{"id":"f7ce765b-a037-42d2-abd7-e2ab500b945a","arxiv_id":"2506.06323","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A PPO-based adaptive filter with a composite reward is claimed to beat classical filters on synthetic noise, but the abstract's ECG results are absent from the body and the paper has internal inconsistencies.","lead":"This paper trains a PPO reinforcement learning agent to adjust digital filter coefficients for noise removal, using a composite reward that balances SNR gain, mean squared error, and output smoothness. The goal is an adaptive filter that works without manual retuning in changing noise, which could benefit wireless, biomedical, and sensor signal processing.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract asserts ECG experiments and real-world outperformance, but Section VIII-B says evaluation is simulated-only; the central 'necessarily surpasses' claim is therefore unsupported and internally inconsistent.","rationale":"I read the paper as a letter-style proposal with a synthetic proof-of-concept. Given the abstract and conclusion, the central claim is real-world viability. The single most load-bearing condition is that a policy trained on Gaussian-noise-corrupted synthetic sinusoids transfers to real physiological signals. The paper's own limitations section explicitly denies that this has been tested. This is not a matter of outside consensus; it is an internal contradiction between the abstract's ECG claim and Section VIII-B. The reader's weakest_assumption identifies the same point, and I agree. I also note that the synthetic results are reported as point estimates with no seed variance or statistical comparison, and the ablation narrative conflicts with Table I, so even the synthetic leg is not fully reproducible as written. I credit the paper's concrete choices (PPO clipping, LSTM policy, 64-sample window, grid-searched baselines) as a plausible starting point, and the reported sub-millisecond inference is useful if reproducible. But the central claim as stated does not survive: 'necessarily surpasses' is too strong for the evidence, and the ECG claim is unsupported. Therefore I do not move the reader's verdict; REJECT remains appropriate.","tokens_in":5392,"tokens_out":3591,"duration_ms":40861,"concrete_test":"Run the full PPO pipeline and the four grid-searched baselines on the MIT-BIH Noise Stress Test Database (for example, record 118 clean ECG with added baseline wander, electrode motion, and muscle artifact), using the exact reward in Eq. 3, multiple random seeds, and reporting per-seed output SNR, residual total variation, and inference latency. If PPO does not significantly outperform the best classical baseline on all three noise types, or if no such experiment is provided, the abstract's real-world outperformance claim and the conclusion's 'necessarily surpasses' claim should be withdrawn.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest claim is that PPO-driven filtering 'necessarily surpasses' LMS, RLS, Wiener, and Kalman on output SNR and signal smoothness. This claim is supposed to be backed by the abstract's ECG results on the MIT-BIH Noise Stress Test, but no ECG experiment appears in the body. Section VIII-B explicitly states: 'Our evaluation has so far been limited to simulated data.' The only quantitative evidence is synthetic sum-of-sinusoids with additive noise, so the synthetic-to-real transfer is an unsupported load-bearing assumption. The central claim is also phrased as a necessity ('necessarily surpasses'), but no proof or exhaustive analysis supports a universal ordering; the reported numbers are point estimates from a single run without error bars, seeds, or statistical tests. Even the synthetic ablation is internally inconsistent: Table I says 'SNR + MSE (No TV)' diverges with '<0 (diverged)', while Section VIII-A says removing either term costs only 2-4 dB of generalization loss. These inconsistencies mean the headline result cannot be independently checked as reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The letter proposes an adaptive filtering framework based on Proximal Policy Optimization (PPO). Adaptive filtering is cast as an MDP (Section III), and the agent is trained with the composite reward R(t)=αΔSNR(t)-βMSE(t)-γTV(r(t)) (Eq. 3), where α, β, γ are balancing weights. The agent is trained on synthetic sum-of-sinusoids corrupted with Gaussian noise and evaluated on Gaussian, Laplacian, impulse, pink, brown, and uniform noise (Section VI). The paper claims that the PPO policy outperforms LMS, RLS, Wiener, and Kalman filters in output SNR and smoothness while running in under 1 ms per step, and that it generalizes to noise types unseen in training. The abstract additionally claims ECG validation on the MIT-BIH Noise Stress Test Database, but no such experiment appears in the body; Section VIII-B states 'Our evaluation has so far been limited to simulated data.'","tokens_in":5614,"tokens_out":8836,"duration_ms":86057,"significance":"If the empirical claims were supported by complete and reproducible experiments, the paper would offer a useful demonstration that a policy-gradient RL agent can act as a low-latency adaptive filter for nonstationary noise, with a clear MDP formulation and a reasonable cross-noise generalization protocol. The runtime comparison and the ablation of reward components are also relevant. However, as submitted, the significance is not established: there are no error bars or statistical tests, the reward weights are not reported, the Table I ablation conflicts with the discussion in Section VIII-A, and the abstract's ECG claim is contradicted by the body. The contribution is therefore best viewed as a pilot study that needs substantial revision before its claims can be evaluated.","major_comments":[{"comment":"The abstract states that 'real-world analysis is made and evaluated on ECG recordings from the MIT-BIH Noise Stress Test Database' and that the PPO policy 'slightly outperforms strong classical baselines on ECG denoising.' The body contains no ECG experiment, no MIT-BIH results, and no description of how ECG signals were filtered. Section VIII-B states explicitly that 'Our evaluation has so far been limited to simulated data,' and Section VIII-C lists physiological sensor data as future work. This direct contradiction means the abstract's real-world result cannot be independently checked and should be removed or supported with the missing experiment.","section":"Abstract vs. Sections VIII-B and VIII-C"},{"comment":"The conclusion says that the PPO-driven filter 'necessarily surpasses traditional adaptive methods (LMS, RLS, Wiener, Kalman) in terms of output SNR and signal smoothness.' The evidence consists of a single synthetic experiment with no reported number of runs, no error bars, no seeds, and no statistical tests. The word 'necessarily' asserts a universality that no finite set of simulations can establish, and the paper provides no proof or exhaustive analysis. The claim should be restricted to the specific tested configurations and supported by repeated runs with confidence intervals.","section":"Section IX"},{"comment":"The composite reward in Eq. (3) is the central design contribution, but the values of α, β, and γ are never reported, nor is the selection procedure described. Table I compares reward components, but without the weights the ablation cannot be reproduced or interpreted quantitatively. In addition, ΔSNR and MSE in Eq. (3) are the same metrics used for evaluation in Section VII; the reported SNR improvement therefore reflects the training objective directly rather than an emergent property. The authors should report the weights and add at least one evaluation metric that is not part of the reward, such as spectral distortion or a downstream task accuracy.","section":"Section V, Eq. (3)"},{"comment":"Table I reports that the 'SNR + MSE (No TV)' variant 'diverged' with '<0 (diverged)' SNR, while Section VIII-A states that removing the smoothness term 'resulted in degraded performance' with 'a generalization loss of approximately 2–4 dB in output SNR on unseen noises.' A divergent run with negative SNR is not described by a 2–4 dB generalization loss; these two statements are mutually inconsistent. This discrepancy must be resolved before the ablation can be used to support the composite-reward design.","section":"Section VII-E (Table I) vs. Section VIII-A"},{"comment":"The baseline comparisons and reported results are insufficiently specified. Grid-search ranges for the LMS step size, RLS forgetting factor, Kalman process/measurement noise, and Wiener correlation estimates are not given, and the filter order K in the FIR action model is never stated. The text reports only point estimates without standard deviations or test statistics, so the claimed systematic improvement over four classical baselines cannot be quantitatively assessed from the information provided.","section":"Section VI and Section VII-B"}],"minor_comments":[{"comment":"The word 'plateu' should be 'plateau' in the sentence about episodic reward convergence.","section":"Section VII-D"},{"comment":"The phrase 'assumptions of stationary' should read 'assumptions of stationarity'.","section":"Section II"},{"comment":"The figures would benefit from axis labels and units for the horizontal axis (noise level or input SNR); the current panels make it difficult to see the numerical magnitudes behind the claimed SNR gains.","section":"Section VI and Fig. 1"},{"comment":"The architecture description differs between Section IV/V, which mentions two-layer MLPs, and Section VII-A, which introduces an LSTM state; reconcile these descriptions so the network structure is unambiguous.","section":"Section VII-A"}],"recommendation":"major_revision","confidential_remarks":"The abstract-body mismatch over the ECG evaluation is a serious integrity concern that should be addressed head-on in any revision. If the MIT-BIH results cannot be produced, the ECG claim must be removed from the abstract. The authors should also supply reward weights, repeated-seed statistics, and a corrected ablation table. With those changes, the simulation-only contribution might be publishable, but not in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is reasonable: frame adaptive FIR filtering as an MDP, train a PPO agent with a composite reward that balances SNR gain, MSE, and residual smoothness, and test generalization on noise types not seen in training. That combination is not in the cited prior work, so there is a small kernel of novelty. The paper also does a decent job of situating itself against earlier RL-based filtering papers (Oh, Xie, Lin, Luo), and the motivation for moving beyond LMS/RLS/Wiener/Kalman is clearly stated.\n\nThat is where the credit ends. The execution has problems serious enough that I would not trust any of the reported numbers as they stand. The abstract claims ECG evaluation on the MIT-BIH Noise Stress Test and says the method \"slightly outperforms\" classical baselines on ECG, but Section VIII-B explicitly says evaluation has so far been limited to simulated data. No ECG experiment appears anywhere in the body. That is not a minor omission; it is a load-bearing inconsistency between the headline claim and the actual evidence.\n\nThe internal contradictions continue. Section VI describes the policy and value networks as two-layer MLPs with 128 units, while Section VII-A says the actor-critic uses a single-layer LSTM with a two-layer MLP critic. Table I says removing the TV term makes training diverge with SNR under 0 dB, but Section VIII-A says removing either the smoothness term or the SNR/MSE term costs only 2-4 dB of generalization. Those two statements cannot both be true. The reward weights alpha, beta, and gamma are never reported, so the ablation cannot be reproduced. There are no error bars, no seeds, no statistical tests, and no code or data.\n\nThe conclusion's \"necessarily surpasses\" is an overstatement in the literal sense: no proof or exhaustive analysis establishes a necessary ordering. At best, the synthetic results suggest the PPO filter can outperform tuned classical filters on the tested configurations. That may be true, but the paper as written does not demonstrate it reliably.\n\nWho gets value from this? Someone exploring RL-based adaptive filtering might read it as a pointer to a plausible reward design and a reminder of the challenges. But as evidence, it is not currently usable. The fix is straightforward in principle: report reward weights, reconcile the architecture and ablation descriptions, add error bars, and either add the ECG experiments or remove the claim. If the author does that, the paper could deserve peer review. As it stands, I would reject without sending it to referees: the abstract-body mismatch and self-contradictory results suggest it is not ready for external scrutiny.\n\nRecommendation: desk reject with an invitation to resubmit after major revision.","headline":"A plausible RL-for-filtering idea undermined by internal contradictions: the abstract's ECG results don't exist in the body, the architecture changes mid-paper, and the ablation table disagrees with the discussion.","tokens_in":6104,"tokens_out":1914,"would_cite":false,"duration_ms":20299,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A PPO-trained agent with a composite reward is claimed to beat LMS, RLS, Wiener, and Kalman on nonstationary denoising while running in real time.","keywords":["adaptive filtering","reinforcement learning","PPO","composite reward","nonstationary noise","signal denoising","generalization","real-time inference"],"falsifier":"Run the trained PPO policy on an ECG segment from the MIT-BIH Noise Stress Test Database corrupted by baseline wander, electrode motion, and muscle artifacts, and compare output SNR and residual smoothness against grid-tuned LMS, RLS, Wiener, and Kalman filters; the central claim fails if the PPO policy does not beat all four baselines on those real recordings.","tokens_in":5183,"feed_emoji":"📶","tokens_out":6083,"duration_ms":54570,"temperature":0.7,"pith_summary":"This paper claims that an adaptive filter whose coefficients are updated by a Proximal Policy Optimization (PPO) agent can outperform four classical filters—LMS, RLS, Wiener, and Kalman—on nonstationary denoising tasks, while running fast enough for real-time use. The agent is trained only on synthetic sinusoids corrupted by Gaussian noise, yet the experiments report that it retains high output SNR on five unseen noise types. The training signal is a composite reward balancing SNR improvement, mean-squared error, and residual smoothness; ablations show that removing any component degrades stability or generalization. The conclusion goes further, stating that the PPO-driven filter 'necessarily surpasses' classical adaptive methods, although the discussion also notes that real-world validation, including the ECG evaluation mentioned in the abstract, remains outstanding.","feed_headline":"PPO-driven filter beats classical denoisers on unseen noise","feed_subtitle":"Trained only on Gaussian noise, the learned agent claims real-time gains across six noise types.","key_machinery":"The load-bearing object is the composite reward function $R(t) = \\alpha \\Delta\\text{SNR}(t) - \\beta \\text{MSE}(t) - \\gamma TV(r(t))$, where $\\Delta\\text{SNR}$ is the signal-to-noise improvement at the filter output, $\\text{MSE}$ is the mean-squared error against a reference, and $TV(r)$ is the total variation of the residual $r(t)=y(t)-\\hat{x}(t)$. This reward is combined with PPO's clipped surrogate objective $L^{\\text{CLIP}}(\\theta)$ to keep each policy update conservative. The agent observes a 64-sample sliding window plus the current residual, and its LSTM policy outputs a Gaussian over FIR coefficient updates. The reward's job is to define what 'good filtering' means during training; the ablation study in the paper shows that dropping the MSE term or the TV term leads to divergence or a 2–4 dB generalization loss.","core_discovery":"The central claim is that framing adaptive filtering as a Markov decision process and training a PPO policy with the composite reward $R(t) = \\alpha \\Delta\\text{SNR}(t) - \\beta \\text{MSE}(t) - \\gamma TV(r(t))$ yields a filter that adapts to changing noise without manual re-tuning. The paper argues that PPO's clipped surrogate objective prevents the instability that classical gradient-based updates suffer under nonstationary noise, and that the smoothness penalty keeps residuals small and artifact-free. On the paper's synthetic benchmark, the learned policy beats grid-tuned LMS, RLS, Wiener, and Kalman baselines in output SNR across Gaussian, impulsive, Laplacian, pink, brown, and uniform noise, and runs in under one millisecond per time step. The author states this surpasses the classical filters in both SNR and smoothness, not only on seen noise.","pith_inferences":["A natural test the paper leaves implicit: run the same trained policy on the MIT-BIH Noise Stress Test ECG recordings mentioned in the abstract and report per-noise-type SNR; if it underperforms tuned classical filters there, the synthetic-to-real transfer assumption fails.","The composite reward is essentially a regularized objective; one could expect similar gains from other policy-gradient algorithms, so the paper's contribution is the reward design more than PPO itself.","If transfer to real signals succeeds, the paper's own caveat that the training phase is more demanding points to a deployment trade-off: training cost is high, but inference cost is small enough for embedded use only if the network is simplified."],"forward_implications":["If the central claim holds, a single PPO policy can replace tuned LMS, RLS, Wiener, and Kalman filters in dynamic noise environments, eliminating manual re-calibration.","A filter trained on only one noise type can generalize to several unseen noise distributions, so training data need not cover every deployment condition.","The composite reward is not optional: the paper reports that removing the MSE term makes training diverge and removing the TV term causes borderline generalization, so any reimplementation needs all three terms.","At under 1 ms per inference step, the learned policy is fast enough for low-latency streaming applications such as radio-signal cleaning and biomedical monitoring.","Because the method does not need ground truth at runtime, it can be applied online in settings where clean signals are unavailable."],"supporting_citations":[{"why":"Supplies the PPO clipped-surrogate objective that stabilizes policy updates in continuous action spaces.","marker":"[4]"},{"why":"Defines the LMS adaptive filter used as the classical baseline that the PPO agent must beat.","marker":"[1]"},{"why":"Provides the adaptive filter theory and RLS/Wiener formulations used as classical baselines.","marker":"[2]"},{"why":"Defines the Kalman filter with state-space model used as a baseline.","marker":"[3]"},{"why":"Establishes the precedent that a model-free RL agent can outperform classical estimators in uncertain channel conditions.","marker":"[5]"},{"why":"Shows an RL-driven filter designed for impulsive noise, the closest prior approach to PPO-based adaptive filtering.","marker":"[7]"},{"why":"Demonstrates deep RL applied to active noise control, supporting the paper's claim that RL is viable for filtering.","marker":"[10]"}],"fun_headline_variants":["Reinforcement learning filter beats classical denoisers on ECG","PPO adaptive filter generalizes to unseen noise types in real time","Composite reward helps AI filter outclass LMS, RLS, and Kalman","Policy-gradient denoiser matches Kalman on ECG with less tuning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that a policy trained only on synthetic sinusoids plus Gaussian noise will transfer to real-world signals such as ECG, even though its own discussion states that evaluation has so far been limited to simulated data.","fun_headline_variants_meta":{"raw":{"variants":["Reinforcement learning filter beats classical denoisers on ECG","PPO adaptive filter generalizes to unseen noise types in real time","Composite reward helps AI filter outclass LMS, RLS, and Kalman","Policy-gradient denoiser matches Kalman on ECG with less tuning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000469,"raw_usage":{"total_tokens":2330,"prompt_tokens":934,"completion_tokens":1396,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":1320}},"tokens_in":550,"tokens_out":1396,"duration_ms":12972,"temperature":1.0,"reasoning_tokens":1320,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:36:20.983101+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained PPO policy on an ECG segment from the MIT-BIH Noise Stress Test Database corrupted by baseline wander, electrode motion, and muscle artifacts, and compare output SNR and residual smoothness against grid-tuned LMS, RLS, Wiener, and Kalman filters; the central claim fails if the PPO policy does not beat all four baselines on those real recordings.","supporting_citations":[{"cited_title":"Widrow and S","cited_arxiv_id":null,"evidence_quote":"Defines the LMS adaptive filter used as the classical baseline that the PPO agent must beat."},{"cited_title":"Haykin, Adaptive Filter Theory , 5th ed","cited_arxiv_id":null,"evidence_quote":"Provides the adaptive filter theory and RLS/Wiener formulations used as classical baselines."},{"cited_title":"A new approach to linear filtering and prediction problems,","cited_arxiv_id":null,"evidence_quote":"Defines the Kalman filter with state-space model used as a baseline."},{"cited_title":"Channel estimation via successive denoising in MIMO- OFDM systems: a reinforcement learning approach,","cited_arxiv_id":null,"evidence_quote":"Establishes the precedent that a model-free RL agent can outperform classical estimators in uncertain channel conditions."},{"cited_title":"A fractional filter based on reinforcement learning for effective tracking under impulsive noise,","cited_arxiv_id":null,"evidence_quote":"Shows an RL-driven filter designed for impulsive noise, the closest prior approach to PPO-based adaptive filtering."},{"cited_title":"GFANC-RL: reinforcement learning-based generative fixed-filter active noise control,","cited_arxiv_id":null,"evidence_quote":"Demonstrates deep RL applied to active noise control, supporting the paper's claim that RL is viable for filtering."}],"review_version":1}