurrent parameters, is 35.7% smaller on disk, and runs at 1.90 to 1.97 times inference speedup.","Matched random pruning also beats the full ESN on most datasets, so part of the benefit is simply smaller reservoir size; DMP's criterion adds a smaller, dataset-dependent margin.","Pruning at 10% removes too little redundancy and 30% often degrades accuracy, so the one-shot Gramian ranking is reliable only for moderate pruning ratios.","Spectral-radius rescaling after pruning does not help; DMP without rescaling is the main configuration."],"supporting_citations":[{"why":"Supplies the Echo State Network formulation and the Mackey-Glass benchmark used for chaotic forecasting.","marker":"[Jaeger, 2001]"},{"why":"Provides the reservoir computing framework with fixed recurrent weights and readout-only training that DMP preserves.","marker":"[Lukoševičius and Jaeger, 2009]"},{"why":"Supplies the leaky-integrator ESN update and spectral-radius stability condition used in the reservoir setup.","marker":"[Jaeger et al., 2007]"},{"why":"Motivates the Jacobian-spectrum view of one-shot recurrent-network pruning that DMP adapts to untrained reservoirs.","marker":"[Zhang and Stadie, 2019]"},{"why":"Shows that dominant dynamical modes can guide structured state-space pruning, the closest dynamical counterpart DMP extends.","marker":"[Gwak et al., 2024]"},{"why":"Provides activation-based reservoir pruning baselines that DMP is intended to improve upon.","marker":"[Dutoit et al., 2009]"},{"why":"Provides an effective connection-pruning criterion for ESNs and an example of the static criteria DMP contrasts.","marker":"[Scardapane et al., 2014]"},{"why":"Motivates task-specific node pruning and the performance-guided removal idea behind DMP.","marker":"[Yadav and Stender, 2025]"},{"why":"Motivates the freshly initialized smaller-ESN baseline by showing compact reservoirs can compete with larger random ones.","marker":"[Martinuzzi, 2025]"},{"why":"Motivates matched random pruning and fresh smaller networks as required controls in the evaluation protocol.","marker":"[Blalock et al., 2020]"}],"fun_headline_variants":["Prune by dynamical impact: cut neurons that don't shape transitions","Dynamical mode pruning keeps accuracy with leaner reservoirs","Rank neurons by transition modes, prune the rest, retain accuracy","Reduce reservoir redundancy without static pruning rules"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the dynamical-mode ranking computed once from the unpruned reservoir remains valid for the pruned reservoir; if pruning shifts the dominant transition directions, the neuron scores become miscalibrated.","fun_headline_variants_meta":{"raw":{"variants":["Prune by dynamical impact: cut neurons that don't shape transitions","Dynamical mode pruning keeps accuracy with leaner reservoirs","Rank neurons by transition modes, prune the rest, retain accuracy","Reduce reservoir redundancy without static pruning rules"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000205,"raw_usage":{"total_tokens":1332,"prompt_tokens":820,"completion_tokens":512,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":436,"completion_tokens_details":{"reasoning_tokens":446}},"tokens_in":436,"tokens_out":512,"duration_ms":6488,"temperature":1.0,"reasoning_tokens":446,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:13:05.871256+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct check: after pruning at 30%, recompute the trajectory-averaged Jacobian Gramian for the pruned reservoir and compare its dominant subspace with the original; if the principal-angle overlap is small, and re-ranking with the new Gramian changes the retained set and improves NRMSE, then the one-shot assumption fails.","supporting_citations":[{"cited_title":"Proceedings of the 2019 11th International Conference on Machine Learning and Computing , year =","cited_arxiv_id":null,"evidence_quote":"Motivates the Jacobian-spectrum view of one-shot recurrent-network pruning that DMP adapts to untrained reservoirs."},{"cited_title":"Advances in Neural Information Processing Systems , volume =","cited_arxiv_id":null,"evidence_quote":"Shows that dominant dynamical modes can guide structured state-space pruning, the closest dynamical counterpart DMP extends."},{"cited_title":"Chaos: An Interdisciplinary Journal of Nonlinear Science , volume =","cited_arxiv_id":null,"evidence_quote":"Motivates task-specific node pruning and the performance-guided removal idea behind DMP."},{"cited_title":"Chaos: An Interdisciplinary Journal of Nonlinear Science , volume =","cited_arxiv_id":null,"evidence_quote":"Motivates the freshly initialized smaller-ESN baseline by showing compact reservoirs can compete with larger random ones."}],"review_version":1}