{"id":"f568e39c-5a49-49fa-a863-0c74f51d969c","arxiv_id":"2412.10153","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A sampling-based training strategy using cached loss ranking, frequency-guided crossover, and mutation reduces INR training time by 48-66% with no loss in quality.","lead":"This paper proposes a way to train implicit neural representations (neural networks that map coordinates to signals) on only a fraction of the data points each step, cutting training time by roughly half while preserving or improving reconstruction quality. The method is relevant to anyone fitting images, audio, or 3D shapes with neural fields, where training cost is a bottleneck.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Paper's own β=0.3 (Table 16) and β=0.05 (Table 12) results contradict the unqualified 'consistently enhances' claim; benefit is conditional on selection ratio.","rationale":"I read the paper in good faith. The core mechanism—cache-based hard-example selection with crossover and mutation—is clearly described and the ablations are informative. The reader's identified weakest assumption (stale cached fitness) is actually contradicted by the paper's own ablation: τ=1 (fresh cache every iteration) gives 35.51 dB vs. τ=100 (stale cache) giving 37.81 dB, so staleness appears beneficial, not harmful. The more serious issue is an internal inconsistency: the headline 'consistently enhances' claim is contradicted by the paper's supplementary results at β=0.3 and β=0.05. This is not a disagreement with consensus; it is a logical contradiction within the paper. It bears directly on the central narrative that sparsification per se improves performance. The acceleration claim (48-66% time reduction) is plausible and supported by the time accounting, but the 'superior convergence' claim is only established at the specific selection ratios tested and is not 'consistent'. A β-sweep with multiple seeds would settle the boundary of the method's benefit and force the authors to qualify their claim. My verdict remains CONDITIONAL, matching the reader's verdict, but for a different principal reason than the reader's stale-cache concern.","tokens_in":22784,"tokens_out":17425,"duration_ms":180386,"concrete_test":"On the same DIV2K 5k-iteration protocol, run EVOS and Standard with constant-scheduler ratios β ∈ {0.2, 0.3, 0.4, 0.5, 0.6, 0.7}, using at least 3 seeds each and reporting mean ± std. Determine the smallest β at which EVOS's PSNR ≥ Standard's. If that threshold is β=0.5 or higher, the 'consistently enhances' claim is falsified as stated and must be revised to a ratio-qualified claim. If EVOS matches or beats Standard at β=0.3 with adequate statistical power, the contradiction is resolved and the original claim may stand.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central insight stated in Sec. 1 and the Conclusion is that sparsification 'consistently enhances' training performance. This is directly contradicted by the paper's own results: Appendix H (Table 16) reports EVOS at β=0.3 reaching 36.24 dB at 5k iterations versus 37.10 dB for Standard, a 0.86 dB deficit, and Table 12 (β=0.05) shows EVOS at 29.74 dB, far below Standard. Appendix H admits 'it fails to surpass standard training performance' at 30% selection, yet the abstract and conclusion retain the unqualified claim. For the central claim to hold, EVOS should beat or match Standard across the operating range; the paper's evidence shows it only does so for β≥0.5. The method's advantage is thus contingent on the selection ratio, and the paper's broader 'sample weighting benefits overfitting' narrative is overstated. This does not refute the 48-66% acceleration at default settings, but it requires the claims to be qualified and the regime of benefit to be characterized.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes EVOS, a training-time coordinate-selection method for implicit neural representations. At each iteration EVOS selects a subset of coordinates using sparse fitness evaluation with cached losses, a frequency-guided crossover based on low- and high-frequency fitness, and a mutation step that re-injects random coordinates; the network is trained with a cross-frequency loss on the selected subset. Experiments on image, audio, text, and SDF fitting tasks report 48%--66% training-time reductions relative to full-coordinate training, with PSNR/SSIM improvements at the default beta=0.5/step-wise configuration, and the paper claims that strategic sparsification 'consistently enhances' INR training.","tokens_in":23003,"tokens_out":7582,"duration_ms":74515,"significance":"EVOS is a simple and practical idea, and the paper contains a substantial experimental effort: multiple backbones, network sizes, modalities, ablations of each component, comparisons with four sampling-based baselines, and released code. The default configuration (constant beta=0.5 or the step-wise scheduler of INT) delivers acceleration close to the advertised range and quality at least matching or exceeding standard training in most reported settings. These strengths make the method worth pursuing. The contribution is nevertheless overstated: the headline 'consistently enhances' claim is contradicted by the authors' own beta=0.3 and beta=0.05 experiments, and the empirical evaluation lacks error bars and relies on reimplemented baselines.","major_comments":[{"comment":"The unqualified claim that strategic sparsification 'consistently enhances' training performance is contradicted by the manuscript's own results. In Appendix H, Table 16 (beta=0.3 constant scheduler), EVOS reaches 36.24 dB at 5k iterations versus 37.10 dB for Standard; the appendix text explicitly concedes that at 30% selection EVOS 'fails to surpass standard training performance.' In addition, Table 12 (beta=0.05) reports EVOS at 29.74 dB, below Uniform (30.14 dB) and Soft Mining (30.73 dB) and far below the Standard reference of 37.10 dB from Table 1. The Abstract, Sec. 1, Sec. 4.2 ('Sample Weighting ALSO Benefits Overfitting'), and Sec. 5 therefore need to be revised to state that the benefit is conditional on the selection ratio/scheduler, and the operating regime (e.g., beta >= 0.5 with constant, or step-wise) should be characterized explicitly. The 48%-66% acceleration claim at default settings is not refuted by this comment, but the current framing is.","section":"Abstract; Sec. 1; Sec. 4.2; App. H, Table 16"},{"comment":"The claim in Sec. 3.5 that EVOS achieves acceleration 'without introducing additional storage overhead' is inconsistent with Eq. (6). The cross-frequency loss term Lh_t uses \\hat F_theta(x \\ z_t), cached full-coordinate network outputs from the last key iteration, and Eq. (1) also caches per-coordinate fitness values. Both require storing O(N) values, which is additional memory relative to standard INR training. The authors should either remove the 'no storage overhead' claim or quantify the memory cost and explain why it is negligible.","section":"Sec. 3.5; Eq. (6)"},{"comment":"All quantitative tables report single runs without error bars or variance estimates. Given the consistent-improvement claim, this is a load-bearing issue: several margins are small (e.g., Table 4 at 10k iterations: 38.71 vs 38.60 dB; Table 1 at 5k: SSIM 0.962 vs 0.964 for Standard), and the comparison includes baselines EGRA and Expan. that were reimplemented by the authors because official implementations are unavailable (Supp. F). Please report multiple seeds or state and justify determinism, and provide confidence intervals or seed-level tables for the main comparisons.","section":"Sec. 4.2; Tables 1-17"},{"comment":"The design of Gamma(t) is motivated by the G_t curves in Appendix I, but the connection between G_t and the accuracy of cached fitness rankings is never made quantitative. Eq. (1) relies on cached values from the last key iteration; Table 5 shows that increasing evaluation frequency (tau=1, tau=10) actually degrades performance, which the paper attributes to instability. This counterintuitive result needs a more direct analysis: for example, measure how often the top-k selected set changes between key iterations, or ablate cache staleness while controlling for evaluation cost. Without this, the sparse-evaluation schedule remains an unexplained empirical choice rather than a validated mechanism.","section":"Sec. 3.2; Appendix I"}],"minor_comments":[{"comment":"With tau=100 and theta=0.01, the interval tau - theta*t/T changes from 100 to 99.99 over T=5000 iterations, which cannot produce the 'more frequent evaluations as training progresses' described in Sec. 3.2. Please check the formula or the reported hyperparameter.","section":"Sec. 3.2, Eq. (2)"},{"comment":"The crossover operation is not fully specified: it is not clear how the balancer Psi guarantees that exactly k offspring are produced, and the notation U_p l and U_(1-p)l is undefined. Please spell out the sampling procedure.","section":"Sec. 3.3, Eqs. (4)-(5)"},{"comment":"The caption and table layout make it difficult to see which rows use the constant vs. step-wise scheduler; consider splitting into two tables or adding a scheduler column.","section":"Table 1"},{"comment":"The 'link' for code in the Abstract and Sec. 4.1 is not a URL; provide the actual repository address.","section":"Abstract; Sec. 4.1"},{"comment":"It is not stated whether the DIV2K and Kodak numbers are averages over a set of images or results on a single image; if averages, report the number of images and the variance.","section":"Sec. 4.1"}],"recommendation":"major_revision","confidential_remarks":"I see no evidence of circularity: the method is empirical and the evaluation schedule is motivated by a measured G_t curve, not by the final results. The main risks are the overstatement in the abstract and conclusion and the absence of variance reporting; both are addressable in revision. The paper is within scope and the default-configuration acceleration claims deserve a fair hearing after the claims are qualified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my read. EVOS is a solid, incremental INR acceleration method. The 48-66% training time reduction at default settings is plausible, and the paper's own ablations show each of the three components earns its keep. But the headline claim that sparsification 'consistently enhances' performance is not supported by the paper's own data: at β=0.3 (Table 16) EVOS is 0.86 dB below Standard at 5k iterations, and at β=0.05 (Table 12) it is far below. The appendix even admits the 30% case 'fails to surpass standard training performance.' The abstract and conclusion need to be qualified to say the benefit is conditional on selection ratio and scheduler. That is a real correction, not a fatal flaw.\n\nWhat is genuinely new is the specific combination: cached sparse fitness evaluation, Laplacian-guided crossover for high-frequency emphasis, and a random mutation step that injects non-survivors. INT is the closest prior work but does not use this combination. The method description is clear, the ablation study is informative, and the experiments across images, audio, text, and 3D shapes give reasonable evidence of generality. The G_t curves in the appendix justify the evaluation schedule without circularity. Credit where due.\n\nWeaknesses beyond the overclaim: all results are single runs with no error bars; for a method that includes stochastic mutation, the variance is unknown. The baseline EGRA and Expansive Supervision are reimplemented, which is fair only if the code is released and checked; the paper points to a project page but I cannot verify a release. The 'evolutionary' framing is mostly packaging—the real content is importance sampling with a cache and a fixed random perturbation. That is fine, but the novelty should be positioned as engineering.\n\nNet: this deserves serious peer review. The acceleration result at β=0.5 is useful and likely reproducible, provided code ships. I would ask for multiple seeds, qualified claims, and a code check before acceptance. Send it to review, and have the referees focus on the β-dependence and the reimplemented baselines.","headline":"EVOS is a solid, incremental INR acceleration method whose default-configuration speedups are plausible, but the paper's unqualified 'consistently enhances' claim is contradicted by its own β=0.3 and β=0.05 experiments.","tokens_in":23513,"tokens_out":2470,"would_cite":true,"duration_ms":23499,"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":"EVOS claims that selecting an evolutionary subset of coordinates each step cuts implicit-neural-representation training time 48–66% while keeping or improving quality — sparsification as implicit sample weighting.","keywords":["implicit neural representation","training acceleration","coordinate selection","evolutionary algorithm","spectral bias","sparse forward passes","sample weighting"],"falsifier":"Run EVOS on a signal whose hardest coordinates move quickly — a video whose sharp edges translate between frames — under the default evaluation schedule, comparing per-iteration PSNR with full-coordinate training at matched wall-clock time. If the selected subset stops tracking the true top-error set and reconstruction quality falls below the full-data baseline, the claim that sparse selection consistently beats full data is refuted; the paper's own statistic $G_t(\\tau,\\sigma)$ (Eq. 12) gives the direct measurement, and freezing the cache ($\\Gamma(t) \\equiv 0$) after some iteration shows exactly when the ranked coordinates diverge from the true hardest points.","tokens_in":22584,"feed_emoji":"🧬","tokens_out":18979,"duration_ms":163366,"temperature":0.7,"pith_summary":"Implicit neural representations — MLPs that map coordinates to signal values — are flexible but slow to train, because every iteration normally evaluates every coordinate. This paper proposes EVOS, which each step feeds the network only a subset of coordinates chosen by an evolutionary process: samples with the largest current reconstruction error are the 'fittest' survivors, they are recombined by a frequency-guided crossover, and a small random mutation keeps the selection honest. The paper's central claim is that this strategic sparsification cuts training time by roughly half (48–66% across settings) while matching or exceeding full-data reconstruction quality, so the acceleration is not a compromise but an improvement. If the claim holds, INR-based compression, novel-view synthesis, and physics simulation could run at about twice the speed without sacrificing fidelity.","feed_headline":"Train on fewer points, beat full-data fitting, run twice as fast","feed_subtitle":"Feeding only the hardest coordinates each step cuts training time 48–66% without hurting fit quality.","key_machinery":"The load-bearing object is the evolved coordinate subset $z_t \\subset x$ that replaces the full coordinate set in each forward pass, built as $z_t = M(x'_t \\odot x''_t)$ where $\\odot$ is crossover and $M$ is mutation. Selection is driven by the sparse fitness function $f_t(x) = \\Gamma(t)\\, D(F_\\theta(x), y) + (1-\\Gamma(t))\\,\\hat f(x)$ (Eq. 1): a full ranking of coordinates by reconstruction error is computed only at key iterations, with spacing shrinking linearly through $\\Gamma(t)$ in Eq. 2, and the ranking is cached in between. To counter spectral bias, two parent sets are drawn per Eq. 3 — low-frequency parents ranked by square error and high-frequency parents ranked by Laplacian error — and crossed by a balancer $\\Psi$ that allocates offspring slots according to the ratio $p = f^{\\mathrm{low}}_t/(f^{\\mathrm{low}}_t + f^{\\mathrm{high}}_t)$ (Eqs. 4–5). A cross-frequency loss (Eq. 6) applies the Laplacian to predictions on $z_t$ augmented with cached predictions for unselected coordinates, since the spatial convolution filter needs the full volume of attributes. Augmented unbiased mutation (Eq. 7) then injects $\\alpha k$ uniformly random non-survivors with $\\alpha = 0.5$, preventing the deterministic cached rankings from entrenching selection bias.","core_discovery":"The central claim is that strategically dropping training samples does not trade away quality: it systematically improves per-iteration performance while cutting wall-clock cost. In direct comparison on 3×256 SIREN over DIV2K images, EVOS without its cross-frequency supervision reaches 37.49 dB PSNR at 5,000 iterations versus 37.10 dB for full-coordinate training while performing roughly half the forward passes, and the full EVOS reaches 37.81–38.43 dB depending on scheduler; time-to-target reductions are 47.7%, 59.4%, and 66.5% for 25, 30, and 35 dB PSNR. The paper explains the gain through sample weighting: reweighting coordinates by their current error implicitly regularizes the loss, and this benefit carries over from generalization tasks to the pure overfitting regime of signal fitting.","pith_inferences":["Because the fitness ranking is just current per-coordinate error, the benefit may come mostly from the cache-plus-mutation scheme rather than the evolutionary framing; a stripped-down baseline that ranks coordinates by cached error and injects random points could isolate that, which the paper does not run.","The paper's own statistic $G_t(\\tau,\\sigma)$ (Eq. 12) could drive an adaptive evaluation schedule: when the measured top-error set shifts faster than the linear schedule assumes, key iterations should arrive sooner — a testable extension the paper leaves implicit.","Dense-coordinate fits where the hard set moves over time, such as physics-informed networks on collocation points or dynamic scenes, are natural next targets, since the same reweighting argument should apply there; nothing in the paper tests this.","The fixed hyperparameters — mutation ratio $\\alpha = 0.5$ and the crossover balance $p$ — are tuned empirically, and an adaptive version that derives them from the gap between cached and fresh fitness could extend the gains to loss landscapes that change faster than the default schedule."],"forward_implications":["Wall-clock training time drops by roughly half across all tested backbones (SIREN, PEMLP, WIRE, GAUSS, FINER) and network sizes, with the relative gain growing as the network gets larger, so existing INR applications inherit the speedup without changing architecture.","The sample-weighting view implies that 'use every training point' is not automatically the right default for signal fitting: per-iteration quality can rise when hard coordinates are over-sampled, even though fitting a signal is a pure overfitting task with no test set.","The same evolutionary selection transfers across modalities — 2D images (DIV2K and Kodak), 1D audio (LibriSpeech), 2D text, and 3D signed-distance shapes — with consistent efficiency gains, so the mechanism is not tuned to a single signal type.","Evaluating fitness too often ($\\tau = 1$ in Eq. 2) costs time and degrades final quality, so sparse cached evaluation is not merely an engineering shortcut but part of what keeps training stable.","Selection overhead is only 1.34% of training time, so nearly all of the 48–66% saving appears in the bottom line, placing EVOS ahead of the closest prior method (INT), which shows unstable early-training quality."],"supporting_citations":[{"why":"INT, the closest prior sampling-based INR accelerator; supplies the greedy coordinate-selection baseline EVOS must beat and the step-wise selection scheduler EVOS adopts.","marker":"[61]"},{"why":"Soft Mining, a Monte-Carlo mining method for neural-field acceleration; the strongest recent sampling baseline, reimplemented and tuned for general INR tasks in the comparison.","marker":"[21]"},{"why":"EGRA, the edge-guided ray-allocation baseline; one of the static-prior sampling methods EVOS is compared against.","marker":"[15]"},{"why":"Expansive Supervision, the frequency-prior sampling baseline from the same research group; adapted from NeRF training for direct comparison.","marker":"[63]"},{"why":"SIREN, the periodic-activation INR backbone used across the main experiments and the source of the signal-fitting task formulation.","marker":"[50]"},{"why":"Establishes spectral bias of neural networks, the low-frequency preference that motivates EVOS's frequency-guided crossover and Laplacian-based selection.","marker":"[39]"},{"why":"RINAS, cited for the claim that deterministic cached or shuffled sample selection introduces training bias, motivating EVOS's augmented unbiased mutation.","marker":"[66]"},{"why":"Model-agnostic random weighting, cited to support the paper's interpretation of sparsified training as implicit sample weighting that regularizes the loss.","marker":"[16]"}],"fun_headline_variants":["EVOS cuts training time 48–66% by picking fittest points","Select fittest coordinates, cut training time up to 66%","Evolutionary sample selection speeds INR training 48–66%","Fitness-based sampling: half the passes, better fits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the error ranking cached at the last key iteration stays accurate enough, between evaluations, to identify the coordinates the network currently fits worst, and that injecting 50% random points each step is enough to keep any staleness in that cache from steering training astray.","fun_headline_variants_meta":{"raw":{"variants":["EVOS cuts training time 48–66% by picking fittest points","Select fittest coordinates, cut training time up to 66%","Evolutionary sample selection speeds INR training 48–66%","Fitness-based sampling: half the passes, better fits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000316,"raw_usage":{"total_tokens":1774,"prompt_tokens":911,"completion_tokens":863,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":789}},"tokens_in":527,"tokens_out":863,"duration_ms":8453,"temperature":1.0,"reasoning_tokens":789,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:15:59.385089+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run EVOS on a signal whose hardest coordinates move quickly — a video whose sharp edges translate between frames — under the default evaluation schedule, comparing per-iteration PSNR with full-coordinate training at matched wall-clock time. If the selected subset stops tracking the true top-error set and reconstruction quality falls below the full-data baseline, the claim that sparse selection consistently beats full data is refuted; the paper's own statistic $G_t(\\tau,\\sigma)$ (Eq. 12) gives the direct measurement, and freezing the cache ($\\Gamma(t) \\equiv 0$) after some iteration shows exactly when the ranked coordinates diverge from the true hardest points.","supporting_citations":[{"cited_title":"Nonparametric teaching of implicit neural representations","cited_arxiv_id":null,"evidence_quote":"INT, the closest prior sampling-based INR accelerator; supplies the greedy coordinate-selection baseline EVOS must beat and the step-wise selection scheduler EVOS adopts."},{"cited_title":"Accelerating neural field training via soft mining","cited_arxiv_id":null,"evidence_quote":"Soft Mining, a Monte-Carlo mining method for neural-field acceleration; the strongest recent sampling baseline, reimplemented and tuned for general INR tasks in the comparison."},{"cited_title":"Egra-nerf: Edge-guided ray allocation for neural radiance fields","cited_arxiv_id":null,"evidence_quote":"EGRA, the edge-guided ray-allocation baseline; one of the static-prior sampling methods EVOS is compared against."},{"cited_title":"Expansive Supervision for Neural Radiance Field","cited_arxiv_id":"2409.08056","evidence_quote":"Expansive Supervision, the frequency-prior sampling baseline from the same research group; adapted from NeRF training for direct comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SIREN, the periodic-activation INR backbone used across the main experiments and the source of the signal-fitting task formulation."},{"cited_title":"Hamprecht, Yoshua Bengio, and Aaron C","cited_arxiv_id":null,"evidence_quote":"Establishes spectral bias of neural networks, the low-frequency preference that motivates EVOS's frequency-guided crossover and Laplacian-based selection."},{"cited_title":"RINAS: Training with Dataset Shuffling Can Be General and Fast","cited_arxiv_id":"2312.02368","evidence_quote":"RINAS, cited for the claim that deterministic cached or shuffled sample selection introduces training bias, motivating EVOS's augmented unbiased mutation."},{"cited_title":"Model-agnostic random weighting for out-of-distribution generalization","cited_arxiv_id":null,"evidence_quote":"Model-agnostic random weighting, cited to support the paper's interpretation of sparsified training as implicit sample weighting that regularizes the loss."}],"review_version":1}