{"id":"1a9f3253-1162-40c0-9b8f-1246c8ee123b","arxiv_id":"2508.18954","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Koopman embeddings pretrained on next-state prediction transfer to safety-function prediction on the Lorenz system, outperforming PCA baselines and retaining performance with a frozen transformer.","lead":"This paper tests whether an embedding learned from predicting the next state of the chaotic Lorenz system can be reused for a different task: estimating the control effort needed to keep the system safe. It finds that Koopman-based embeddings transfer better than PCA-style baselines, and that freezing the pretrained model does not hurt performance.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Safety function depends only on query state q; transformer context encodes only past states, so frozen vs fine-tuned comparison does not test representation transfer.","rationale":"The reader's weakest_assumption focused on the underpowered Wilcoxon test, but the more fundamental issue is that the chosen downstream task cannot probe representation transfer at all. The safety function U(q) is a pure function of the current state, and the model is given q directly, so the transformer's representation of past states is causally irrelevant. Thus, even a perfectly powered statistical test would not validate the central claim; the null result is expected regardless of whether the representations are reusable. This invalidates the experiment as a test of the stated hypothesis. I recommend UNVERDICTED because the paper's strongest claim is not supported by the current experimental design, though the paper may have other merits. The reader's conditional verdict was too generous in treating the p-value as the main weakness. The concrete ablation test would settle the issue by showing whether the transformer contributes anything to Task B; if it does not, the authors must redesign the downstream task to actually depend on the learned dynamics representation.","tokens_in":16713,"tokens_out":6284,"duration_ms":59178,"concrete_test":"Retrain the Task B safety head with the transformer context ablated: replace the input sequence with zeros (or a fixed random sequence) while keeping the same q input, for both Koopman (F) and Koopman (U). If performance on the 252 test trajectories remains statistically unchanged (e.g., Wilcoxon p > 0.05) relative to the full models, the frozen-vs-fine-tuned result does not depend on the pre-trained representation. Additionally, train a head on q alone (3D-to-1D) as a control; if it matches Koopman (F), the claimed transfer advantage is an artifact of the task design.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The central claim that freezing pre-trained transformer weights causes no degradation presupposes that the downstream task actually requires the pre-trained representation. In Task B (Section 3.3), the safety head receives the concatenation of the transformer's final hidden state (encoding past states) and the raw query state q. The safety function U(q) (Eq. 3, Section 2.3) is defined as the worst-case control effort to keep a trajectory starting at q inside Q; the Lorenz system is Markovian, so U(q) depends only on q, not on any past trajectory. Consequently, the transformer's context vector carries no information about the target, and the head could solve the task from q alone. The frozen/unfrozen comparison (p=0.335, Table 2) therefore shows only that the transformer is irrelevant to this task, not that its representations are reusable. This also confounds the Koopman-vs-PCA comparisons: all models can ignore their sequence embedding and regress q to U(q); observed differences may stem from head capacity (32 vs 9 vs 3 embedding dimensions) rather than embedding quality. To support the transfer claim, the downstream task must either depend on history or feed q through the learned representation (e.g., phi_e(q)) rather than as raw input.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a three-stage transfer-learning pipeline for chaotic systems: learn a Koopman autoencoder on Lorenz trajectories, pre-train a transformer for next-state prediction in the Koopman embedding space, then fine-tune (or freeze) the transformer for a safety-function regression task. The authors report that their Koopman-based representations outperform PCA-based baselines and that freezing the pre-trained transformer causes no significant degradation, which they interpret as evidence that the learned representations capture reusable dynamical structure. The paper includes detailed hyperparameter and statistical appendices, paired Wilcoxon tests with Bonferroni correction, and a public project page.","tokens_in":17009,"tokens_out":6481,"duration_ms":60240,"significance":"If the central claim were established, the paper would be a useful step toward foundation-model-style transfer in physics-informed ML, showing that a self-supervised Koopman representation can support a downstream control-oriented task without retraining the backbone. Strengths of the manuscript include the physically structured Koopman operator (diagonal plus banded skew-symmetric), the clear three-stage methodology, and the careful paired statistical testing. However, as detailed in the major comments, the downstream task as designed does not actually test representation transfer: the safety function depends only on the raw query state, which is provided directly to the head. Consequently the headline claims about reusable representations, and the comparisons against PCA baselines, are not currently supported. The issues are identifiable and, in principle, fixable with additional control experiments, so the contribution is conditional rather than conclusive.","major_comments":[{"comment":"The downstream task does not test representation transfer. U(q) in Eq. (3) depends only on the query state q, and the Lorenz dynamics are Markovian. In Stage 3, the head receives the transformer's final hidden state (encoding past states) concatenated with the raw query q, so the target is computable from q alone without using any information from the transformer context. The frozen-vs-fine-tuned comparison in Table 2 (p=0.335) therefore only shows that an irrelevant frozen context does not hurt; it cannot establish that Koopman representations are reusable. To make the claim testable, add a q-only head (no transformer context), or feed q through the learned encoder phi_e(q) rather than as raw input, or use a downstream task that genuinely depends on history.","section":"Sec. 3.3, Fig. 3, Eq. (3)"},{"comment":"The standard PCA baseline failed to converge on Task A (Table B.1, 'Converged? No') yet is retained as a Task B comparator. Its poor downstream performance could be due to an untrained transformer backbone or smaller input/head capacity, not to embedding quality. The head input dimensions differ across models (35 for Koopman, 12 for PCA-PI, 6 for PCA), and the safety-head widths also differ, so the observed performance gaps in Table 1 are confounded by capacity. Please match head capacities across models, include a PCA representation that successfully trains on Task A, or explicitly analyze these confounds.","section":"Sec. 3.4, Table B.1"},{"comment":"The conclusion of 'no degradation' treats a non-significant Wilcoxon p-value as evidence of equivalence. With 252 test trajectories and per-model MSE standard deviations that are large relative to the mean differences (Koopman (F): 3.08±6.55; Koopman (U): 5.59±17.14), the test is likely underpowered. Please report a confidence interval for the difference and an equivalence test (e.g., two one-sided tests or a bootstrap CI) with a pre-specified equivalence bound, or weaken the wording to 'no statistically significant difference detected.'","section":"Table 2, Appendix B.2"},{"comment":"The Koopman (F) and Koopman (U) fine-tuning settings are not matched. They differ in learning rate (6.83e-3 vs 1.04e-3), number of epochs (80 vs 50), optimizer (Adam vs AdamW), and safety-head width (128 vs 112). Thus the frozen-vs-fine-tuned comparison is confounded by multiple hyperparameter changes, not just the freeze flag. Use identical training configurations except for the freeze flag, or explicitly analyze the confounding effects.","section":"Table A.3"}],"minor_comments":[{"comment":"The Stage 2 sequence length is inconsistent: Section 3.3 says '64-step non-overlapping sequences' while Appendix A.6 says 'non-overlapping 256-step sequences'; Table A.2 lists context length 64. This ambiguity affects reproducibility.","section":"Sec. 3.3 vs Appendix A.6"},{"comment":"The caption mentions 'Region R Boundary (x = 0)' but the text uses region Q throughout; unify the notation.","section":"Figure 4 caption"},{"comment":"The word 'dissaptivity' appears to be a typo for 'dissipativity'.","section":"Section 4.2"},{"comment":"The author names are spelled inconsistently (Capeáns vs Capeans, Sanjuán vs Sanjuan); please standardize.","section":"References [11,12]"},{"comment":"The footnote clarifies that 'control' is used by convention, but the paper should state explicitly that no control inputs are applied in the experiments; the safety function is a preparatory component.","section":"Section 2.3 footnote"}],"recommendation":"major_revision","confidential_remarks":"The core concern in the reader's report is correct and, in my assessment, decisive for the paper's central claim. The downstream safety task is Markovian in the raw query state, so the frozen-vs-fine-tuned experiment cannot support the abstract's claim about reusable dynamical structure. The PCA comparison is further weakened by the failed convergence of the standard PCA baseline and by mismatched head capacities. These are not merely presentation issues: they affect the main conclusion. I recommend major revision rather than rejection because the flaws are identifiable and fixable with additional experiments (q-only baseline, phi_e(q) input, matched hyperparameters, equivalence testing). If the authors cannot add such experiments, the representation-transfer claim should be removed or substantially weakened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Take a quick look at the Task B setup before reading the fine print: the safety head receives the transformer's final hidden state concatenated with the raw query state q. Since U(q) is defined by the Markovian Lorenz dynamics, q alone determines the target. The past trajectory adds nothing. So when the paper reports that freezing the transformer weights makes no difference (p=0.335), the most plausible reading is that the context vector is simply not used—the head can regress q to U(q) regardless. That is not a test of representation transfer; it's a test of whether the head can ignore the transformer. The stress-test note is right.\n\nThat said, the paper has genuine merits. The three-stage pipeline is spelled out carefully, the hyperparameter and data-generation details are in the appendices, and the statistical testing is solid (paired Wilcoxon with Bonferroni). The authors also honestly admit that the plain PCA baseline failed to converge. The architecture—Koopman autoencoder with banded skew-symmetric operator, transformer pretraining, then fine-tuning—is a reasonable assembly of existing ideas.\n\nThe problems are concentrated in interpretation. First, the frozen-vs-fine-tuned equality is over-interpreted as evidence of reusable structure; as argued above, the null result is expected if the target ignores history. Second, the Koopman-vs-PCA performance gaps are confounded by embedding dimension: the Koopman head sees 35 inputs, PCA(PI) sees 12, PCA sees 6. A head that receives more irrelevant inputs might do worse if it can't fully ignore them, so the 'Koopman advantage' could be a capacity or regularization effect, not a representation effect. The authors do not ablate a head that takes only q. Third, the abstract claims 'data-efficient' performance, but no data-scaling study appears in the paper.\n\nNone of this means the paper is worthless. It is a clear, honest empirical study of a sensible idea, and the frozen-backbone variant is a useful design to test on a task that actually requires temporal context. The error maps are a nice diagnostic. But the central claim, as written, does not hold up. I'd send it to a good referee—the authors deserve a chance to redesign the downstream task, e.g., feeding phi_e(q) or using a partially observed variant where history matters. As is, it's a conditional reject with a clear route to revision.","headline":"Task design sinks the transfer claim: the safety function depends only on q, so the frozen vs fine-tuned result shows the transformer context is irrelevant, not that the representation is reusable.","tokens_in":17486,"tokens_out":5431,"would_cite":false,"duration_ms":49665,"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":"This paper claims that a Koopman embedding of a chaotic system, learned once for next-state prediction, can be reused for a downstream safety-control task with no loss of accuracy when the pretrained transformer is frozen.","keywords":["Koopman operator","chaotic dynamical systems","Lorenz system","transfer learning","transformer","safety function","representation learning","physics-informed machine learning"],"falsifier":"Give the frozen and fine-tuned backbones a much larger evaluation set (thousands of test trajectories, multiple seeds) and compute a confidence interval for the difference in safety-function MSE. If the interval excludes zero or crosses a pre-specified equivalence bound, the no-degradation claim fails. A second check: replace the Koopman encoder with an unconstrained autoencoder of the same capacity; if it transfers equally well, the Koopman structure is not the cause.","tokens_in":16666,"feed_emoji":"🦋","tokens_out":4365,"duration_ms":39602,"temperature":0.7,"pith_summary":"This paper argues that a Koopman embedding of a chaotic system, learned once from next-state prediction, carries enough dynamical structure to support a different downstream task—estimating a global safety function for control—without retraining the sequence model. The authors test this on the Lorenz system through a three-stage pipeline: a Koopman autoencoder creates a latent space where dynamics are approximately linear; a transformer is pretrained to predict next states in that space; and only a small safety head is trained on top of the frozen transformer. They report that this frozen configuration matches the fine-tuned one, and that both beat PCA-based embeddings on safety-function accuracy. The point of the claim is that physics-informed representations can be amortized across tasks, much like pretrained language models are reused in NLP.","feed_headline":"Frozen transformer matches fine-tuning on chaotic safety task","feed_subtitle":"Koopman embeddings pretrained on next-state prediction generalise to safety-function control on the Lorenz system.","key_machinery":"The load-bearing object is the learned Koopman operator K, a 32-by-32 matrix decomposed into a diagonal part D (growth and decay) plus a banded skew-symmetric part S_band (rotation), so that the encoder maps a state s_t to a latent y_t and the next latent is K y_t. This linearising structure is what makes the transformer's next-state pretraining capture durable dynamical content; at transfer time the encoder and transformer are frozen and only a multilayer safety head takes the transformer's final hidden state concatenated with the query state to predict U(q).","core_discovery":"On the Lorenz system, embeddings produced by a structured Koopman autoencoder transfer from short-horizon next-state prediction to a long-horizon safety function that quantifies minimum control effort to stay in a safe region. The paper's central quantitative result is that a transformer frozen after pretraining and a transformer fine-tuned on the safety task perform indistinguishably (Wilcoxon p = 0.335), while both significantly outperform standard PCA and physics-informed PCA baselines. The authors interpret this as evidence that the Koopman latent space encodes reusable dynamical structure—the attractor's geometry and long-term behaviour—rather than task-specific patterns learned for pre","pith_inferences":["The equal-frozen claim rests on accepting a null result; a higher-powered test with more trajectories or repeated seeds and pre-specified equivalence bounds could reveal a real degradation, since the paper does not report such a test.","Because the safety region Q covers only the right lobe of the attractor, the measured transfer may be specific to that region; testing the left lobe or the full attractor would show whether the embedding's geometry generalises globally.","The comparison to PCA may be generous to Koopman: the physics-informed PCA baseline includes time derivatives, which the paper itself notes can produce phantom oscillations, so a stronger baseline would be an unconstrained nonlinear autoencoder without the Koopman constraint.","An untested extension is whether the same frozen embedding transfers across different chaotic systems or across parameter regimes; if it does, the amortisation argument becomes much stronger."],"forward_implications":["If true, a single pretrained Koopman backbone can serve multiple downstream tasks on the same system, so the cost of learning a physics-informed representation is paid once.","Frozen-backbone transfer means safety heads can be trained cheaply and with less data; the authors report a 4-layer frozen transformer uses less than half the peak GPU power of the 11-layer PCA baseline during fine-tuning.","Koopman embeddings should be preferred over PCA-derived features when downstream tasks require global, long-horizon understanding rather than local reconstruction.","The result supports treating next-state prediction as a viable pretext task for physical systems, analogous to next-token prediction in language models.","The comparison validates the structured Koopman operator design (diagonal plus skew-symmetric bands) as the source of the advantage, rather than the transformer architecture alone."],"supporting_citations":[{"why":"Introduces the Koopman operator that this work approximates as a linearising embedding.","marker":"[2]"},{"why":"Establishes applied Koopman theory, motivating the claim that linear latent dynamics capture nonlinear structure.","marker":"[3]"},{"why":"Supplies the structured Koopman autoencoder design (diagonal plus banded skew-symmetric operator) and the transformer pretraining protocol for physical systems.","marker":"[4]"},{"why":"Provides the NLP pretraining-then-fine-tuning paradigm the three-stage pipeline is modelled on.","marker":"[7]"},{"why":"Defines the sculpting algorithm used to compute ground-truth safety values in the discretised state space.","marker":"[10]"},{"why":"Extends partial-control safety theory to the Lorenz equations, grounding Task B.","marker":"[11]"},{"why":"Frames the safety-function prediction task and transformer-based control for Lorenz, the direct baseline this work extends.","marker":"[12]"},{"why":"Documents phantom oscillations in PCA, the artefact the authors invoke to interpret PCA baseline behaviour.","marker":"[25]"}],"fun_headline_variants":["Koopman embeddings transfer from prediction to chaotic control","Frozen transformer equals fine-tuning on chaotic safety task","Pretrained Koopman latents generalize to Lorenz safety control","Same Koopman representation works for prediction and control","Chaos control via Koopman: frozen transformer matches fine-tuning"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The claim that freezing the transformer costs nothing rests on treating a non-significant Wilcoxon p-value (0.335) as evidence of equivalence; with 252 trajectories and error variances comparable to the observed differences, the test may simply lack the power to detect a practically meaningful drop.","fun_headline_variants_meta":{"raw":{"variants":["Koopman embeddings transfer from prediction to chaotic control","Frozen transformer equals fine-tuning on chaotic safety task","Pretrained Koopman latents generalize to Lorenz safety control","Same Koopman representation works for prediction and control","Chaos control via Koopman: frozen transformer matches fine-tuning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000225,"raw_usage":{"total_tokens":1256,"prompt_tokens":655,"completion_tokens":601,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":399,"completion_tokens_details":{"reasoning_tokens":520}},"tokens_in":399,"tokens_out":601,"duration_ms":5960,"temperature":1.0,"reasoning_tokens":520,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:02:47.398700+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Give the frozen and fine-tuned backbones a much larger evaluation set (thousands of test trajectories, multiple seeds) and compute a confidence interval for the difference in safety-function MSE. If the interval excludes zero or crosses a pre-specified equivalence bound, the no-degradation claim fails. A second check: replace the Koopman encoder with an unconstrained autoencoder of the same capacity; if it transfers equally well, the Koopman structure is not the cause.","supporting_citations":[{"cited_title":"Hamiltonian Systems and Transformation in Hilbert Space","cited_arxiv_id":null,"evidence_quote":"Introduces the Koopman operator that this work approximates as a linearising embedding."},{"cited_title":"Applied koopmanism","cited_arxiv_id":null,"evidence_quote":"Establishes applied Koopman theory, motivating the claim that linear latent dynamics capture nonlinear structure."},{"cited_title":"Transformers for modeling physical systems","cited_arxiv_id":null,"evidence_quote":"Supplies the structured Koopman autoencoder design (diagonal plus banded skew-symmetric operator) and the transformer pretraining protocol for physical systems."},{"cited_title":"Dynamics of partial control","cited_arxiv_id":null,"evidence_quote":"Defines the sculpting algorithm used to compute ground-truth safety values in the discretised state space."},{"cited_title":"Partially controlling transient chaos in the Lorenz equations","cited_arxiv_id":null,"evidence_quote":"Extends partial-control safety theory to the Lorenz equations, grounding Task B."},{"cited_title":"AI-driven control of chaos: A transformer-based approach for dynami- cal systems","cited_arxiv_id":null,"evidence_quote":"Frames the safety-function prediction task and transformer-based control for Lorenz, the direct baseline this work extends."},{"cited_title":"Phantom oscillations in principal component analysis","cited_arxiv_id":null,"evidence_quote":"Documents phantom oscillations in PCA, the artefact the authors invoke to interpret PCA baseline behaviour."}],"review_version":1}