{"id":"f15203f5-fbb6-4712-92a1-79c01fb77d7b","arxiv_id":"2506.16495","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A per-model scalar quantization transform aligns heterogeneous feature distributions so that a single learned codec can compress features from LLaMA3, DINOv2, and Stable Diffusion 3 with better rate-accuracy than task-specific baselines.","lead":"This paper introduces a pre-processing transform that reshapes feature values from different large AI models (LLaMA3, DINOv2, Stable Diffusion 3) into a shared 8-bit space, allowing one learned codec to compress features across models and tasks. The result is better rate-accuracy tradeoffs than task-specific codecs, including zero-shot transfer to new tasks and a CNN model.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Scalar transform fit on 10 features may not align joint feature statistics; universality claim rests on untested marginal-only alignment.","rationale":"The reader identified the 10-feature scalar transform as the weakest assumption; I agree that this is the least-secure link, but I frame the concern more sharply as a marginal-versus-joint alignment problem. The paper's own evidence for alignment is exclusively marginal: the KL divergences in Table 1 are computed on pooled scalar histograms, and Figure 2 visualizes pooled distributions. The codec, however, must predict the joint distribution of transformed feature tensors. A universal codec can generalize across models only if the full tensor statistics, not just the scalar marginals, are brought into a common space. The cross-task universality experiments in Section 5.4 provide some direct evidence that the approach works in practice, which is genuine credit: the CSR-trained codec applied to Seg and TTI features still outperforms baselines. Nevertheless, those experiments cover only three model-task combinations, with no error bars and no stability analysis over the 10-feature transform fit. The CNN transfer result in Section 5.7 is encouraging but uses a single architecture and a single task, so it does not close the gap. Thus the central claim is plausible and partly supported, but its most general form—effective generalization to arbitrary different models and tasks—remains conditional on an untested assumption about joint-distribution alignment. The proposed check would settle whether this concern lands; until then, keeping the conditional verdict is appropriate without escalating to rejection.","tokens_in":20049,"tokens_out":8806,"duration_ms":100764,"concrete_test":"Run the full pipeline with T fitted on 1, 5, 10, 50, and 200 randomly drawn features, with several seeds per count, and for each fit evaluate (a) zero-shot cross-task rate-accuracy, e.g., a CSR-trained Hyperprior codec applied to Seg and TTI transformed features, and (b) per-channel KL or MMD between transformed feature distributions across models. If the rate-accuracy curves shift materially across fits, or if per-channel divergence remains large while pooled KL is small, the scalar-transform assumption is load-bearing and the universality claim needs qualification. A comparison against per-channel transforms or direct CDF histogram equalization would further isolate whether the 1D pooled Lloyd-Max transform is necessary.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that a universal codec trained on the balanced target distribution generalizes across models and tasks rests on the sufficiency of the transform T in Section 4.2. T is a single 1D Lloyd-Max quantizer learned on the pooled scalar values of only 10 randomly selected feature tensors and applied element-wise to every coordinate. Equation (3) optimizes only marginal scalar MSE; Table 1 measures alignment via KL divergence between pooled histograms. But the codec operates on high-dimensional tensors with spatial layout and channel structure. A scalar transform can equalize marginals while leaving joint statistics—spatial autocorrelation, channel covariances, conditional dependencies—unaligned across models. The paper provides no per-channel distribution analysis and no sensitivity analysis for the 10-feature fit. If different channels have different scales or shapes, the single T maps them into different effective codebook regions, and the reported zero-shot gains may depend on the particular 10-feature draw. Every downstream result, including the cross-task universality curves and the CNN transfer experiment, inherits this unverified assumption.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses universal feature coding for large models, proposing DT-UFC: a per-model scalar transformation learned by Lloyd-Max quantization that maps raw feature values into 0..255, with the goal of making feature distributions more balanced and aligned across models. A single learned codec (Hyperprior or ELIC) is then trained on transformed features and evaluated on common-sense reasoning (LLaMA3), semantic segmentation (DINOv2), and text-to-image synthesis (SD3). Experiments include hybrid and task-specific training, zero-shot cross-task evaluation, a control using transformed features with an unmodified baseline codec, and a cross-architecture test on ResNet-50 features. The paper claims that the transform enables better rate-accuracy and cross-model generalization than task-specific baselines.","tokens_in":20291,"tokens_out":7134,"duration_ms":75498,"significance":"The problem is timely and the proposed solution is refreshingly simple: a plug-and-play scalar transform trained with a well-defined Lloyd-Max objective, plus shared codec training. Strengths include a clean formulation, release of source code, evaluation across three tasks, two codecs, hybrid/task-specific training, and zero-shot cross-task and cross-architecture settings, which makes the central empirical claim partly independent of circularity. The main weaknesses are the gap between the stated 'balanced' design criterion and the actual objective, and the absence of error bars and sensitivity analyses; these need to be addressed before the universality claim is fully supported.","major_comments":[{"comment":"The optimization objective in Eq. (3) minimizes reconstruction MSE only; it contains no term that encourages the 256 output bins to be equally occupied, so the 'balanced distribution' criterion stated in Section 4.2 is not actually enforced. For a highly peaky input, Lloyd-Max allocates many bins to high-density regions, and the transformed histogram generally remains far from uniform; Table 1 reports KL divergences between the three transformed distributions but not the divergence from uniformity, and Fig. 2 is illustrative rather than quantitative. Since the universality argument rests on the balanced target space, please either add an explicit balance measure (e.g., entropy of the transformed histogram or KL to uniform) and report it, or modify the transformation to enforce balancing, and rerun the key experiments to confirm the gains remain.","section":"Section 4.2, Eq. (3)"},{"comment":"The transform is fitted using only 10 randomly selected features, and it is applied element-wise to every scalar of the feature tensor. The paper does not report the size of these 10 features (number of scalar samples), any per-channel distributional analysis, or sensitivity of the downstream results to the particular 10-feature draw. If different channels have very different scales or shapes, a single scalar transform may align pooled marginals while leaving per-channel and spatial statistics misaligned, which could affect the universality claim. Please add: (i) the actual number of scalar samples used in fitting, (ii) a sensitivity study varying the number of features (e.g., 1, 10, 100) and repeated random draws, and (iii) if possible, per-channel KL or histogram comparisons.","section":"Section 4.2 and Section 5"},{"comment":"All rate-accuracy comparisons are reported as single runs without error bars or multiple seeds. Because learned codecs are stochastic and some reported differences are small (e.g., Table 3, TTI rows), the claim that DT-UFC 'consistently outperforms' baselines is not statistically supported. Please retrain with at least 3 seeds and report mean and standard deviation (or show dominant curves), and state the variance of the BPFP and accuracy/FID/mIoU metrics.","section":"Tables 2-4 and Fig. 5"},{"comment":"The isolation experiments in Table 5 are not presented at matched rates, which makes the claimed improvement ambiguous. For example, in the CSR rows the transformed-feature entries start at BPFP 2.08 while the baseline entries start at 1.35; in the TTI rows the transformed features use more bitrate than the baseline at similar FID values. Please interpolate both curves to a common set of rates (or equivalent operating points) before claiming that the transform alone improves rate-accuracy.","section":"Section 5.5, Table 5"},{"comment":"The bitrate term BPFP is never defined in the paper, and it is unclear whether it includes the cost of transmitting the transform parameters {c_k} (256 floats per model) or only the quantization indices. Please define BPFP precisely and state whether the transform side information is amortized across the test set; if it is not included, quantify the overhead and justify its omission.","section":"Section 5.2, Eq. (6)"}],"minor_comments":[{"comment":"The text refers to 'GTP-4'; this should be 'GPT-4'.","section":"Section 1"},{"comment":"The KL divergence formula contains a corrupted symbol ('˝') and should be written as a sum or integral.","section":"Section 3.2"},{"comment":"It would help to add a 'Uniform' column so the reader can see how far the transformed distributions are from the claimed balanced target.","section":"Table 1"},{"comment":"The caption and axis labels in the supplied version contain stray '/uni...' tokens; please ensure the final PDF renders correctly.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"I see no evidence of misconduct. The main concerns are the mismatch between the claimed 'balanced' target and the actual MSE objective, and the lack of statistical robustness in the empirical comparisons. If the authors add the requested sensitivity and significance analyses, the paper could become suitable for publication. The phrase 'first systematic study' should also be toned down unless the related-work positioning is strengthened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe real contribution here is modest but legitimate: the paper takes Lloyd-Max scalar quantization—classical 1950s signal processing—and uses it as a per-model pre-transform so a single learned codec can compress features from LLaMA3, DINOv2, and SD3. The empirical story is stronger than the theory, and that is okay.\n\nWhat the paper does well: the evaluation is broad for this area. Three tasks across reasoning, segmentation, and text-to-image; two codecs (Hyperprior and ELIC); hybrid and task-specific training; zero-shot cross-task transfer; and a ResNet-50 test. Code is public, which is real evidence. The KL-divergence numbers in Table 1 show the marginal distributions align, and the cross-task curves in Fig. 5 show the codecs actually generalize. The central claim holds up as demonstrated.\n\nWhere the soft spots are, in order of importance:\n\n- The 10-feature fit is the weakest link. A single scalar transform equalizes pooled marginals, but if feature channels have different scales or shapes, the codec sees a mixture of effective codebook regions. The paper reports no per-channel distribution analysis and no sensitivity analysis for the 10-feature draw. The stress-test note is right that joint statistics are unexamined. That said, the cross-task experiments are the real test, and they pass; so this is a gap in analysis, not a fatal flaw.\n\n- No comparison with plain histogram equalization, which is the obvious classical baseline for this idea.\n\n- No error bars or multiple seeds. For a method that begins by fitting a transform to 10 random features, this matters.\n\n- Bitrate does not include the cost of transmitting the 256 centroids per model. Likely negligible, but should be stated.\n\nWho this is for: anyone working on coding for machines, distributed inference, or feature compression. It deserves a serious referee. I would send it out; a minor revision should add a histogram-equalization baseline, a robustness check on the 10-feature draw, and clearer accounting of side information.","headline":"A modest but real contribution: Lloyd-Max scalar quantization as a per-model pre-transform enables one codec to serve heterogeneous large-model features, and the empirical evaluation is broad enough to support it.","tokens_in":20791,"tokens_out":2733,"would_cite":true,"duration_ms":28222,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A universally trained feature codec can compress and reconstruct intermediate representations from different large models and tasks, provided each model's features are first reshaped by a learned peaky-to-balanced distribution transform.","keywords":["universal feature coding","peaky-to-balanced distribution transformation","large models","feature compression","cross-model generalization","rate-accuracy","DINOv2","LLaMA3"],"falsifier":"Take DINOv2 features and compute per-channel histograms after applying the paper's transform learned from 10 features; if channels retain visibly different scales or shapes, or if re-learning the transform from 1000 features materially changes rate-accuracy, the single-scalar-transform premise is violated.","tokens_in":19895,"feed_emoji":"⚖️","tokens_out":4061,"duration_ms":38352,"temperature":0.7,"pith_summary":"The paper tries to establish that a single feature codec can serve many large models and tasks if the model-specific feature distributions are first reshaped into a common balanced space. It identifies distributional heterogeneity—peaky LLaMA3 and DINOv2 features versus flatter SD3 features—as the obstacle, and proposes a learned non-uniform scalar transform that maps every feature value into 256 bins of roughly equal mass. Once transformed, features from different models look alike to the codec, so one hybrid-trained codec generalizes across tasks and even architectures. The paper reports rate-accuracy gains over task-specific baselines and shows zero-shot cross-task compression works after alignment.","feed_headline":"One codec compresses features from many large models","feed_subtitle":"A learned peaky-to-balanced transform aligns LLaMA3, DINOv2, and SD3 features so one codec generalizes across tasks.","key_machinery":"The load-bearing object is the learned non-uniform scalar transformation $T: \\mathbb{R} \\to \\{0,1,\\dots,255\\}$ with inverse $T^{-1}(k) = c_k$. It partitions the original feature-value range into 256 regions and maps each to an integer, choosing the regions and centroids to minimize squared transformation distortion via alternating assignment and update steps, like scalar Lloyd-Max quantization. This reshapes peaky per-model histograms into a near-flat distribution over $\\{0,\\dots,255\\}$, so the codec sees inputs with a consistent range and near-uniform occupancy; the paper claims this improves latent utilization, stabilizes entropy modeling, and aligns distributions across models.","core_discovery":"The central claim is that distributional alignment, rather than a more powerful codec, is the key enabler of universal feature coding. The paper defines universal feature coding as compressing intermediate features from diverse large models with one codec and shows that directly training a codec on heterogeneous features fails because encoders internalize distribution-specific priors. Its proposed solution is the peaky-to-balanced transform $T: \\mathbb{R} \\to \\{0,1,\\dots,255\\}$, a data-driven partition of the feature value range into 256 intervals whose inverse values are the interval centroids; the transform is fit on only 10 randomly chosen feature vectors by alternating assignment and centroid update. After transformation, the KL divergences between feature distributions from different models drop sharply, and a codec trained on transformed features outperforms task-specifically trained baselines on common-sense reasoning, semantic segmentation, and text-to-image generation, and also compresses ResNet-50 features without retraining.","pith_inferences":["If the scalar transform generalizes to models beyond the three studied, the same plug-and-play module could be shipped as a per-model header of 256 centroids with one shared codec, a practical deployment pattern for distributed inference.","The 256-level target space suggests direct compatibility with existing image codecs and entropy coders; one could test whether the transform also improves coding of natural images whose histograms are strongly skewed.","Because the transform is fitted on only 10 features, per-channel or covariate shift is a plausible failure mode; a natural extension the paper does not explore is a per-channel or learned parametric transform.","The method's success on ResNet-50 implies the transform may act as a general normalizer for learned representations, potentially useful beyond coding, for example in federated or heterogeneous-device settings where models exchange intermediate tensors."],"forward_implications":["A single hybrid-trained codec can compress features from LLaMA3, DINOv2, and SD3 with better rate-accuracy than codecs trained per task.","The transform alone, applied before an untrained baseline codec, improves compressibility and accuracy in the CSR task and causes no degradation in the Seg and TTI tasks.","Codecs trained on transformed features generalize to unseen tasks in a zero-shot fashion; for example, models trained on Seg or TTI compress CSR features well after alignment.","The approach transfers across architectures: hybrid-trained codecs handle ResNet-50 features with near-lossless classification accuracy without retraining.","Codecs trained on transformed features remain effective on imperfect or truncated inputs, indicating they learned general feature structure rather than memorized input patterns."],"supporting_citations":[{"why":"Supplies the Hyperprior learned codec used as both baseline and DT-UFC's compression backbone, including the rate-distortion training objective and entropy model.","marker":"[6]"},{"why":"Provides the large-model feature coding benchmark, including the dataset, unified test conditions, feature packing method, and evaluation protocol.","marker":"[19]"},{"why":"Supplies the ELIC high-capacity learned codec used to verify that DT-UFC works across different codec architectures.","marker":"[23]"},{"why":"Source of the DINOv2 visual features whose highly peaky distribution motivates the transform in the segmentation task.","marker":"[41]"},{"why":"Source of the Stable Diffusion 3 features whose more uniform distribution is contrasted with DINOv2 and LLaMA3 features.","marker":"[12]"},{"why":"Source of the LLaMA features used in the common-sense reasoning task, representing the textual modality.","marker":"[54]"}],"fun_headline_variants":["One codec, all large-model features, via peaky-to-balanced","Peaky-to-balanced transform enables one universal codec","Universal feature coding: just align the distributions","One codec for LLaMA, DINO, and SD3 features","From peaky to balanced: one codec generalizes across models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"One scalar transform fit on only ten sampled feature vectors is assumed to capture the distribution of every feature channel and every sample from a model, so that all dimensions land in the same balanced space.","fun_headline_variants_meta":{"raw":{"variants":["One codec, all large-model features, via peaky-to-balanced","Peaky-to-balanced transform enables one universal codec","Universal feature coding: just align the distributions","One codec for LLaMA, DINO, and SD3 features","From peaky to balanced: one codec generalizes across models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00047,"raw_usage":{"total_tokens":2378,"prompt_tokens":1020,"completion_tokens":1358,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":1272}},"tokens_in":636,"tokens_out":1358,"duration_ms":12176,"temperature":1.0,"reasoning_tokens":1272,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:24:50.487669+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take DINOv2 features and compute per-channel histograms after applying the paper's transform learned from 10 features; if channels retain visibly different scales or shapes, or if re-learning the transform from 1000 features materially changes rate-accuracy, the single-scalar-transform premise is violated.","supporting_citations":[],"review_version":2}