{"id":"db9dc431-8023-4a80-8cf5-1741362adf00","arxiv_id":"2506.20469","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A linear genetic programming encoding of multi-branch DNN topologies, combined with a semantics-based Kriging surrogate and a pre-selection initialization, improves evolved network accuracy on two of three image datasets at reduced GPU cost.","lead":"This paper proposes NeuroLGP-MB, an evolutionary method that uses linear genetic programming to encode multi-branch connections in deep neural networks, guided by a Kriging surrogate model built from network output semantics. A pre-selection variant trains 150 candidate networks briefly to initialize the surrogate, and the authors report accuracy gains on two breast-cancer histology datasets with 12 to 17 percent GPU-time savings.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central efficiency claim depends on 10-epoch partial training preserving 30-epoch rank order, yet the paper never directly validates this transfer; with only four runs and overlapping error bars, the headline Sur-PS advantage is not established.","rationale":"The reader's weakest assumption is exactly the load-bearing condition for the paper's central claim: partial 10-epoch training must rank architectures similarly to full 30-epoch training, or the surrogate-assisted search is buying time at the cost of selection quality. The paper provides no direct evidence for that transfer; Table 1's surrogate fit metrics describe the KPLS model's calibration, not the partial-to-full rank agreement. This is more fundamental than the four-run issue: even with many more runs, a weak transfer correlation would undermine the method's efficiency-versus-quality trade-off. The four-run comparison with overlapping error bars and no significance tests makes the observed mean accuracy differences uninterpretable as real advantages, and the mixed surrogate metrics in Table 1 contradict the abstract's blanket 'outperforms simpler surrogate' wording. All of these weaknesses are addressable by a focused transfer experiment followed, if warranted, by larger replicated comparisons; they do not demonstrate that the method is wrong, only that the central claim is not yet established. The conditional verdict is therefore appropriate, and no verdict change is needed.","tokens_in":6325,"tokens_out":9411,"duration_ms":111778,"concrete_test":"Run a dedicated transfer experiment on BreakHis×40: sample roughly 100 architectures from the same LGP search space, train each to 10 and 30 epochs, and compute Kendall's tau between the 10-epoch and 30-epoch validation accuracies. Then train a KPLS model on the 30-epoch semantic vectors and test its predictions when candidate semantic vectors are taken from the 10-epoch checkpoints, reporting predicted-vs-actual R2 and Kendall's tau. If tau is below about 0.6 or the KPLS R2 is below about 0.5, the partial-training surrogate cannot be trusted to rank candidates, and the central efficiency claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that Surrogate-PS 'produced the best networks on average' while cutting GPU time depends on a transfer assumption that the paper never directly tests: the 10-epoch partial-training semantic vectors used as query points in the model-management loop (Section 2.3, annotations (i) and (ii)) must rank architectures the way full 30-epoch training would. Table 1 reports KPLS fit quality (MSE, Kendall's tau, R2), but no experiment measures the correlation between an architecture's validation accuracy at epoch 10 and its accuracy at epoch 30. If that correlation is weak, the KPLS model can be well calibrated on fully evaluated individuals yet mis-rank the partially trained candidates that drive the search; the reported 11.8-16.6% GPU-time saving would then come at the cost of selection quality, and the headline accuracy advantage on BreakHis×40 and ×200 could disappear. This risk is compounded by the small experimental basis: four runs, no significance tests, and overlapping ±std intervals (e.g., 0.919±0.032 vs 0.907±0.029 on ×40). The abstract's claim that the advanced surrogate 'outperforms baseline, computationally expensive, and simpler surrogate models' is also not supported by Table 1, where Sur-PS is not uniformly better than Sur in all surrogate metrics (e.g., BreakHis×200 R2 = 0.746 for Sur-PS vs 0.937 for Sur).","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NeuroLGP-MB, a Linear Genetic Programming representation for encoding multi-branch connections in deep neural networks, and combines it with a Kriging Partial Least Squares surrogate model. A pre-selection variant, Sur-PS, initializes the surrogate with high-performing partially trained individuals and uses a 40/60 split between fully evaluated and partially trained candidates. The method is compared with a baseline random-search approach, a fully trained evolutionary approach, and a standard surrogate-assisted evolutionary approach on BreakHis×40, BreakHis×200, and Chest X-Ray datasets over four runs. The authors report that Sur-PS achieves the highest average accuracy on BreakHis×40 and BreakHis×200 while reducing GPU time by 11.8–16.6% relative to the expensive model, and they analyze elite-population depth and topological complexity.","tokens_in":6623,"tokens_out":4073,"duration_ms":43574,"significance":"If the empirical claims are supported, the paper offers a practically relevant contribution: an LGP encoding that sidesteps graph-edit distance for multi-branch topologies, semantic vectors that allow surrogate-assisted search over thousands of candidate networks, and a pre-selection strategy that improves search initialization. The paper also provides an explicit analysis of how elite depth and complexity evolve, which is useful for understanding search behavior. However, the current evidence base is narrow: four runs per condition, no significance tests, and an unvalidated transfer assumption between 10-epoch and 30-epoch training. The central efficiency claim depends on this transfer, so the paper's contribution is currently promising but not fully established.","major_comments":[{"comment":"The central efficiency claim relies on the assumption that a 10-epoch partial training run ranks architectures the way a full 30-epoch training run would. In the model-management loop described in Section 2.3, annotations (i) and (ii), the KPLS surrogate is trained on fully evaluated individuals but queried on partially trained individuals. The paper never directly measures the correlation between an architecture's validation accuracy at epoch 10 and its accuracy at epoch 30. Table 1 reports KPLS fit quality (MSE, Kendall's Tau, R2) on the surrogate's training data, not the epoch-10-to-epoch-30 transfer. If partial-training rankings diverge from full-training rankings, the reported accuracy advantage of Sur-PS on BreakHis×40 and BreakHis×200, and the accompanying 11.8–16.6% GPU-time saving, would not transfer to real deployment. This should be tested directly, for example by reporting rank correlations between epoch-10 and epoch-30 accuracies over a sample of architectures on each dataset, or by comparing final selected networks against a full-training control.","section":"Section 2.3 and Section 3.1"},{"comment":"The headline claim that Sur-PS 'outperforms' the baseline, expensive, and simpler surrogate models is not statistically supported. Each condition is evaluated with only four runs, no significance tests are reported, and the standard deviations overlap. For BreakHis×40, Sur-PS gives 0.919±0.032 versus 0.907±0.029 for the expensive model and 0.904±0.029 for the surrogate model; for BreakHis×200, Sur-PS gives 0.939±0.010 versus 0.930±0.030 for the expensive model. On Chest X-Ray, the expensive method has the highest mean accuracy (0.917±0.016 versus 0.914±0.014 for Sur-PS), directly contradicting the abstract's statement that the advanced surrogate outperforms the computationally expensive model. The paper should either provide significance tests, confidence intervals, or an explicit statement that the differences are not significant and that the contribution is efficiency rather than accuracy gains.","section":"Section 4.1"},{"comment":"There is an internal inconsistency between the abstract's claim that the more advanced surrogate model 'outperforms baseline, computationally expensive, and simpler surrogate models' and the results in Table 1. On BreakHis×200, the standard surrogate has R2 of 0.9373 while Sur-PS has R2 of 0.7458; on BreakHis×40, the standard surrogate has slightly higher Kendall's Tau (0.6536 versus 0.6480) and R2 (0.6239 versus 0.6185). Only on Chest X-Ray are all three surrogate metrics better for Sur-PS. The claim of uniform superiority over the simpler surrogate is therefore not supported by the reported data. The abstract and conclusions should be tempered, or the surrogate-quality comparison should be analyzed separately from end-to-end accuracy and runtime.","section":"Table 1 and Abstract"},{"comment":"The experimental setup is not specified precisely enough to establish a fair comparison across methods. Population size and generation count are stated only for the expensive approach (Pop. size = 50, Gen. size = 15); the surrogate and Sur-PS variants do not clearly state their population and generation budgets. The baseline 'random search' method also needs a clear statement of how many architectures it evaluates and whether it uses the same encoding and evaluation protocol. The time-analysis sentence mentions '150 individuals trained to 10 epochs' for Sur-PS, which implies a specific population size, but this is not connected to the stated population/generation parameters. Without this information, the reported runtime savings and accuracy comparisons cannot be independently assessed.","section":"Section 3.1"}],"minor_comments":[{"comment":"There are several typos: 'Chext X-Ray' should be 'Chest X-Ray', and 'Kendell's Tau' should be 'Kendall's Tau'.","section":"Section 4.1"},{"comment":"Equation (1) contains an extra closing parenthesis: D(s(p_i), s(p_j))) should be D(s(p_i), s(p_j)).","section":"Equation (1)"},{"comment":"The table in Figure 1 has two 'I2' column headers; the second should presumably be 'I3'. Please clarify the register operand naming.","section":"Figure 1"},{"comment":"The KPLS method is referenced to [1] but no KPLS hyperparameters (e.g., number of PLS components, kernel parameters) are reported. Providing these would help reproducibility.","section":"Section 2.1"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant problem and the proposed NeuroLGP-MB representation has clear potential, but the empirical evidence for the central claim is currently too thin. The lack of significance testing and the unvalidated epoch-10-to-epoch-30 transfer assumption are the main obstacles. I would encourage the editor to request a revision that either adds the missing validation and statistical analysis or substantially softens the claimed advantages."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, incremental extension of the authors' own NeuroLGP-SM work, not a breakthrough. The genuinely new pieces are an LGP encoding for multi-branch DNNs and a pre-selection scheme that initializes the KPLS surrogate with high-performing individuals. The semantic-vector distance idea is real and it does let the surrogate scale to thousands of sample points, which is useful for DNN neuroevolution.\n\nWhat the paper does well: the encoding is clever and the reverse-traversal construction for concat layers is clearly explained. The pre-selection idea is sensible—seed the surrogate with good individuals instead of random ones—and it does buy an 11.8–16.6% GPU-time reduction over full evolution while roughly holding accuracy on two of three datasets. The depth/complexity analysis of elite individuals is a nice extra. There is no circularity problem; the surrogate is fit to measured fitness values and used for standard model management.\n\nThe soft spots are real but addressable. The central claim that Sur-PS \"produced the best networks on average\" rests on four runs with overlapping standard deviations and no significance tests. On Chest X-Ray the expensive baseline has the highest mean accuracy. The abstract's claim that the advanced surrogate \"outperforms baseline, computationally expensive, and simpler surrogate models\" is not uniformly supported by Table 1: on BreakHis×200 the plain surrogate has better R² (0.937 vs 0.746). The biggest unaddressed risk, and the one I'd push on in review, is the transfer assumption: the surrogate is queried on semantic vectors from 10-epoch partial training, but the fitness target is 30-epoch accuracy. The paper never directly measures the correlation between epoch-10 and epoch-30 rankings, so we don't know whether the time saving comes at the cost of selection quality. That is testable and should be done. Also, the semantic vectors are built from the test set's outputs; the paper doesn't clarify whether test data leaks into surrogate training. No code or data is released, which makes the four-run statistics harder to evaluate.\n\nBottom line: the method is plausible, internally consistent, and honestly described. The flaws are not fatal. I'd send it to a serious referee, but the revision needs significance tests, a direct epoch-10-vs-epoch-30 rank-correlation analysis, and a clear statement on test-set usage. Worth a reading-group slot if you want to discuss surrogate model management in neuroevolution.","headline":"A legitimate incremental extension of the authors' own NeuroLGP-SM work, but the headline efficiency-accuracy claim rests on an untested partial-training transfer assumption and four-run statistics.","tokens_in":7182,"tokens_out":1912,"would_cite":true,"duration_ms":20028,"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 semantic surrogate and pre-selection allow multi-branch DNN architectures to evolve competitively while cutting GPU time by 11.8–16.6%.","keywords":["neuroevolution","neural architecture search","linear genetic programming","multi-branch connections","surrogate-assisted evolutionary algorithms","semantic vectors","Kriging partial least squares","pre-selection"],"falsifier":"Take one generation of evolved architectures, train each for 10 and 30 epochs, and compare the KPLS-predicted ranking with the full-training ranking; if Kendall's tau drops below roughly 0.5, or the pre-selected best architecture's full-training accuracy underperforms a randomly sampled baseline, then the time savings of Surrogate-PS do not transfer to deployment.","tokens_in":6136,"feed_emoji":"🧬","tokens_out":7758,"duration_ms":77135,"temperature":0.7,"pith_summary":"Multi-branch connections make deep networks more expressive but harder to search, and full training of every candidate is expensive. This paper argues that both problems can be eased at once: a Linear Genetic Programming encoding expresses multi-branch topologies, while a semantic surrogate model estimates fitness from partially trained networks. The proposed pre-selection variant, which initializes the surrogate with high-performing individuals, produced the best average test accuracy on BreakHis ×40 and BreakHis ×200 and cut GPU time by 11.8–16.6% relative to the fully trained evolutionary search. If the claim holds, surrogate-assisted neuroevolution can move from dozens or hundreds of evaluated points to thousands on modern DNNs.","feed_headline":"Multi-branch neural nets evolve with 16% less GPU time","feed_subtitle":"A 10-epoch ranking plus semantic vectors makes neural architecture search faster and still accurate.","key_machinery":"The machinery is a Linear Genetic Programming (LGP) encoding—a linear sequence of register-based instructions that builds a DNN graph in reverse—combined with a semantic distance: each network's output vector is the flattened final softmax layer over all test images times the number of classes. Because these vectors are fixed-length, architectures of different depth and branch count can be compared by ordinary distance, and Kriging Partial Least Squares (KPLS) can regress fitness onto them without maximum-likelihood estimation over huge parameter spaces. The pre-selection management strategy supplies the surrogate's initial training data by sampling the best partially trained individuals from a first generation rather than random individuals, and it carries the argument by bypassing genotype-based topology distances and allowing thousands of sample points.","core_discovery":"The central discovery is that the semantic output of a network—the flattened softmax predictions over the test set—provides a fixed-length distance vector for surrogate modeling regardless of how many layers or branches a network has. NeuroLGP-MB encodes multi-branch topologies with Linear Genetic Programming, reading the genotype in reverse so that a CONCAT instruction marks a split; the evolutionary search therefore avoids expensive graph-edit genotypic distances. The proposed pre-selection surrogate pipeline splits the population (40% fully trained to 30 epochs, 60% partially trained to 10 epochs), feeds the semantic vectors of fully evaluated individuals into Kriging Partial Least Squares, estimates fitness of new individuals, and uses expected improvement for selection. Across four runs, this Surrogate-PS variant achieved the highest mean accuracy on BreakHis ×40 (0.919 ± 0.032) and BreakHis ×200 (0.939 ± 0.010), matched the other methods on Chest X-Ray, and cut GPU hours by 11.8–16.6% relative to the expensive approach. Elite-population analysis shows the search favors deeper networks with relatively few concatenation layers.","pith_inferences":["Inference: If the partial-training ranking transfers to full training (Kendall's tau 0.63–0.74 in the paper), the same pre-selection scheme could lower cost in other neural architecture search spaces where genotype comparison is expensive, including graph-based architecture search.","Inference: The fixed-length semantic vector could double as an explicit diversity objective in multi-objective neuroevolution, not only as a distance for surrogate fitness.","Inference: Because the semantic vector's length grows with dataset size times class count, a testable extension is to use per-batch or class-conditional semantic summaries to keep the surrogate scalable to larger datasets."],"forward_implications":["Multi-branch topologies can be compared for surrogate modeling using fixed-length semantic output vectors, so no graph-edit distance over variable-length genotypes is required.","Pre-selection with 10-epoch partial training provides a cheaper fitness estimate that preserves or improves mean accuracy, with 11.8–16.6% GPU-time savings over full training.","Surrogate-assisted neuroevolution can scale to thousands of evaluated sample points, moving beyond the dozens or hundreds typical of earlier surrogate neuroevolution.","The evolutionary search under this setup tends to produce deeper networks with relatively few branch/concatenation layers, an observation that can guide future mutation and crossover design."],"supporting_citations":[{"why":"Supplies the Kriging Partial Least Squares surrogate model used to estimate fitness from high-dimensional semantic vectors.","marker":"[1]"},{"why":"Supplies the Linear Genetic Programming representation that encodes multi-branch DNN topologies.","marker":"[2]"},{"why":"Provides the Chest X-Ray dataset used as one of the three test benchmarks.","marker":"[6]"},{"why":"Provides the BreakHis breast-cancer histopathology datasets on which the pre-selection surrogate achieves its best results.","marker":"[8]"},{"why":"Establishes the NeuroLGP-SM semantic surrogate approach that this paper extends with multi-branch encoding and pre-selection.","marker":"[9]"},{"why":"Earlier scalable surrogate-assisted neuroevolution work that motivates scaling from dozens/hundreds to thousands of sample points.","marker":"[10]"}],"fun_headline_variants":["Surrogate evolution slashes DNN design GPU time by 16%","Semantic vectors make neural architecture search cheaper","NeuroLGP-MB: efficient multi-branch DNN evolution","Evolving multi-branch nets with 16% less GPU time","Surrogate-assisted evolution speeds up multi-branch DNN design"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method depends on the assumption that a network trained for only 10 epochs—ranked through the KPLS surrogate—predicts which architectures would perform best after the full 30 epochs, and the reported rank correlations are moderate and vary by dataset.","fun_headline_variants_meta":{"raw":{"variants":["Surrogate evolution slashes DNN design GPU time by 16%","Semantic vectors make neural architecture search cheaper","NeuroLGP-MB: efficient multi-branch DNN evolution","Evolving multi-branch nets with 16% less GPU time","Surrogate-assisted evolution speeds up multi-branch DNN design"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000739,"raw_usage":{"total_tokens":3310,"prompt_tokens":966,"completion_tokens":2344,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":2258}},"tokens_in":582,"tokens_out":2344,"duration_ms":15734,"temperature":1.0,"reasoning_tokens":2258,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:47:13.405346+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one generation of evolved architectures, train each for 10 and 30 epochs, and compare the KPLS-predicted ranking with the full-training ranking; if Kendall's tau drops below roughly 0.5, or the pre-selected best architecture's full-training accuracy underperforms a randomly sampled baseline, then the time savings of Surrogate-PS do not transfer to deployment.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Kriging Partial Least Squares surrogate model used to estimate fitness from high-dimensional semantic vectors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Linear Genetic Programming representation that encodes multi-branch DNN topologies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Chest X-Ray dataset used as one of the three test benchmarks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the BreakHis breast-cancer histopathology datasets on which the pre-selection surrogate achieves its best results."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the NeuroLGP-SM semantic surrogate approach that this paper extends with multi-branch encoding and pre-selection."}],"review_version":1}