{"id":"91e72eb3-ee22-4675-b10c-1c6d1e6b5d17","arxiv_id":"2607.17526","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"FlowSonic combines deterministic rectified-flow inversion, cached cross-attention injection, and a 'seeded' third-order Adams-Bashforth solver to report better timbre and genre edits on small datasets.","lead":"FlowSonic edits real music recordings by inverting them into a music-generation model's latent space, injecting the original's attention patterns during generation, and using a third-order ODE solver seeded with inversion velocities. The paper reports improved timbre- and genre-edit quality over existing zero-shot baselines, but the solver's mathematical justification has a gap.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cached inversion velocities are not valid AB3 history: Eq. (14) mixes source-prompt, reverse-time derivatives into the target-prompt forward step, so the claimed third-order accuracy is unsupported.","rationale":"After reading the manuscript, the load-bearing claim is that Seeded-AB3 with DHC provides third-order accuracy without warm-start. The key equation, Eq. (14), reuses the last two inversion velocity evaluations as f_{-1}, f_{-2}. For a genuine Adams–Bashforth method, these must be evaluations of the ODE's vector field at t_{n}-h and t_{n}-2h on the same integral curve being advanced. During generation, the vector field is conditioned on the target prompt with CFG>1; during inversion, it is conditioned on the source prompt with CFG=1, and the times correspond to points just after t=0 on the reverse curve. Thus the consistency conditions for AB3 are violated. The observed improvements in Tables I–IV therefore cannot be attributed to 'higher-order trajectory integration' as claimed. The empirical evaluation (40 clips, no error bars, subjective test with 20 participants) is too weak to independently support the numerical claim. A proper warm-start comparison would isolate whether DHC provides any benefit, and a local-truncation-error measurement would settle the order. For these reasons I agree with the reader's REJECT verdict; the paper would need a corrected numerical analysis or a softened claim to become acceptable.","tokens_in":22111,"tokens_out":6413,"duration_ms":54712,"concrete_test":"Run the timbre-transfer experiment (40 clips) twice with identical hyperparameters and CFG: (A) the paper's Seeded-AB3 as in Eq. (14), and (B) a proper warm-start where the first step uses Heun's method (second-order) followed by standard AB3 with the same NFE. Compare CLAP, Chroma, CQT-PCC, and FAD per clip with a paired test. If (A) and (B) are statistically indistinguishable, the DHC history is not responsible for the reported improvements and the high-order consistency claim is empirically unsupported. If (A) is better, the gain must be attributed to the specific cached values, not to AB3 order; the authors would need to explain why non-AB3 seeding helps.","verdict_should_be":"REJECT","load_bearing_attack":"Section III-C presents DHC in Eq. (14): z1 = z0 + Δt[(23/12)v0 − (16/12)v_K_inv + (5/12)v_{K−1}_inv]. A third-order Adams–Bashforth step requires the three latest derivatives evaluated on the solution curve being integrated, at times t0, t0−h, t0−2h. The cached v_K_inv, v_{K−1}_inv are evaluations from the source-prompt inversion path, at times just after 0 (i.e., positive time points on the reverse direction), under a vector field with source CFG=1, whereas the generation step uses target-prompt CFG (5–15). These are not evaluations of the same vector field at preceding points on the generation trajectory. Consequently Eq. (14) is not an AB3 step; it is an inconsistent one-step update with an error that need not scale as h^4. The paper's headline claim that DHC 'eliminates lower-order warm-start initialization' while preserving third-order accuracy is therefore a claim-without-derivation. This is load-bearing because the numerical solver is the main novel contribution; the attention injection is a known adaptation from image editing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents FlowSonic, a zero-shot text-guided music editing framework built on a pretrained rectified-flow transformer (FluxMusic). The method deterministically inverts a real-world recording into the latent space, preserves musical structure by reusing cached cross-attention key/value representations during generation, and introduces a 'Seeded-AB3' solver with Dynamic History Caching (DHC). DHC initializes the third-order Adams–Bashforth integrator using velocity evaluations obtained during inversion, allegedly avoiding lower-order warm-start steps and improving numerical stability. The paper reports experiments on timbre-transfer and genre-transfer tasks, comparing against AudioLDM2, MusicGen, ZETA, and a FluxMusic editing baseline, and claims consistent improvements in semantic alignment, structural preservation, and perceptual quality. Code is released publicly.","tokens_in":22284,"tokens_out":7113,"duration_ms":62137,"significance":"If the numerical claims were sound, DHC would be an attractive, training-free way to initialize high-order multi-step solvers in inversion-based editing, with no extra network evaluations. The paper also contributes a systematic comparison of integration schemes and a reproducible codebase. However, the central theoretical claim—that Eq. (14) is a third-order AB3 step with O(h^4) local error—is invalid because the cached velocities come from a different vector field and a reversed integration path. This flaw undermines the main novel contribution, leaving the attention-injection mechanism as a known adaptation from image editing. The empirical support is further weakened by a per-solver CFG equalization protocol, a small subjective study, and an overclaim contradicted by the paper's own FAD results.","major_comments":[{"comment":"The update z1 = z0 + Δt[(23/12)v0 − (16/12)v_K_inv + (5/12)v_{K−1}_inv] is not a third-order Adams–Bashforth step. AB3 requires the three latest derivatives evaluated at preceding points on the same solution curve (t0, t0−h, t0−2h) of the same vector field. The cached v_K_inv and v_{K−1}_inv are computed on the source-prompt inversion path in the reverse direction with source CFG=1, not on the target-prompt forward trajectory. The local truncation error O(h^4) therefore does not follow, and the claim that DHC 'preserves third-order accuracy' is unsupported. Since the solver is the main novel contribution, this is load-bearing. The authors should either provide a correct multi-step error analysis for the proposed update or remove the order guarantees and reframe DHC as a heuristic.","section":"§III-C, Eq. (14)"},{"comment":"The Equitonal Transfer protocol adjusts the CFG scale per solver so that all configurations achieve 'comparable' CLAP similarity. This makes the CLAP differences in Tables I and II uninformative by construction and introduces a per-solver free parameter that is not reported. Without the adjusted CFG values, the comparison is not reproducible, and the claim that Seeded AB3 achieves the 'highest CLAP similarity' (Table I, 0.238 vs 0.236) is not meaningful evidence. Please report the CFG values used for each solver and analyze sensitivity, or use a protocol that does not equalize the headline semantic metric.","section":"§IV / §V-B, Equitonal Transfer"},{"comment":"The claim that FlowSonic 'consistently outperforms existing music editing methods across ... perceptual audio quality' is contradicted by the paper's own FAD results: AudioLDM2 achieves the lowest FAD (3.623) on the timbre-transfer task, beating FlowSonic KV (3.938) and V (3.887). The text acknowledges AudioLDM2's low FAD, but the abstract and conclusion still assert consistent superiority. Please qualify the claims and discuss why the proposed method trails on this metric.","section":"Abstract / §V-C, Table III"},{"comment":"The subjective evaluation uses 20 participants, each rating one timbre and one genre example per condition, and reports only mean scores without variance or significance testing. Differences such as MOS-T 4.00 vs 4.10 are within sampling noise. The claims of 'dramatic improvement' and 'consistently' are not statistically supported. Please add more participants/items, report score distributions, and use inferential statistics (e.g., paired tests with multiple-comparison correction).","section":"§V-D and Appendix A"}],"minor_comments":[{"comment":"The notation t_{−1}, t_{−2} and the integration grid are undefined. Please specify the grid spacing and the exact time indices of the cached inversion velocities used in Eq. (14).","section":"§III-C, Eq. (12)"},{"comment":"The PCA components explain only 8.1%/3.7% and 11.5%/4.4% of the variance, respectively. Please report the total explained variance and note that two-dimensional projections may not faithfully represent trajectory stability.","section":"Figures 2 and 11"},{"comment":"The caption describes a comparison of numerical integration strategies, but the text says the figure examines cross-attention feature injection. Please align the caption with the actual content.","section":"§V-A, Figure 5"},{"comment":"References [20] and [34] are cited for supervised editing and MusicGen, but [20] is M2UGen and [34] is MusicGen-stem. Please ensure the citations match the statements they support.","section":"References"}],"recommendation":"reject","confidential_remarks":"The manuscript's central numerical contribution rests on an incorrect identification of cached inversion velocities as valid Adams–Bashforth history. This is not a minor presentation issue: it invalidates the stated high-order accuracy guarantee. The empirical evaluation, while extensive in scope, is underpowered and contains an overclaim contradicted by its own FAD results. Although the authors release code and compare several solvers systematically, the core technical flaw and the weak statistical support justify rejection. If the authors substantially rewrite the method as a heuristic without order guarantees and strengthen the evaluation, a future submission could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things before reading. First, the paper's central numerical contribution—seeding Adams–Bashforth with cached inversion velocities—is not actually AB3. Eq. (14) mixes velocities from the source-prompt inversion path into the target-prompt forward step. AB3's order guarantee requires a consistent history from the same vector field on the same trajectory. The inversion path uses a different prompt, a different CFG scale, and the opposite time direction, so the cached values are not valid multistep history. The paper quietly assumes this away and claims third-order accuracy without a derivation. That is load-bearing, not cosmetic. The title itself promises high-order trajectory integration.\n\nSecond, the paper still has genuine value. The observation that inversion already computes velocities you could reuse to skip the warm-start phase is new, and the empirical sweep across Euler, midpoint, Heun, and AB3 in a real music-editing setting is useful. The attention-injection part is a straightforward adaptation of prompt-to-prompt and plug-and-play feature injection, but it is executed carefully and the ablations distinguish K, V, and KV injection sensibly. Code is released. The authors also admit in the appendix that Seeded AB3 and unseeded AB3 trajectories nearly overlap in the genre task, which is more honest than most papers.\n\nNow the soft spots, in proportion. The evaluation is thin: 40 clips per task, no error bars, and the Equitonal Transfer protocol tunes CFG per solver to equalize CLAP similarity—so the primary semantic metric is effectively held constant by construction. The subjective study is 20 participants, each hearing one example per task. The FluxMusic baseline scores around 1.0 MOS, which is suspicious enough to make me question the baseline implementation, though the gap to the proposed method is so large that some real effect probably exists. The DHC improvement might be a form of regularization or feature-matching rather than higher-order accuracy. The paper never tests that distinction.\n\nWho is this for? Anyone working on zero-shot audio editing or on numerical integrators for flow models. It deserves a serious referee, not a desk reject: the theoretical flaw is central but clearly fixable by reframing DHC as an empirically motivated initialization or by proving a weaker consistency result. The empirical comparison, despite its weaknesses, is informative enough to warrant revision. Recommend the editor send it out, with instructions to make the authors either prove the order claim or drop it and reposition the contribution as a stable warm-start heuristic, and to add error bars and more evaluation clips.","headline":"The AB3 claim doesn't survive contact with Eq. (14), but the paper is a useful empirical study of a possibly-effective heuristic—worth refereeing if the authors can reposition it honestly.","tokens_in":22903,"tokens_out":2443,"would_cite":false,"duration_ms":25222,"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":"Zero-shot editing of real music recordings is made practical by reusing inversion velocities to seed a high-order ODE solver.","keywords":["music editing","rectified flow","zero-shot","ODE solver","Adams-Bashforth","dynamic history caching","cross-attention injection"],"falsifier":"Run the editing pipeline with the cached history taken from a different recording's inversion, or with the two cached velocities exchanged in time order; if edit quality and structural preservation do not degrade, the claimed trajectory-specific benefit of Dynamic History Caching is not the operative mechanism.","tokens_in":21824,"feed_emoji":"🎵","tokens_out":5948,"duration_ms":49566,"temperature":0.7,"pith_summary":"The paper sets out to show that a pretrained rectified-flow text-to-music model can edit real-world recordings — changing timbre or genre — without any training, fine-tuning, or per-sample optimization. The route is a two-stage process: deterministic inversion of the audio into the model's latent space, with cross-attention features and velocity evaluations cached along the way; then generation under a new prompt that reuses the cached attention features to preserve structure and seeds a third-order Adams-Bashforth solver with the cached inversion velocities via Dynamic History Caching. The paper argues that this seeding removes a numerical discontinuity at the inversion-generation boundary, allowing high-order integration to start at the very first step. Reported results show consistent gains over existing zero-shot editing methods across semantic alignment, harmonic preservation, structural consistency, and perceptual quality, and the whole recipe requires no model modification or extra optimization.","feed_headline":"Inversion velocities seed a high-order solver for stable music edits","feed_subtitle":"Reusing the inversion's velocity history avoids warm-start instability and preserves the original musical structure.","key_machinery":"The two load-bearing mechanisms are the third-order Adams-Bashforth multi-step ODE integrator (AB3), which computes the next latent state from the current velocity plus two previous velocities, and Dynamic History Caching (DHC), which supplies those previous velocities from the inversion stage instead of a lower-order warm-start. DHC makes the first generation step third-order accurate and removes the startup asymmetry between inversion and generation. Structural preservation is carried separately by reusing cached key-value cross-attention tensors during generation, with full key-value replacement giving the best balance of fidelity and editability.","core_discovery":"The central claim is that the failure mode of inversion-based rectified-flow editing is largely a startup problem in numerical integration. Multi-step solvers like third-order Adams-Bashforth need derivative history that does not exist at the beginning of generation, so conventional pipelines fall back to lower-order warm-start steps; this introduces a 'Multi-Step Startup Asymmetry' whose errors propagate through the entire trajectory. FlowSonic's Dynamic History Caching closes that gap by taking the final two velocity evaluations from the deterministic inversion and feeding them directly into the AB3 formula for the first generation step. Combined with reuse of cached key-value attention te","pith_inferences":["Whether the cached velocities need to come from the same recording is testable: replacing them with velocities from another song's inversion (or reversing their order) and measuring whether edit quality drops would isolate the trajectory-specific vs generic-smoothing role of DHC.","The seeding crosses a direction reversal and a prompt change, so the classic AB3 error bounds cited in the paper do not directly apply; a numerical check of the local truncation error at the first seeded step on the target trajectory would clarify how far the heuristic departs from the theory.","The same inversion-history seeding could be carried over to image and video editing with rectified flow, where the same warm-start problem appears.","Comparing against a fourth-order single-step solver (e.g., classical Runge-Kutta) at matched cost would separate the benefit of higher order from the benefit of reusing inversion-specific history."],"forward_implications":["A pretrained rectified-flow model can be converted into an editor of arbitrary recordings by adding a caching pass, with no retraining or per-sample optimization.","The seeding strategy should transfer to other multi-step solvers and to future rectified-flow models, as the paper states.","Most of the perceptual improvement comes from the numerical seeding alone, with attention injection adding a smaller complementary gain.","The gains are obtained at the same computational cost as the original pipeline, since DHC reuses velocity evaluations that inversion already computed."],"fun_headline_variants":["Startup fix: reuse inversion velocities for stable music edits","Dynamic history caching stabilizes zero-shot music editing","High-order ODE solver fixes music edit startup drift","Multi-step startup asymmetry? Reuse inversion history","Warm-start drift fixed by reusing inversion velocities"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Velocities cached under the source-prompt inversion are treated as valid Adams-Bashforth history for a generation trajectory that runs under a different prompt and in the reversed time direction.","fun_headline_variants_meta":{"raw":{"variants":["Startup fix: reuse inversion velocities for stable music edits","Dynamic history caching stabilizes zero-shot music editing","High-order ODE solver fixes music edit startup drift","Multi-step startup asymmetry? Reuse inversion history","Warm-start drift fixed by reusing inversion velocities"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000705,"raw_usage":{"total_tokens":3004,"prompt_tokens":720,"completion_tokens":2284,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":464,"completion_tokens_details":{"reasoning_tokens":2209}},"tokens_in":464,"tokens_out":2284,"duration_ms":14117,"temperature":1.0,"reasoning_tokens":2209,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T17:42:56.638061+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the editing pipeline with the cached history taken from a different recording's inversion, or with the two cached velocities exchanged in time order; if edit quality and structural preservation do not degrade, the claimed trajectory-specific benefit of Dynamic History Caching is not the operative mechanism.","supporting_citations":[],"review_version":1}