{"id":"719c69b8-fe13-4a0c-bd3d-18644c8e4659","arxiv_id":"2605.25170","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A network that grows, prunes, and freezes its own layers during training reaches 94% source-finding success in simulated plume navigation, though the spectral guarantee rests on unverified assumptions.","lead":"A new training method lets a robot's neural network add, trim, and lock layers while learning, and the authors test it on simulated smell-based navigation. The method reaches 94% source-finding success, but the theoretical guarantee and the experimental evidence are not yet airtight.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1 is vacuous as stated: U⊂C\\R cannot contain the real spectrum of M^(ℓ), and even with a corrected domain the contraction/homogeneity conditions are never verified on pruned/frozen GPF layers.","rationale":"I read the paper as making two separable claims: a practical adaptive RL method for olfactory navigation, and a theoretical RMT guarantee that eigenvalue composition is preserved under GPF growth. The empirical claim is already weakened by single-seed evaluation and checkpoint selection, but the more load-bearing vulnerability is the theoretical claim. The formal domain issue in Theorem 1 makes it vacuous as written; even if that is repaired as a typo, the transition from homogeneous iid-initialized layers to sparse, pruned, frozen GPF layers is missing and explicitly deferred in the supplementary material. Because the theorem is the stated justification for why growing depth is spectrally safe, the central theoretical contribution is not currently established. The paper does release code, and the key quantities are empirically checkable, so the gap is addressable rather than terminal. This matches the reader's conditional assessment: a revision that corrects the theorem statement, verifies the contraction condition on actual GPF checkpoints, and adds multi-seed results would license acceptance. Thus I keep the verdict unchanged.","tokens_in":18368,"tokens_out":7442,"duration_ms":87191,"concrete_test":"Using the released code, rerun the plume experiment with the reported hyperparameters. At each grow/prune/freeze event, save W^(ℓ), compute the empirical spectrum of M^(ℓ), and attempt to find any open U⊂C\\R containing it; the attempt will fail, confirming the formal defect. Then, with the theorem restated on U=C\\R, estimate the layer-wise contraction factor empirically: for a small perturbation δs in the Stieltjes transform, measure sup_z |Ψ_ℓ(s+δs)-Ψ_ℓ(s)|/|δs| on a grid of z∈C\\R using the next layer's transform as Ψ_ℓ. If this sup is ≥1 for any pruned/frozen layer, Condition 2 fails for actual GPF weights, so Theorem 1 cannot ground the spectral-preservation claim. If the sup is <1 across all layers and the heterogeneous product bound holds, the theorem can be salvaged.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim—that Pennington & Worah's single-layer spectral analysis extends to GPF's growing depth—rests entirely on Theorem 1 (§3.2), but the theorem is not connected to the actual GPF dynamics. First, its hypothesis 'an open set U ⊂ C\\R containing the spectral support of M^(ℓ)' is unsatisfiable for any non-degenerate Gram matrix: the spectrum of M^(ℓ) is real, and C\\R contains no real points. Second, even with a corrected domain, Conditions 1–2 require homogeneous layers ψℓ=Ψ and a uniform contraction ∥D_sΨ∥ ≤ ρ(z)<1. GPF layers after pruning/freezing are sparse, non-iid, and layer-dependent (Table 4 shows 47.9%, 19.6%, and 9.9% weight retention at successive prune passes), so the conditions are neither implied by Pennington & Worah nor verified. Supplementary §11.1 explicitly defers nonhomogeneous layers to future work. Hence the 'preserved eigenvalue composition' claim is not established for GPF-grown networks. The MP-shaped eigenvalue plots and 94% navigation result are empirical evidence, but they do not supply the missing theorem.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Grow-Prune-Freeze (GPF) networks, which add hidden layers when validation loss plateaus, prune low-belief / low-magnitude neurons, and freeze stable layers as a continual-learning mechanism. The method is instantiated with an Expected SARSA policy on a Farrell–Murlis turbulent plume simulator, reporting 98% on the best training checkpoint and 94% on 100 held-out episodes. The theoretical section claims to extend Pennington & Worah (2017) from single hidden layers to multi-layer continual-learning models via composition of Stieltjes transforms, asserting that eigenvalue composition is preserved as layers are added. Additional experiments in the Supplementary Material apply GPF to CIFAR-10 image classification, Atari Ms Pacman, and GPT-2 language modeling.","tokens_in":18721,"tokens_out":5763,"duration_ms":70666,"significance":"If the theoretical and empirical claims were established, GPF would be a practically useful adaptive continual-learning wrapper for edge robotics, and the spectral-stability result would be a valuable extension of random matrix theory to growing depth. The paper deserves credit for releasing code and data, for using a clearly described plume simulator with a sensible reward design, and for attempting to connect architectural adaptation to an RMT analysis. However, as presented, the main theoretical claim is not established: Theorem 1 is mathematically misstated, and its assumptions are not connected to the actual GPF-grown, pruned, and frozen weight matrices. The central empirical claim rests on a single training run with no variance estimate or quantified baseline. The significance of the work is therefore contingent on substantial revision.","major_comments":[{"comment":"The hypothesis of Theorem 1 states that there exists an open set U ⊂ C\\R containing the spectral support of M^(ℓ). This is unsatisfiable for the Gram matrices M^(ℓ) defined in Eq. (5): these matrices are real symmetric and positive semidefinite, so their spectrum is contained in the real line, while C\\R excludes every real point. The theorem is therefore vacuous as stated. If the intended reading is that U is the domain of z for the Stieltjes transform, the statement must be rewritten, and the 'Consequently' clause about convergence of the empirical spectral distribution under a joint m, n, L limit does not follow from the fixed-point argument alone. This invalidates the abstract's claim that Pennington & Worah is extended to n-layer continual-learning models.","section":"Section 3.2, Theorem 1"},{"comment":"Theorem 1 assumes homogeneous layers ψℓ = Ψ and a uniform contraction bound ||D_sΨ|| ≤ ρ(z) < 1. The GPF procedure after pruning and freezing produces sparse, non-iid, layer-dependent weight matrices: Table 4 reports 47.9%, 19.6%, and 9.9% weight retention at successive prune passes, and different layers have different sparsity and activation patterns. The paper does not verify the contraction bound or homogeneity on any actual GPF-trained model, and Supplementary §11.1 explicitly defers the formal analysis of nonhomogeneous layers to future work. Consequently, the spectral fixed-point theorem is not connected to the architectures used in the experiments. The authors should either prove or numerically verify Conditions 1–2 on the GPF-trained matrices, or explicitly restrict the RMT claim to the idealized homogeneous setting and remove it from the general abstract claim.","section":"Section 3.2, Conditions 1–2; Supplementary §11.1"},{"comment":"The 94% held-out success rate is based on a single training run (environment seed 42, agent seed 43) and the best retained checkpoint (episode 3000). Table 4 shows substantial variance during training (e.g., 95.5% at episode 1000, 40.5% at episode 1500) and degradation to 68.0% by episode 5000 in the four-layer configuration. No mean or confidence interval over multiple seeds is reported, and the text never states the baseline success rate corresponding to the orange dotted line in Figure 3. With a single seed and checkpoint selection, the result cannot be separated from initialization luck. Multiple seeds, variance estimates, and a quantified baseline comparison are required to support the empirical claim.","section":"Section 5, Table 4"},{"comment":"The abstract states that further experiments 'amount evidence' that GPFs generalize to Atari, image classification, and language modeling, but the Supplementary Material does not report quantitative results for these tasks. For Atari, the text says 'Conclusions are difficult to draw'; for GPT-2, the GPF loss plot is referenced only as 'Figure??' and no numeric comparison to the baseline is given. Either report quantitative performance with baseline comparisons, or remove the generalization claim from the abstract.","section":"Supplementary §17"}],"minor_comments":[{"comment":"'Pennington & Worth' should be 'Pennington & Worah'; the same misspelling appears in the contribution list.","section":"Abstract"},{"comment":"Threshold notation is inconsistent: the text uses ω_l in Eq. (1) and surrounding discussion, while Table 3 and Supplementary Table 5 use ω_v. Please standardize.","section":"Section 2, Eq. (1); Table 3"},{"comment":"The algorithm title 'Grow-Pruen-Freeze Network' contains a typo: 'Pruen' should be 'Prune'.","section":"Algorithm 1"},{"comment":"'The bottom access indicates layer count' should be 'bottom axis'; 'Further experiments amount evidence' is ungrammatical and should be 'amount to evidence' or 'provide evidence'.","section":"Figure 3 caption; Abstract"},{"comment":"The GPF GPT-2 loss plot is referenced as 'Figure??'; the figure reference is unresolved. Also, 'Lewandowsi' in Section 12 is a typo for 'Lewandowski'.","section":"Supplementary §17.3"},{"comment":"The text switches to first person ('allows me implement a form of control'); please use impersonal style for consistency with the rest of the paper.","section":"Section 12"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is not ready for acceptance. The main theoretical result needs substantial reworking: Theorem 1 is mathematically misstated, and its assumptions are not verified on GPF-trained networks. The empirical evaluation also needs multiple seeds and a quantified baseline. I would consider a revised version if the authors can repair the theorem and either verify its conditions or substantially weaken the RMT claims, and if the navigation experiments are rerun with proper statistical controls."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Kordel and Daescu have written a paper about a genuinely new adaptive continual-learning scheme: GPF networks grow depth when validation loss plateaus, prune low-belief weights, and freeze stable layers. That scheduling combination, and the belief-based hardening idea, are the real contributions. The paper is also honest about its limits—the auxiliary experiments on CIFAR, Atari, and GPT-2 are explicitly exploratory, and the limitations section says more evaluation is needed. The plume navigation result (94% on held-out trials) is plausible: the simulator is described in enough detail to be reproducible, the failure modes are all timeouts, and the code is open source.\n\nBut the central theoretical claim does not hold up as written. Theorem 1 assumes an open set U ⊂ C\\R containing the spectral support of M^(ℓ). The spectrum of a Gram matrix is real, so that hypothesis can never be satisfied for any non-degenerate M^(ℓ). The theorem is vacuous as stated. Even if you fix the domain by using something like U ⊂ C\\R^+, the conditions require homogeneous layers and a uniform contraction bound that is never checked on the pruned, frozen, layer-dependent weight matrices that GPF actually produces. The supplementary remarks explicitly defer non-homogeneous layers to future work, so the 'preserved eigenvalue composition' claim is not proven by the theorem. The Marchenko-Pastur plots in the appendix are consistent with the idea, but they are empirical illustrations, not a proof.\n\nThe empirical section is also thinner than the abstract suggests. The headline numbers come from one training run with seeds 42 and 43, retaining the best checkpoint, with no error bars or variance estimate. The text says the baseline is in Figure 3 but never gives its numeric success rate. That is fixable—a few seeds and a table would do—but as it stands the 94% is a point estimate without uncertainty.\n\nNone of this is fatal in the sense that the framework is incoherent. The GPF dynamics are clearly defined, the pseudocode is precise, and the method is easy to implement. The authors are engaging seriously with the literature—the citations to LeCun, Kirkpatrick, Dohare, and Pennington & Worah are appropriate. The problem is that the paper overpromises on the RMT theory and under-delivers on the statistics.\n\nI'd send this to peer review with a clear message: fix or delete Theorem 1, verify or drop the spectral-invariance claim, and add multi-seed results with explicit baselines. The core idea is worth giving referee time to. My own verdict is conditional, not reject. I would not yet cite the theorem in my own work, but I would follow the code and the revision.","headline":"GPF is a promising adaptive-learning framework, but the central spectral theorem is vacuous as stated and the headline result lacks variance—worth refereeing, not yet citable.","tokens_in":19185,"tokens_out":2976,"would_cite":false,"duration_ms":30331,"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":"A network that grows, prunes, and freezes its own layers maintains its weight spectrum and finds turbulent odor sources in 94% of held-out trials.","keywords":["continual learning","random matrix theory","olfactory navigation","grow-prune-freeze networks","temporal difference learning","Expected SARSA","plume navigation","adaptive networks"],"falsifier":"Compute the empirical Stieltjes transform of the Gram matrix at each GPF layer before and after a grow-prune-freeze cycle, and measure the spectral radius of the Jacobian of the layer-to-layer map; if that radius is not < 1 on the learned sparse weights—or if the empirical spectral distribution shifts materially when a layer is added—then eigenvalue composition is not preserved and the central claim fails.","tokens_in":18227,"feed_emoji":"🤖","tokens_out":12968,"duration_ms":115416,"temperature":0.7,"pith_summary":"The paper introduces Grow-Prune-Freeze (GPF) networks, a continual-learning architecture that adds hidden layers when validation loss stagnates, prunes low-salience neurons by magnitude and belief, and freezes stable layers to prevent catastrophic forgetting. Its central theoretical claim is that a known single-layer result from nonlinear random matrix theory extends to a growing, indefinitely deep network: the eigenvalue composition of the weight Gram matrices is preserved as layers are added, because the layer-to-layer Stieltjes transform composes into a contraction with a unique fixed point. If that claim holds, a robot learning in a partially observable, non-stationary environment such as a turbulent odor plume can adapt its network capacity online without a separate training/test phase and without collapsing its learned representations. The accompanying experiments show an Expected SARSA GPF agent reaching 94% held-out source-finding success in a turbulent plume simulator, with additional evidence on image classification, Atari, and language modeling.","feed_headline":"Grow, prune, freeze: adaptive network finds odor sources 94%","feed_subtitle":"Self-adapting network adds, prunes, and freezes layers while learning; finds odor source in 94% of held-out trials.","key_machinery":"The load-bearing object is the Stieltjes transform s^(ℓ)(z) of the empirical Gram matrix M^(ℓ) of each hidden layer, together with the layer-wise nonlinear spectral transformer Ψ_ℓ that maps the previous layer's Stieltjes transform to the next. In the theory, adding depth composes these operators into Ψ^(L:1) = Ψ_L ∘ ⋯ ∘ Ψ_1, and Theorem 1 shows that if Ψ is an analytic contraction (Fréchet derivative bounded by ρ(z) < 1), the iterated transform reaches a unique fixed point, so the eigenvalue composition of network weights is preserved as layers are added. The GPF mechanism itself—belief values that harden frequently used neurons, a stagnation threshold for growth, magnitude-based pruning, a","core_discovery":"The paper's central claim is that the grow-prune-freeze cycle—grow a layer when validation loss plateaus, prune 'soft' neurons below a belief threshold, freeze layers whose weights have stabilized—lets a policy network adapt indefinitely while preserving the spectral distribution of its weights. The theoretical result (Theorem 1) states that for homogeneous layers whose nonlinear spectral transformer Ψ is an analytic contraction, the Stieltjes transform of the layer-ℓ Gram matrix is obtained by composing Ψ ℓ times, and the iteration converges to a unique fixed point; hence the empirical spectral distribution of the weight Gram matrices converges as depth, width, and data grow jointly. This e","pith_inferences":["One testable extension is to track the empirical Stieltjes transform of a GPF's Gram matrices across a grow event; if the spectral radius of the layer-to-layer map exceeds 1 after pruning and freezing, the theory would need a heterogeneous-layer version to cover the actual agent.","The recursive single-hidden-layer variant sketched in the paper's limitations—reusing one layer's weights with index-dependent loading—could reduce model size on edge robots even further; this is the paper's ongoing work, not a demonstrated result.","The 94% figure comes from one simulator and one sensor model; a natural extension is to ask whether the same GPF success survives real chemical plumes, or whether the freshness of the belief-hardening signal depends on the plume's intermittency statistics.","If the spectral-preservation claim carries over to non-homogeneous layers, it would provide a principled criterion for when to grow: add capacity when the current layer's Stieltjes map approaches a fixed point, rather than on a fixed loss-stagnation schedule."],"forward_implications":["If the spectral-stability claim is correct, a GPF agent can grow its policy network during deployment and retain the eigenvalue structure that supports stable gradient propagation, directly addressing plasticity loss in continual learning.","The 94% held-out success on turbulent plume navigation indicates that an edge-deployable, adaptively growing Expected SARSA policy is a practical candidate for robotic olfactory search under sparse, intermittent, non-stationary signals.","The same grow-prune-freeze cycle applied to CIFAR-10, Atari Ms Pacman, and a GPT-2-style model shows lower validation volatility and stable loss decay relative to static baselines, supporting GPF as a general continual-learning wrapper rather than a plume-specific trick.","Because pruning and freezing selectively remove low-information folds while preserving hardened neurons, GPFs can compress a capable policy into a sparse four-layer network, cutting onboard memory and compute for robot inference.","The theoretical extension from one hidden layer to L layers implies that depth can be added incrementally—rather than retraining from scratch—as long as layer homogeneity and the contraction condition hold."],"fun_headline_variants":["Adaptive grow-prune-freeze nets hit 94% on odor tracking","Odor-seeking AI learns continuously via grow-prune-freeze","Grow, prune, freeze: continual learning for smell navigation","Random matrix theory backs adaptive odor-tracking nets"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proof of the spectral-stability claim assumes every layer is generated by the same random-matrix law and that the layer-to-layer spectral transformer is a strict contraction, but the paper does not verify those conditions on the actual pruned, frozen, non-iid weight matrices that GPF produces.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive grow-prune-freeze nets hit 94% on odor tracking","Odor-seeking AI learns continuously via grow-prune-freeze","Grow, prune, freeze: continual learning for smell navigation","Random matrix theory backs adaptive odor-tracking nets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000262,"raw_usage":{"total_tokens":1438,"prompt_tokens":756,"completion_tokens":682,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":623}},"tokens_in":500,"tokens_out":682,"duration_ms":6685,"temperature":1.0,"reasoning_tokens":623,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T13:10:48.515877+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the empirical Stieltjes transform of the Gram matrix at each GPF layer before and after a grow-prune-freeze cycle, and measure the spectral radius of the Jacobian of the layer-to-layer map; if that radius is not < 1 on the learned sparse weights—or if the empirical spectral distribution shifts materially when a layer is added—then eigenvalue composition is not preserved and the central claim fails.","supporting_citations":[],"review_version":2}