{"id":"721e0432-d911-4ab0-9464-133e2e66c666","arxiv_id":"2606.20671","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"SCALED surrogate gradient is reinterpreted as a projection-based first-order local approximation of non-differentiable video codecs, enabling effective training of full neural wrappers with BD-Rate gains up to 23.59% on x264.","lead":"This paper reinterprets the SCALED surrogate gradient as a first-order local approximation of video codecs to enable end-to-end training of neural pre- and post-processors. A smart generalist might read it to understand practical ways to boost compression efficiency in existing video systems without replacing the codecs.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Reinterpretation of SCALED as first-order local approximation requires verifying that the codec's discrete decisions admit a consistent linear model whose gradient matches the reparameterization exactly.","rationale":"The reader's weakest assumption directly identifies the same point: whether the reinterpretation remains valid and property-preserving for full neural wrapping. Because the provided abstract contains only the claim without the supporting derivation, the load-bearing risk is precisely the unverified step from reparameterization to local linear model; the concrete test above would falsify or confirm that step without requiring the full experimental suite.","tokens_in":1772,"tokens_out":349,"duration_ms":10594,"concrete_test":"Extract the exact SCALED gradient formula and the claimed first-order approximation from the methods section; recompute both on a single-frame toy codec (identity + uniform quantizer) with a known analytic gradient and check whether they coincide to machine precision for at least three different input perturbations of size < 1e-3.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on showing that the SCALED surrogate (originally a reparameterization using true compression error) equals the gradient of a first-order Taylor expansion of the codec around the operating point. For this to hold when extending to full pre/post neural wrappers, the codec must be locally approximable by a differentiable function whose Jacobian is independent of the wrapper networks; otherwise the mathematical properties (e.g., unbiasedness or variance) inherited from the reparameterization may change. The abstract provides no derivation or explicit statement of the local model, leaving open whether the equality is identity-level or only approximate under unstated regularity conditions on the quantization and mode decisions.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper claims that the SCALED surrogate gradient, originally introduced as a reparameterization trick, admits a projection-based reinterpretation as a first-order local approximation to the video codec. It extends the approach from downscaling to full neural codec wrappers (pre- and post-processing networks) and reports generalization across codecs (x264, VVenC), quality factors, and downscaling ratios, with BD-Rate (PSNR) gains up to -23.59% and -20.07% relative to standard resampling baselines.","tokens_in":1928,"tokens_out":397,"duration_ms":10446,"significance":"If the local-approximation interpretation is shown to hold exactly (or under clearly stated conditions), the work would supply a useful analytic lens for surrogate gradients in hybrid neural-conventional codecs and justify extending SCALED to more complex wrapper architectures. The reported BD-Rate numbers are concrete and the generalization claim is testable; both would be strengthened by the promised mathematical grounding.","major_comments":[{"comment":"The central claim equates the SCALED reparameterization to the gradient of a first-order Taylor expansion of the codec. The abstract states the interpretation but supplies neither the explicit local model nor the derivation showing that the surrogate matches the Jacobian of that model; without this step the reinterpretation remains heuristic rather than identity-level.","section":"Abstract / §3 (interpretation)"},{"comment":"When extending to full pre- and post-processing networks, the manuscript must verify that the codec's local linear model (and therefore the surrogate gradient) remains independent of the wrapper parameters; otherwise the unbiasedness or variance properties inherited from the original reparameterization may no longer hold. No such verification or regularity condition is referenced in the provided abstract.","section":"§4 (full neural wrapping)"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments, which help strengthen the mathematical grounding of our claims. We address each major comment below and indicate the revisions we will make.","responses":[{"response":"The explicit local linear model (first-order Taylor expansion of the codec around the input point) and the derivation showing that the SCALED surrogate equals the gradient of this approximation (via the projection interpretation) are provided in full in Section 3. The abstract summarizes the result at a high level. We will revise the abstract to include one additional sentence referencing the local model and directing readers to the derivation in §3, thereby making the identity-level nature of the claim under the stated conditions explicit.","revision_made":"yes","referee_comment":"[Abstract / §3 (interpretation)] The central claim equates the SCALED reparameterization to the gradient of a first-order Taylor expansion of the codec. The abstract states the interpretation but supplies neither the explicit local model nor the derivation showing that the surrogate matches the Jacobian of that model; without this step the reinterpretation remains heuristic rather than identity-level."},{"response":"The local linear model is constructed exclusively from the fixed, non-differentiable codec function and is therefore independent of the neural wrapper parameters by definition. The surrogate gradient continues to use the true compression error, preserving the unbiasedness and variance properties of the original reparameterization. We will add an explicit verification paragraph in §4 stating this independence and the regularity condition that the codec remains unchanged.","revision_made":"yes","referee_comment":"[§4 (full neural wrapping)] When extending to full pre- and post-processing networks, the manuscript must verify that the codec's local linear model (and therefore the surrogate gradient) remains independent of the wrapper parameters; otherwise the unbiasedness or variance properties inherited from the original reparameterization may no longer hold. No such verification or regularity condition is referenced in the provided abstract."}],"tokens_in":1439,"tokens_out":424,"duration_ms":20427,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main takeaway is that this work gives a projection-based reading of the SCALED surrogate gradient as a first-order local approximation to the video codec. They then show the same gradient works for the harder case of full neural wrapping with both pre- and post-processing networks, not just downscaling.\n\nWhat is new is the explicit local-linear interpretation and the extension to complete wrappers. The paper reports concrete BD-Rate (PSNR) numbers: up to -23.59% on x264 and -20.07% on VVenC versus standard resampling baselines. They also test across multiple codecs, quality factors, and downscaling ratios, which is a reasonable check for generalization.\n\nThe results are useful for anyone training networks around conventional video codecs. Using the true compression error directly, without a separate mimic network, is a practical advantage, and the numbers suggest the method scales to the full pre/post setting.\n\nThe soft spot is the central claim about the approximation. The stress-test note is on point: for the surrogate to equal the gradient of a first-order Taylor expansion, the codec's discrete decisions (quantization, mode choices) need to admit a consistent linear model whose Jacobian does not depend on the wrapper networks. The abstract states the interpretation but does not show the derivation. If the full paper only offers a heuristic argument rather than an identity-level proof under stated regularity conditions, the insight is weaker than presented. That is the main thing a referee would need to verify.\n\nThis paper is for people working on neural wrappers for video compression. A reader already familiar with surrogate gradients will see the projection angle and the empirical extension as the useful parts.\n\nIt deserves peer review. The empirical results are grounded enough and the angle on an existing method is clear enough that referees should see the details.","headline":"The paper reinterprets SCALED as a first-order local approximation of the codec and extends it to full pre/post neural wrappers, with reported BD-rate gains but the exact match to the Taylor expansion needs checking in the math.","tokens_in":2407,"tokens_out":458,"would_cite":false,"duration_ms":15552,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"The SCALED surrogate gradient for neural codec wrappers is a first-order local approximation of the codec.","keywords":["neural wrappers","surrogate gradients","video codecs","non-differentiable optimization","SCALED gradient","pre- and post-processing","BD-Rate"],"falsifier":"A direct comparison, on identical training runs, between the SCALED gradient and the true partial derivative of a differentiable proxy codec; if the performance gap widens substantially for full wrapping, the local-approximation claim does not hold.","tokens_in":2696,"feed_emoji":"📹","tokens_out":609,"duration_ms":13366,"temperature":0.7,"pith_summary":"The paper establishes that the SCALED surrogate gradient, previously introduced only as a reparameterization trick, admits an interpretation as a first-order local approximation to the non-differentiable video codec. This view explains its effectiveness for end-to-end training of neural pre- and post-processors without needing an auxiliary mimic network. The same gradient supports both simple downscaling tasks and the harder case of full neural wrapping. Experiments show consistent gains across codecs, quality settings, and downscaling ratios, with BD-Rate reductions reaching -23.59 percent on x264.","feed_headline":"Surrogate gradient equals first-order codec approximation","feed_subtitle":"Reinterpretation explains why SCALED training succeeds for full neural pre- and post-processors around x264 and VVenC.","key_machinery":"The SCALED gradient, which replaces the non-differentiable codec with a scaled version of the true compression error.","core_discovery":"The SCALED surrogate gradient can be reinterpreted as a first-order local approximation of the video codec. This reinterpretation supplies insight into why the gradient enables effective training of neural wrappers and remains valid when the wrapper consists of both pre- and post-processing networks.","pith_inferences":["The local-approximation view may extend to other non-differentiable signal-processing blocks whose error surface is locally smooth.","One could test whether replacing the true error with a cheaper estimate still preserves the first-order property.","The same gradient might be applied to audio or image codecs whose internal decisions are also discrete."],"forward_implications":["Training becomes possible for complete neural wrappers rather than only downscalers.","Performance improvements generalize across different video codecs and quantization parameters.","Multiple downscaling ratios can be handled by the same gradient construction.","BD-Rate gains of up to 23.59 percent on x264 are obtained relative to standard resampling."],"fun_headline_variants":["SCALED as first-order local codec approximation","Projection-based SCALED gradient interpretation","First-order codec approx via surrogate projection","Local approximation view of neural codec gradients","SCALED reinterpreted through codec projection"],"cache_read_input_tokens":64,"weakest_assumption_plain":"The scaled gradient introduced as a reparameterization trick continues to function as a mathematically valid first-order approximation even when the wrapper includes both pre- and post-processing networks.","fun_headline_variants_meta":{"raw":{"variants":["SCALED as first-order local codec approximation","Projection-based SCALED gradient interpretation","First-order codec approx via surrogate projection","Local approximation view of neural codec gradients","SCALED reinterpreted through codec projection"]},"model":"grok-4.3","cost_usd":0.00628,"raw_usage":{"total_tokens":2961,"prompt_tokens":683,"num_sources_used":0,"completion_tokens":54,"cost_in_usd_ticks":62799500,"prompt_tokens_details":{"text_tokens":683,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2224,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":683,"tokens_out":54,"duration_ms":13582,"temperature":1.0,"reasoning_tokens":2224,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T05:10:26.234478+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A direct comparison, on identical training runs, between the SCALED gradient and the true partial derivative of a differentiable proxy codec; if the performance gap widens substantially for full wrapping, the local-approximation claim does not hold.","supporting_citations":[],"review_version":1}