{"id":"3c29efde-ac56-4ae3-9b01-a424c5ca80e6","arxiv_id":"2505.09196","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper names the 'gene effect' for the observation that random parameters sometimes outperform trained ones in low-light enhancement, and introduces parameter dynamic evolution with orthogonal basis generation to counter it, yielding modest PSNR gains.","lead":"Researchers at Beijing Institute of Technology report that resetting selected trained parameters to random values can improve low-light image enhancement for some images, a phenomenon they call the gene effect. They propose a dynamic parameter generation module that adapts weights per image to counter this effect and claim roughly 1 dB PSNR improvement on the LOL-v1 benchmark.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DGE metric in Eq. 6 is inverted relative to the gene-effect definition: it measures output sensitivity, not whether random resets improve enhancement, so the central claim that PDE mitigates the gene effect is not supported.","rationale":"The reader correctly identified the DGE metric as the load-bearing weakness. My reading agrees and sharpens it: DGE is not merely an unvalidated proxy; its sign is inverted relative to the paper's own definition of the gene effect. The gene effect is defined as random parameter resets improving enhancement performance, which requires the reset to change the output, often substantially. DGE measures the opposite direction — how much the output changes — and the paper treats low output change as 'more gene effect.' This makes the evidence for the central causal claim unsound. The phenomenon itself is still plausible and Table 1 directly measures it with POI, but that direct metric is abandoned in the mitigation evaluation. The proposed PDE module does show small, consistent PSNR gains on standard benchmarks, so a rejection is not warranted; the paper should be conditional on a direct, ground-truth-relative evaluation of gene-effect mitigation and on the release of code so the DGE numbers can be recomputed. For these reasons I leave the verdict unchanged.","tokens_in":13923,"tokens_out":4148,"duration_ms":43144,"concrete_test":"Directly measure the gene effect as defined. Using the same trained checkpoints and reset protocol as Table 1 and Figure 3, for each of the n reset layers and each test image, compute the PSNR to the ground-truth reference for the original model and for the reset model. Define POI as the fraction of images where the reset model's PSNR exceeds the original model's PSNR, and the mean improvement on those images. Report POI for Restormer and Restormer+PDE, and also compute the correlation between DGE and POI across layers. If POI is not substantially lower for PDE, or if high DGE is not associated with low POI, then Eq. 6 is not a valid proxy and the claim that PDE mitigates the gene effect fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that PDE mitigates the gene effect rests entirely on the DGE metric in Eq. 6. DGE is the average PSNR between the original model output and the reset-model output. But the gene effect, as defined in the abstract and Section 3, is the event that resetting parameters to random values improves enhancement performance with respect to the ground truth. DGE never involves the ground truth. A reset that leaves outputs nearly unchanged gives a high DGE; such a reset cannot produce the improvement that defines the gene effect. Conversely, a reset that produces large output changes gives a low DGE, and this is the only regime in which random parameters can beat learned ones, as in Figure 3. The paper's interpretation in Section 5.2 ('A larger value of DGE ... indicating more gene effects. Conversely, a smaller DGE implies weaker gene effects') is therefore backwards. Restormer+Ours lowering DGE from 48.94 to 45.09 means its output is more sensitive to random resets, not that the gene effect is smaller. Because every gene-effect-mitigation claim in Tables 3-4 and the ablation reasoning rests on DGE, the causal claim 'PDE mitigates the gene effect' is not established. The enhancement gains in Table 5 are independent evidence for PDE as a module, but do not validate the biological analogy or the mitigating claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a phenomenon dubbed the 'gene effect' in low-light image enhancement: for certain images, resetting trained network parameters to random values yields better enhancement (higher PSNR) than the learned parameters. The authors attribute this to static parameters, propose a 'parameter dynamic evolution' (PDE) module with 'parameter orthogonal generation' (POG) to mitigate the effect, and report PSNR improvements on LOL-v1, LOL-v2-real, and LOL-v2-synthetic benchmarks. The central conceptual claim is that PDE mitigates the gene effect, evaluated with a newly defined DGE metric (Eq. 6) that measures output difference before/after parameter reset.","tokens_in":14181,"tokens_out":4559,"duration_ms":44352,"significance":"If the core claims were established, the paper would contribute a new observation (random parameter resets sometimes improve LLIE outputs) and a practical plug-and-play dynamic convolution module that yields consistent, if modest, PSNR gains on standard paired benchmarks. The method is simple and the empirical enhancement improvements in Table 5 are interesting independent of the biological analogy. However, the significance is substantially weakened because the paper's evaluation of gene-effect mitigation rests on a metric (DGE) that appears to measure the wrong quantity and is interpreted in a way that is internally inconsistent with the paper's own definition of the gene effect. The orthogonality construction in POG also lacks a clear mathematical guarantee in the main text. These issues affect the manuscript's central narrative, not just its presentation.","major_comments":[{"comment":"The DGE metric does not measure the gene effect as defined. The gene effect (Abstract; Section 3) is the event that resetting parameters to random values improves enhancement performance with respect to the ground truth. Eq. (6) computes the average PSNR between the original model output F(xj) and the reset-model output Fi(xj); it never involves the ground truth. A high DGE means the reset changes the output little, which is exactly the regime where a random reset cannot improve quality over the learned model. A low DGE means the reset changes the output substantially, which is a necessary condition for the gene effect (improvement) to occur, but it can also mean large degradation. The paper's interpretation in Section 5.2 ('A larger value of DGE ... indicating more gene effects. Conversely, a smaller DGE implies weaker gene effects.') is therefore backwards relative to the definition. Since every gene-effect-mitigation claim in Tables 3 and 4 and the ablation reasoning are based on DGE reduction, the central claim that PDE mitigates the gene effect is not supported by the presented evidence.","section":"Section 5.2, Eq. (6)"},{"comment":"The orthogonal basis construction is mathematically unclear and likely inconsistent. The paper defines Np as normalized embeddings of shape N × De and then writes Bp = I − 2 Np Np^T. With Np ∈ R^{N×De}, the product Np Np^T is N × N, so Bp is an N × N matrix, not the claimed Bp ∈ R^{N×De×De}. The Householder-reflection form I − 2 v v^T is defined for a single unit vector v, not for a matrix, and no argument is given for why this operation yields orthogonal bases for each parameter. The main text defers the 'theory guarantee' to the supplementary material, but the property is load-bearing for POG's claimed ability to 'prevent the excessive expression of similar parameters.' Please clarify the dimensions and provide a proof or a precise statement in the main text.","section":"Section 4.2, Eq. (2)"},{"comment":"The DGE values for Restormer+Ours are reported as 45.09 on both LOL-v1 and LOL-v2-real. Given that these are different test sets with different image distributions, obtaining identical values to two decimal places is implausible and suggests a typo or a calculation error. Please verify the reported DGE values and correct any mistakes, as this table is central to the gene-effect-mitigation claim.","section":"Table 3"},{"comment":"The observation that 'dynamic parameters exhibit weaker gene effects compared to static parameters' is presented through Table 2, but the table does not define what the numeric entries mean (e.g., PSNR change after reset? improvement fraction? units?) and the sign convention is ambiguous. Without a clear definition, the reader cannot assess the strength of the observation, and the subsequent causal attribution of the gene effect to static parameters in Section 3.1 rests on this unspecified evidence.","section":"Table 2"}],"minor_comments":[{"comment":"The text says 'gama correction' in the first paragraph; it should read 'gamma correction.'","section":"Section 2.1"},{"comment":"The DGE formula in Eq. (6) is introduced after the sentence 'we use the method in motivation experiments and observations (Section 3.1) to detect gene effect,' but Section 3.1 does not define DGE; consider moving the metric definition earlier or explicitly referencing Eq. (6) in Section 3.1.","section":"Section 5.2"},{"comment":"Entries such as '11.00 / ——-' and '7.98 / ——-' are used without explanation; please add a footnote stating what the dash denotes (e.g., failed to produce meaningful results or omitted due to collapse).","section":"Table 4"},{"comment":"The main quantitative tables report single-run PSNR/SSIM values without standard deviations or significance tests. Given that some improvements are small (e.g., 20.91 to 21.88 dB), please report variance across at least three runs or a statistical test to confirm the gains are not due to random variation.","section":"Tables 5–8"},{"comment":"The notation 'Np^T' is ambiguous: if Np is N × De, then Np^T is De × N, and the product Np Np^T is N × N as noted; please clarify the intended operation and correct the dimensions in Eq. (2) and the surrounding text.","section":"Section 4.2"},{"comment":"The figure caption lists numeric values (e.g., '12.24 dB 9.39 dB') without explaining what they represent; adding a legend or clarifying the PSNR values in each panel would improve readability.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical enhancement gains are plausible and the plug-and-play module could be a useful contribution, but the conceptual framing around the 'gene effect' is not validated by the DGE metric as defined. Authors should either correct the metric to measure the frequency or magnitude of random-reset improvements relative to ground truth, or substantially soften the gene-effect-mitigation claims. The identical DGE values in Table 3 and the unclear orthogonality construction in Eq. (2) are additional concerns that should be addressed before publication. If the gene-effect narrative is revised, the paper may still be acceptable for a vision venue; in its current form, the central claim is not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: the empirical observation is real and worth knowing. On standard low-light benchmarks, resetting some trained parameters to random values improves PSNR for a large fraction of test images; the authors document this with a POI table and examples. The proposed plug-in module POG is a credible dynamic-convolution variant with an orthogonality constraint, and it gives consistent gains of roughly 0.5–1 dB on LOL-v1/v2 when added to Restormer, Retinexformer, and CIDNet. That part is solid and useful for people in LLIE.\n\nThe problem is the metric used for the central claim. DGE in Eq. 6 is the PSNR between original and reset model outputs; it never touches ground truth. A reset that leaves outputs almost unchanged gives high DGE, and such a reset cannot produce the improvement that defines the gene effect. A reset that changes outputs a lot gives low DGE, and that is the only regime where random parameters can beat learned ones. So Section 5.2's interpretation—'larger DGE indicates more gene effects'—is backwards relative to the paper's own definition. The claim that PDE reduces DGE and therefore mitigates the gene effect is unsupported; if anything, their numbers show PDE makes outputs more sensitive to resets. The right measurement would be the percentage of improved images (their own Table 1) or average PSNR change against ground truth, before and after PDE.\n\nOther soft spots: no error bars or seeds, so the PSNR gains could be partly noise; Table 3 shows identical DGE values for LOL-v1 and LOL-v2-real for Restormer+Ours, which looks like a copy-paste slip; and Table 5 labels RUAS as [70] when [70] is KinD. The biological evolution analogy is decorative but not itself a problem.\n\nBottom line: the observation and the module are worth conditional acceptance, not rejection. But the gene-effect story needs a correct metric or it collapses. I would send this to a serious referee, and I would not cite it until the metric issue is fixed.","headline":"Real observation and a useful dynamic-conv module, but the gene-effect mitigation claim is unsupported because the DGE metric is interpreted backwards.","tokens_in":14714,"tokens_out":5003,"would_cite":false,"duration_ms":46775,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","68U10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Resetting trained parameters to random values can improve low-light enhancement for over 30% of images, and per-image parameter evolution counters the effect.","keywords":["low-light image enhancement","gene effect","dynamic parameters","parameter orthogonal generation","parameter resetting","static parameters","image restoration","transformer networks"],"falsifier":"On a fixed low-light enhancement model and test set, compute per-image PSNR for the original model and for models with each candidate layer reset to random values, and count how many images improve; if that fraction is near zero across layers, the gene effect as defined does not exist. For the PDE claim, repeat the same count after adding PDE: if the fraction of improving images does not decrease while DGE decreases, the paper's mitigation claim is measuring output sensitivity rather than the gene effect itself.","tokens_in":1652,"feed_emoji":"🧬","tokens_out":2202,"duration_ms":70707,"temperature":0.7,"pith_summary":"The paper claims that a counterintuitive \"gene effect\" limits low-light image enhancement: for more than 30% of test images, replacing certain well-trained network parameters with random values produces better enhancement, measured by PSNR. It argues that the cause is static parameters, since a single learned weight set applied to all images cannot fit a training set in which visually similar low-light inputs map to different bright images. Drawing an analogy to biological gene mutation and recombination, the paper proposes parameter dynamic evolution (PDE), a plug-in module that generates convolution parameters per input image, using parameter orthogonal generation (POG) to keep the generated parameter candidates from collapsing into a static set. Experiments on Restormer, Retinexformer, and CIDNet report reduced gene-effect levels and higher PSNR, such as Restormer rising from 20.91 to 21.88 dB on LOL-v1.","feed_headline":"Random parameters beat trained ones in low-light enhancement","feed_subtitle":"A gene-effect fix evolves model weights per image, lifting Restormer PSNR from 20.91 to 21.88 dB.","key_machinery":"The load-bearing mechanism is the parameter orthogonal generation (POG) technique. For each target parameter, POG learns an embedding, normalizes it, and builds an orthogonal basis via the reflection matrix $B_p = I - 2 N_p N_p^T$, so the basis embeddings are orthogonal to one another. Given an input feature, a two-layer MLP with softmax produces weights, and each parameter embedding is a weighted sum of the orthogonal bases, decoded by another MLP into the actual convolution weights. Orthogonality is what distinguishes PDE from ordinary dynamic convolution: it prevents the candidate parameter embeddings from becoming similar and collapsing to a static parameter set, which the paper identifies as the failure mode of prior dynamic-parameter methods. The PDE module itself is a plug-in bottleneck block inserted after attention in decoder stages, chosen because attention layers show the strongest gene effect.","core_discovery":"The paper's central claim is that static parameters are the source of the gene effect: after training, a fixed weight set is well matched to some images and maladaptive for others, to the point that random weights can beat the learned weights on particular inputs. The paper documents this by resetting attention-layer parameters of Restormer to random values and finding that over 30% of images improve, with up to 40% improving for the first detected layer. To counteract the effect, PDE evolves parameters for each input image: a bottleneck dynamic block generates per-image convolution weights, and POG constructs orthogonal basis embeddings so that the weighted combination of bases cannot converge to a single similar parameter set. Experiments report that adding PDE to existing low-light enhancement networks lowers the proposed DGE metric, defined as the average logarithmic MSE between original and reset outputs, and improves enhancement quality, with Restormer+Ours reaching 21.88 dB PSNR on LOL-v1 versus 20.91 dB for the base model.","pith_inferences":["The same random-reset probe could be run on other ill-posed image-restoration tasks, such as dehazing, deraining, or super-resolution, to test whether the static-parameter explanation is specific to low-light enhancement or general across restoration domains.","A sharper test than DGE would be to measure the fraction of images whose PSNR improves after random reset before and after PDE training; a method that truly removes the gene effect should lower that fraction, not merely change output distances.","The hyperparameter study suggests a natural extension: making the effective number of candidate parameters, controlled by the embedding dimension $D_e$ and bottleneck width $D_m$, adaptive per image rather than fixed.","If random parameter mutations are sometimes beneficial per image, a stochastic inference strategy that samples several parameter sets and selects the best output might outperform the single deterministic PDE output, though the paper does not explore this."],"forward_implications":["If static parameters cause the gene effect, any low-light enhancement model that applies one learned weight set to all images carries an inherent performance ceiling that random resetting can expose.","Adding PDE as a plug-in module requires about 10k fine-tuning steps versus 320k for original training, so existing models could be upgraded at low cost.","The gene effect appears across several architectures, including SNR-Net, LLformer, Retinexmamba, Retinexformer, and CIDNet, suggesting the phenomenon is general across low-light enhancement models.","Pruning methods do not remove the gene effect and can collapse enhancement quality on synthetic data, implying that the affected parameters cannot simply be deleted.","Orthogonal generation is the key technical ingredient: replacing POG with static convolutions or plain dynamic convolution gives smaller gains in the ablation study."],"supporting_citations":[{"why":"Base model whose attention parameters are reset to demonstrate the gene effect and whose LOL-v1 PSNR of 20.91 dB is the main baseline.","marker":"[67]"},{"why":"Supplies the dynamic-convolution mechanism that PDE compares against and that can degrade to static parameters.","marker":"[5]"},{"why":"Provides the conditionally parameterized convolution formulation cited alongside dynamic convolution as the prior dynamic-parameter approach.","marker":"[63]"},{"why":"Second architecture tested with PDE; the Restormer+Ours and Retinexformer+Ours rows give cross-architecture evidence.","marker":"[3]"},{"why":"Third architecture tested with PDE; the CIDNet+Ours row gives additional cross-architecture evidence.","marker":"[9]"},{"why":"Pruning baseline in Table 4 that fails to reduce the gene effect and degrades enhancement performance.","marker":"[14]"},{"why":"Pruning baseline whose performance collapses on the synthetic dataset, supporting the claim that affected parameters cannot be pruned.","marker":"[16]"},{"why":"Cited for the claim that dynamic parameters can correlate or collapse depending on initialization, motivating the orthogonal generation approach.","marker":"[4]"},{"why":"Source for the biological claim that orthogonal genetic information prevents excessive expression of similar genes, the inspiration for POG.","marker":"[34]"}],"fun_headline_variants":["Gene effect: why random weights beat learned ones in low-light","Per-image weight evolution fixes gene effect in low-light","Evolution of weights beats static parameters in low-light","Random weights win sometimes; dynamic evolution fixes that"],"cache_read_input_tokens":16896,"weakest_assumption_plain":"The load-bearing assumption is that the DGE metric, which averages the output difference between the original model and a model with parameters reset to random values, actually measures the gene effect, because a small DGE is interpreted as less gene effect even though the gene effect is defined as random reset improving some images.","fun_headline_variants_meta":{"raw":{"variants":["Gene effect: why random weights beat learned ones in low-light","Per-image weight evolution fixes gene effect in low-light","Evolution of weights beats static parameters in low-light","Random weights win sometimes; dynamic evolution fixes that"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001714,"raw_usage":{"total_tokens":6779,"prompt_tokens":937,"completion_tokens":5842,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":5779}},"tokens_in":553,"tokens_out":5842,"duration_ms":38478,"temperature":1.0,"reasoning_tokens":5779,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:37:04.362806+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a fixed low-light enhancement model and test set, compute per-image PSNR for the original model and for models with each candidate layer reset to random values, and count how many images improve; if that fraction is near zero across layers, the gene effect as defined does not exist. For the PDE claim, repeat the same count after adding PDE: if the fraction of improving images does not decrease while DGE decreases, the paper's mitigation claim is measuring output sensitivity rather than the gene effect itself.","supporting_citations":[{"cited_title":"Restormer: Efficient transformer for high-resolution image restoration","cited_arxiv_id":null,"evidence_quote":"Base model whose attention parameters are reset to demonstrate the gene effect and whose LOL-v1 PSNR of 20.91 dB is the main baseline."},{"cited_title":"Dynamic convolu- tion: Attention over convolution kernels","cited_arxiv_id":null,"evidence_quote":"Supplies the dynamic-convolution mechanism that PDE compares against and that can degrade to static parameters."},{"cited_title":"Condconv: Conditionally parameterized convolu- tions for efficient inference","cited_arxiv_id":null,"evidence_quote":"Provides the conditionally parameterized convolution formulation cited alongside dynamic convolution as the prior dynamic-parameter approach."},{"cited_title":"Retinexformer: One-stage retinex-based transformer for low-light image enhancement","cited_arxiv_id":null,"evidence_quote":"Second architecture tested with PDE; the Restormer+Ours and Retinexformer+Ours rows give cross-architecture evidence."},{"cited_title":"Learning both weights and connections for efficient neural network","cited_arxiv_id":null,"evidence_quote":"Pruning baseline in Table 4 that fails to reduce the gene effect and degrades enhancement performance."},{"cited_title":"Filter pruning via geometric median for deep convolutional neural networks acceleration","cited_arxiv_id":null,"evidence_quote":"Pruning baseline whose performance collapses on the synthetic dataset, supporting the claim that affected parameters cannot be pruned."},{"cited_title":"Frivolous Units: Wider Networks Are Not Really That Wide","cited_arxiv_id":"1912.04783","evidence_quote":"Cited for the claim that dynamic parameters can correlate or collapse depending on initialization, motivating the orthogonal generation approach."},{"cited_title":"The mechanism of mendelian heredity","cited_arxiv_id":null,"evidence_quote":"Source for the biological claim that orthogonal genetic information prevents excessive expression of similar genes, the inspiration for POG."}],"review_version":1}