{"id":"bbb5b0e6-efd4-432d-843e-2ccfc37e26f8","arxiv_id":"2511.11686","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A regularized Schrödinger bridge training scheme—interpolating training targets between clean speech and a posterior-mean denoiser, with perturbed input states—improves SI-SDR by 0.4–1.0 dB over vanilla SB without hurting perceptual metrics.","lead":"Speech enhancement models that use Schrödinger bridges can sound natural but often distort the signal. This paper adds a training regularizer that blends the clean reference with a denoiser's estimate as the target and perturbs the input states, reporting modest SI-SDR gains (0.4–1.0 dB) over vanilla Schrödinger bridge with perceptual scores roughly unchanged.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DP-curve claim rests on misreading of Freirich et al.; ω_t=t² is an untested heuristic, and the 'prediction drift' analysis mischaracterizes the L2 optimum.","rationale":"The reader's weakest assumption already identified the DP-interpolation misreading; I agree. This is the load-bearing issue because the paper's contribution is framed around 'traversing the DP curve.' The empirical results are credible but small, and the mechanism could be explained by simpler label-smoothing/perturbation effects. Since the reader's verdict is CONDITIONAL and my concern reinforces the need for reframing, the verdict remains CONDITIONAL; I recommend no change.","tokens_in":13334,"tokens_out":5900,"duration_ms":59975,"concrete_test":"Ablate the interpolation schedule on VoiceBank+DEMAND: train RSB with ω_t = t, t², t³, and a constant ω=0.5. If SI-SDR/PESQ are insensitive to the schedule or t² is not best, the claim that the target interpolant tracks the DP curve is unsupported. Separately, train vanilla SB with Eq. (4) and compare its outputs at small t to the conditional mean; if they are not more variable than the mean, the claimed posterior-sample endpoint is contradicted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim—that the interpolation target in Eq. (6) 'traverses the DP curve' and approaches the optimal distortion-perception tradeoff—is not supported by the cited literature. Freirich, Michaeli, and Meir (2021) characterize optimal DP solutions as a posterior-mean estimate followed by an optimal transport map to the clean distribution, not as a per-sample linear blend between the clean signal X and the posterior mean X*. The paper asserts the opposite ('the theoretical optimal solutions on the DP curve take an interpolation form') without derivation, and sets ω_t=t² 'for simplicity.' Moreover, the preceding analysis of prediction drift is flawed: Eq. (4) is a plain L2 regression, so its optimum is the conditional mean E[X|X_t, X_1], not a posterior sample. Thus the claimed shift from distortion-oriented to perception-oriented predictions at t→0 is not established. If the DP-curve narrative is invalid, the observed empirical gains may simply reflect a label-smoothing/input-perturbation heuristic, not a principled traversal of an optimality frontier. The paper's own ablation (Table 1) shows input-only perturbation helps SI-SDR, so the mechanism is ambiguous.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Regularized Schrödinger Bridge (RSB), a training-time modification of the Schrödinger Bridge (SB) framework for inverse problems, applied to speech enhancement. RSB perturbs the training objective in two ways: it replaces the ground-truth target X with a time-varying interpolation (1−ω_t)X + ω_t X*, where X* is a posterior-mean estimate from a pretrained predictive model, and it samples perturbed intermediate states from the corresponding marginal. The authors claim that this interpolation steers model predictions along the optimal distortion–perception (DP) curve and that the input perturbation simulates inference-time errors, thereby mitigating exposure bias. Experiments on three speech enhancement benchmarks (WSJ0+WHAM, WSJ0+Reverb, VoiceBank+DEMAND) show small but consistent improvements over vanilla SB in distortion metrics (SI-SDR, SI-SIR, SI-SAR) with comparable or slightly better perceptual metrics. An ablation study indicates that joint input–target perturbation performs best.","tokens_in":13581,"tokens_out":5140,"duration_ms":47802,"significance":"If the theoretical claims were sound, RSB would offer a principled approach to addressing the distortion–perception tradeoff and exposure bias in SB-based inverse problems. The experimental setup is clear, the code is promised in the supplementary material, and the results are credible as measured. However, the central theoretical motivation is not rigorous: the prediction-drift analysis is based on a misinterpretation of the L2 objective, and the interpolation form is not supported by the cited DP-tradeoff literature. The empirical gains, while consistent, are modest and may partly follow from the training target being closer to X*. Nevertheless, the method is clearly specified and the ablation is informative; the contribution could be salvaged by reframing RSB as an empirical regularization heuristic rather than a principled DP-optimality result.","major_comments":[{"comment":"The analysis of prediction drift is incorrect. Equation (4) is an L2 regression objective, so its optimum for each (X_t, X_1) is the conditional mean E[X|X_t, X_1]. At t→0, X_t approaches X_0 = X, so the optimal prediction tends to X itself, not to a 'posterior sample' or a 'perception-oriented' estimate. The statement that the objective 'enforces p_X0 to approach p_X' is not derived from Eq. (4). Thus the claimed transition from distortion-oriented to perception-oriented predictions is not established, which undermines the motivation for the interpolation perturbation in Eq. (6).","section":"DP Tradeoff & Exposure Bias via Prediction Errors (Eq. (4))"},{"comment":"The claim that 'the theoretical optimal solutions on the DP curve take an interpolation form (Freirich, Michaeli, and Meir 2021)' is not supported by that reference. Freirich et al. characterize optimal DP solutions as the posterior-mean estimate followed by an optimal transport map to the clean-data distribution; they do not derive a per-sample linear blend between X and E[X|Y]. The interpolation in Eq. (6) is therefore an ad hoc choice, and the weight ω_t = t^2 is selected 'for simplicity' with no theoretical or empirical justification. The central claim of approaching the optimal DP tradeoff via this interpolation is unsupported.","section":"§Perturbation Design, Eq. (6)"},{"comment":"The ablation does not isolate the proposed mechanism. Input-only perturbation already improves SI-SDR by 0.3 dB, and joint perturbation improves it by 1.0 dB relative to vanilla SB. Because the target in Eq. (8) is a convex combination of X and X*, and X* has higher SI-SDR than a posterior sample by construction, the SI-SDR gain may be partly a direct effect of the training target rather than a traversal of an optimal DP frontier. The paper should compare with alternative schedules (e.g., constant ω) or analyze the DP tradeoff of the targets themselves to support the claimed mechanism.","section":"Table 1 (Ablation study)"}],"minor_comments":[{"comment":"The full-text title ('Regularized Schrödinger Bridge: Alleviating Distortion and Exposure Bias in Solving Inverse Problems') differs from the arXiv title ('Regularized Schrödinger Bridge via Distortion-Perception Perturbation for High-Fidelity Speech Enhancement'). Please align them, and ensure the two abstracts are consistent.","section":"Title/Abstract"},{"comment":"The sentence 'RSB obtains the best perceptual scores across all MOS-derived metrics' is not accurate for DNSMOS, where RSB (3.73) is slightly below SB (3.74) and NCSN++M (3.77). Specify that the claim refers to PESQ, CSIG, CBAK, and COVL.","section":"Table 2"},{"comment":"Figure 2 is difficult to read; the variable labels (e.g., state names) appear garbled. Please redraw it with clear typography.","section":"Figure 2"},{"comment":"The sampler in Eq. (5) is stated without derivation. Adding a short explanation or a citation to Juki'c et al. (2024) for the first-order SDE discretization would improve reproducibility.","section":"Eq. (5) and sampler"}],"recommendation":"major_revision","confidential_remarks":"The theoretical flaws in the DP-curve narrative are significant and, if not addressed, would prevent acceptance at a serious journal. However, the empirical results and the clear experimental methodology suggest that the paper could be revised to present RSB as a well-motivated heuristic regularization, with the theoretical claims softened and the mechanism more carefully analyzed. The reported source code and reproducibility are assets. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the empirical core is decent and the reported gains are real but modest; the theoretical narrative around the distortion–perception curve does not hold up. Treat this as a training heuristic, not a principled DP-optimality result.\n\nWhat's new: RSB combines a time-varying interpolation between the clean target and a posterior-mean estimate (Eq. 6) with an input-state perturbation (Eq. 7) in Schrödinger bridge training. That specific combination is not in the cited input-perturbation literature. The ablation (Table 1) is the strongest part: input-only perturbation gives +0.3 dB SI-SDR, and joint input+target perturbation gives +1.0 dB, which is evidence that the target interpolation does something beyond simple input noise. The experiments on three benchmarks are clearly described and the improvements, while modest (0.5–1 dB SI-SDR; PESQ changes ≤0.07), are consistent and plausible.\n\nThe soft spot is the theory. Eq. (4) is a plain L2 regression; its optimum is the conditional mean E[X | X_t, X_1]. At t→0 the input X_t is essentially X itself, so the optimal prediction is X, not a posterior sample. The paper's claim that predictions 'resemble posterior samples' at small t is incorrect. The 'prediction drift' narrative is asserted, not derived. More importantly, the citation to Freirich, Michaeli, and Meir (2021) is a misread: that paper derives optimal DP solutions as a posterior-mean estimate followed by an optimal transport map, not as a per-sample linear blend between X and X*. The ω_t = t² schedule is chosen 'for simplicity' and is not tested. So the central claim that RSB 'traverses the DP curve' is unsupported. What remains is a label-smoothing plus input-perturbation heuristic that empirically helps. That is a legitimate, if modest, contribution.\n\nAlso, the code is not actually released (the paper says it will be made available upon acceptance), which limits reproducibility.\n\nWho it's for: practitioners working on diffusion-based speech enhancement or other inverse problems who want a simple training trick to improve fidelity without hurting perceptual scores. Theory-minded readers will be frustrated. It deserves a serious referee because the empirical claims are concrete and the method is easily re-implementable, but the authors need to reframe the paper as an empirical regularization technique and drop the optimality claims.\n\nIf I were the editor, I'd send it to review, but I'd expect a major revision that removes the overclaims.","headline":"Empirically decent SB training heuristic; the DP-curve theory is not supported and should be reframed as label smoothing.","tokens_in":14092,"tokens_out":5835,"would_cite":false,"duration_ms":51150,"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 training-time perturbation lets Schrödinger Bridge speech enhancement gain fidelity without losing perceived quality.","keywords":["Regularized Schrödinger Bridge","speech enhancement","distortion-perception tradeoff","exposure bias","diffusion models","inverse problems","dereverberation","posterior mean"],"falsifier":"For a synthetic inverse problem with a known posterior (e.g., Gaussian prior and linear measurement), compute the true optimal DP curve exactly, then train RSB with ω_t = t² and measure the distortion-perception point achieved at each timestep; if the achieved curve does not coincide with or dominate the theoretical DP frontier, the central interpolation claim is falsified.","tokens_in":13137,"feed_emoji":"🎙️","tokens_out":2945,"duration_ms":29520,"temperature":0.7,"pith_summary":"The paper proposes Regularized Schrödinger Bridge (RSB), a training modification of the Schrödinger Bridge generative model for inverse problems, and claims it improves reconstruction fidelity while preserving perceptual quality. It argues that standard Schrödinger Bridge training induces a systematic prediction drift that causes both the distortion-perception tradeoff and error accumulation across sampling steps. RSB replaces the ground-truth training target with a time-varying interpolation between clean speech and a posterior-mean estimate, and feeds the network perturbed intermediate states. The authors report that RSB improves SI-SDR over vanilla SB on three speech benchmarks while matching or exceeding perceptual metrics such as PESQ and CSIG, and that its prediction-error curves are flatter during sampling, indicating reduced exposure bias.","feed_headline":"Training perturbation lifts Schrödinger Bridge speech fidelity","feed_subtitle":"Interpolating targets toward the posterior mean cuts sampling-error buildup and improves SI-SDR on three benchmarks.","key_machinery":"The central object is the distortion-perception perturbation: a time-varying interpolation between the clean signal X (perception-optimal endpoint) and the posterior mean X* (distortion-optimal endpoint), with weight ω_t = t², applied to both the training target (Eq. 6) and the input state (Eq. 7) inside the Schrödinger Bridge training objective (Eq. 8). This perturbation converts training from predicting ground truth into predicting a point on the hypothesized DP curve at each timestep, while the perturbed inputs mimic the model's own sampling errors and therefore reduce the training-inference mismatch responsible for exposure bias.","core_discovery":"The paper's central claim is that a joint input-target perturbation during training simultaneously mitigates exposure bias and shifts the solution toward the optimal distortion-perception tradeoff. Specifically, RSB defines training targets as eX_tgt = (1 - ω_t)X + ω_t X*, where X is the clean signal, X* is the posterior-mean estimate from a pretrained predictive model, and ω_t = t² grows from 0 to 1 over the bridge time. It then samples perturbed intermediate states from the corresponding Schrödinger Bridge marginal and trains the network to predict the perturbed target. This is argued to simulate the prediction errors the model will encounter at inference, preventing error accumulation, an","pith_inferences":["The fixed ω_t = t² schedule is a heuristic; a testable extension is to learn or adapt the schedule per timestep and compare the achieved distortion-perception frontier with the one reported here.","RSB's gains depend on the quality of the frozen posterior-mean model X*; jointly training or online-updating that model could further shift the tradeoff, but the paper does not investigate this.","The interpolation perturbation is formally a label-smoothing technique, suggesting it may transfer to other bridge-based generative models beyond Schrödinger Bridges.","If the true optimal DP frontier is not a linear blend between X and X*, RSB's claim of 'traversing the DP curve' may reduce to a heuristic regularizer; a synthetic Gaussian experiment could test this directly."],"forward_implications":["RSB improves SI-SDR over vanilla SB on WSJ0+WHAM, WSJ0+Reverb, and VoiceBank+DEMAND, while matching or improving perceptual metrics such as PESQ, CSIG, CBAK, and COVL.","The per-step prediction-error curve for RSB is flatter than for vanilla SB across 50 sampling steps, indicating slower error accumulation and reduced exposure bias.","RSB achieves a more favorable distortion-perception tradeoff as the number of sampling steps varies, reaching high PESQ early while preserving SI-SDR better than SB.","Because the posterior-mean model is used only during training, RSB adds no inference-time dependency, in contrast to cascade strategies that feed the predictive estimate into the sampler.","The same regularized-training formulation is proposed as applicable to other inverse problems, including image denoising, deblurring, and inpainting."],"fun_headline_variants":["Perturbed targets fix Schrödinger Bridge drift","Time-varying targets curb exposure bias in speech","Bridge training with perturbation boosts fidelity","Distortion-perception perturbation tames bridge drift"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the linear interpolation (1 - ω_t)X + ω_t X* with ω_t = t² accurately proxies the optimal distortion-perception curve; if that proxy is wrong, the method is an arbitrary label-smoothing heuristic rather than a way to approach the theoretical tradeoff.","fun_headline_variants_meta":{"raw":{"variants":["Perturbed targets fix Schrödinger Bridge drift","Time-varying targets curb exposure bias in speech","Bridge training with perturbation boosts fidelity","Distortion-perception perturbation tames bridge drift"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00028,"raw_usage":{"total_tokens":1530,"prompt_tokens":807,"completion_tokens":723,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":668}},"tokens_in":551,"tokens_out":723,"duration_ms":9171,"temperature":1.0,"reasoning_tokens":668,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T06:48:29.800216+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a synthetic inverse problem with a known posterior (e.g., Gaussian prior and linear measurement), compute the true optimal DP curve exactly, then train RSB with ω_t = t² and measure the distortion-perception point achieved at each timestep; if the achieved curve does not coincide with or dominate the theoretical DP frontier, the central interpolation claim is falsified.","supporting_citations":[],"review_version":1}