{"id":"9ccad14b-1e94-4356-a718-88005f97327b","arxiv_id":"2501.01420","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Ladon is the first multi-task supervised compression model for split computing, sharing one encoder and one preprocessing pipeline to run classification, detection, and segmentation in a single inference.","lead":"This paper presents Ladon, a model that runs the early, compressed layers of a neural network on a mobile device and sends a tiny representation to an edge server that produces classification, detection, and segmentation outputs in one pass. Tests on ImageNet, COCO, and PASCAL VOC show Ladon matches or beats lightweight mobile models while cutting end-to-end latency by up to 95.4% and mobile energy use by up to 88.2% in multi-task edge scenarios.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unified preprocessing pipeline that underpins Ladon's single-inference efficiency is never specified or ablated; without it, the accuracy-parity claim cannot be checked.","rationale":"The reader's weakest assumption—that a single unified preprocessing pipeline is acceptable for all three tasks—is also the load-bearing point I identify. The paper's efficiency advantage comes from running one encoder and one transmission instead of three, and that is only valid if the same preprocessed input feeds all three task heads without a significant accuracy penalty. The authors explicitly claim 'without model accuracy degradation' but never specify the pipeline or provide an ablation against task-specific preprocessing for the same Ladon checkpoints. Without that specification, the accuracy-parity result is not independently verifiable, and the risk is concrete: object detection and semantic segmentation are known to be sensitive to input resolution and aspect ratio, while image classification is conventionally center-cropped. I considered other potential concerns. The missing full-offloading baseline is less load-bearing because at the evaluated 100 Kbps and 37.5 Kbps channel rates, transmitting a raw image is dominated by communication delay and would be noncompetitive. The 'end-to-end' overclaim about training is a naming/scope issue rather than a threat to the measured latency and energy improvements. Missing error bars would matter for the device measurements, but the reported savings are large in magnitude and directionally robust. The unified preprocessing issue is the one that, if real, would undermine the central 'multi-task without degradation' claim. The reader already reached CONDITIONAL, and my analysis does not move that verdict; it sharpens the specific condition that should be satisfied before the claim is accepted.","tokens_in":13326,"tokens_out":4389,"duration_ms":46361,"concrete_test":"Download the released Ladon code and checkpoints; inspect the actual inference preprocessing (resolution, resize/crop, aspect-ratio handling) used for the shared encoder. Then evaluate the COCO 2017 detection head and PASCAL VOC 2012 segmentation head under two conditions: (a) the released unified preprocessing, and (b) the task-specific preprocessing used for the Table 2 baselines (e.g., aspect-preserving resize to 320/416) with the same checkpoints. If the mAP or mIoU difference between (a) and (b) exceeds roughly 1-2 points, the claim that the unified pipeline serves all tasks 'without model accuracy degradation' is unsupported; if the difference is within noise, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central functional claim—one shared encoder and one transmission serving classification, detection, and segmentation with no accuracy loss—rests on the unified preprocessing pipeline asserted in Section 4.1. The paper never states what this pipeline is: no unified input resolution, no cropping/resizing policy, no aspect-ratio handling, and no mapping of detection/segmentation outputs back to the original image. The Table 2 baselines use task-specific input shapes: 224x224 center-crop for classification, 300/320 and 320+/416+ aspect-preserving inputs for detection/segmentation. If Ladon's unified pipeline uses a fixed square resize/crop (e.g., 224x224 or 320x320), small-object detection and pixel-level segmentation are at direct risk, since Faster R-CNN with FPN and DeepLabv3 are resolution- and aspect-ratio-sensitive. Section 6.1's qualitative statement that segmentation is only 'comparable' is consistent with this risk. Moreover, 'without model accuracy degradation' is never tested against the same Ladon heads with task-specific preprocessing; the reported comparisons are only against different baseline models in Table 2, so the phrase conflates 'rivals lightweight models' with 'the unified pipeline does not degrade accuracy.' The public code and models make this testable, but as written the claim is underspecified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Ladon, a supervised compression model for split computing that shares a single encoder and a unified preprocessing pipeline across three vision tasks: image classification (ILSVRC 2012), object detection (COCO 2017), and semantic segmentation (PASCAL VOC 2012). The encoder is executed on a mobile device and transmits a compressed representation to an edge server, which runs task-specific heads. The authors evaluate predictive performance against lightweight single-task baselines, and measure end-to-end latency and mobile energy consumption on NVIDIA Jetson devices, reporting large reductions relative to local computing and to a split-computing baseline (Entropic Student). The paper claims to present the first end-to-end multi-task supervised compression model for split computing.","tokens_in":13488,"tokens_out":5068,"duration_ms":48510,"significance":"If the central claims hold, Ladon would be a useful contribution: a single encoded transmission serving three tasks in one inference, with publicly available code and models, and real-device measurements of latency and energy. The use of the SC2 benchmark and the inclusion of Jetson measurements are strengths. However, several load-bearing points are underspecified or unsupported as written, notably the unified preprocessing pipeline, the 'end-to-end' training description, and the baseline selection for isolating the effect of the proposed design choices.","major_comments":[{"comment":"The unified preprocessing pipeline is never specified. The text describes typical task-specific preprocessing (square center-crop for classification, aspect-preserving resizing for detection and segmentation) but does not state which policy Ladon uses: the unified input resolution, the resize/crop rules, how aspect ratio is handled, or how detection/segmentation outputs are mapped back to the original image coordinates. The claim that this unified pipeline works 'without model accuracy degradation' is also not tested against a control: the only predictive performance comparisons in Table 2 are against different single-task lightweight baselines with task-specific preprocessing. To support the central claim that one encoded transmission can serve all tasks, the authors should specify the pipeline and report an ablation comparing Ladon with the unified pipeline against Ladon with task-specific preprocessing for each head.","section":"Section 4.1 (Unified preprocessing)"},{"comment":"The paper calls Ladon an 'end-to-end multi-task supervised compression model' (Introduction, Section 1) and an 'end-to-end multi-task model' (Figure 1), but the training procedure in Steps 1-3 is sequential: encoder-decoder pre-training, then classification fine-tuning, then freezing all parameters and separately fine-tuning the detection and segmentation heads. There is no joint multi-task loss and no simultaneous optimization of the shared representation for the three tasks. The phrase 'end-to-end multi-task' is therefore not supported by the described training protocol. Please either revise the claim to 'multi-task-head supervised compression model' or provide a genuinely joint training scheme and demonstrate its impact.","section":"Section 4.3 (Training) / Introduction"},{"comment":"The predictive performance evaluation uses only lightweight single-task baselines (MobileNetV2/MobileNetV3, MNASNet, SSD, DeepLab, LRASSP) as references, as summarized in Table 2. This does not isolate the effects of the two central design choices: the unified preprocessing pipeline and the shared multi-task backbone. To support the claim that the unified pipeline does not degrade accuracy and that parameter sharing does not hurt task-specific performance, the paper should add ablations comparing Ladon against: (a) Ladon with task-specific preprocessing for each head, (b) the same backbone (ResNet-50 or ResNeSt-269e) trained as independent single-task SC2 models without sharing, and (c) full-offloading execution with the same backbones. Without these controls, the reported 'matching or outperforming' results cannot be attributed to the proposed architecture and preprocessing choices.","section":"Section 5.2 / Table 2 / Figures 2-4"}],"minor_comments":[{"comment":"The latency and energy experiments report point estimates without confidence intervals or number of repeated runs. Given that the central quantitative claims are reductions of up to 95.4% in latency and 88.2% in energy, adding error bars or standard deviations would strengthen the reliability of these measurements.","section":"Section 6.2 / 6.3"},{"comment":"The paper uses the notation gϕ both for the 'remaining layers' that produce a prediction and for the decoder in Equation (1). Clarifying this notation would avoid ambiguity about whether the decoder is part of the encoder-side or the task-specific modules.","section":"Section 4.3 / Supplementary A"},{"comment":"The claim that 'Ladon models either outperformed or rivaled' the baselines is supported by Figures 2-4, but the semantic segmentation results are described as 'comparable' rather than superior. Please state explicitly in the text which Ladon configurations (backbone and β) achieve which qualitative conclusion on each dataset, since the figures alone do not convey the exact performance ranges.","section":"Section 6.1"},{"comment":"The comparison with Chimera is described only in terms of its bottleneck assumption. Since Ladon is claimed to be the 'first' end-to-end multi-task supervised compression model, consider adding a sentence explaining explicitly why Chimera (which also serves multiple tasks with a shared feature extractor) does not fall under this definition, beyond the bottleneck criterion.","section":"Section 3 / Introduction"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and leverages the authors' prior SC2 benchmark work, but the 'first' claim and the unified-preprocessing design need to be checked carefully against the presented evidence. The main issues are verifiability and attribution of gains: the pipeline is unspecified, the training is not end-to-end as described, and the baselines do not isolate the design choices. These are fixable with additional experiments and revised claims, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take on arXiv:2501.01420 (Ladon). The paper is worth your time if you work on split computing or edge inference. The core idea is new: unlike Entropic Student, which shares an encoder but runs separate preprocessing and separate inferences per task, Ladon shares one encoder, one preprocessing pipeline, and one forward pass to serve classification, detection, and segmentation. The reported latency and energy reductions on Jetson hardware are large (up to ~95% latency, ~88% energy), and the encoder is tiny (0.5-0.9 MB). Code and trained models are public. That is real, reproducible evidence.\n\nThe softest spot is exactly where the stress-test points: the unified preprocessing pipeline is asserted but never specified. We are not told the input resolution, the cropping/resizing policy, the aspect-ratio handling, or how detection/segmentation outputs are mapped back to the original image. The baselines use task-specific shapes (224x224, 300/320, 416+). If Ladon uses a fixed square crop, small-object detection and pixel-level segmentation are at direct risk. And the phrase 'without model accuracy degradation' is never tested against the same heads with task-specific preprocessing; it is only compared to different baselines. This is a load-bearing gap because the entire runtime advantage comes from one encoded transmission serving all tasks.\n\nThe 'end-to-end' label also overstates the training: it is three sequential steps, with the encoder frozen before the detection and segmentation heads are added. That is a shared-encoder approach, not joint multi-task training. Minor, but the wording should change.\n\nNo error bars on the latency and energy numbers, and no full-offloading baseline, though the local-computing and split-computing baselines are reasonable. The segmentation results are honestly described as 'comparable,' which fits the concern about resolution.\n\nOverall, this is a well-executed systems paper with a genuine architectural novelty and public code. The central engineering claim—one encoder, one transmission, large savings—is well supported. The accuracy-parity claim is plausible but not fully checkable as written. I'd send it to review, but with a firm request to specify the preprocessing pipeline, ideally ablate it against task-specific preprocessing, and report variance.\n\nRecommendation: worth a serious referee.","headline":"A solid multi-task split-computing architecture with real hardware gains, but the unified preprocessing that underpins the single-inference claim is never specified.","tokens_in":14072,"tokens_out":3445,"would_cite":true,"duration_ms":33125,"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":"Ladon's single shared encoder serves three vision tasks at once, cutting end-to-end latency by up to 95.4% and mobile energy by up to 88.2% while matching lightweight baselines.","keywords":["split computing","supervised compression","multi-task learning","edge computing","rate-distortion tradeoff","image classification","object detection","semantic segmentation"],"falsifier":"Run Ladon's detection and segmentation heads on COCO 2017 and PASCAL VOC 2012 under the paper's unified preprocessing, then rerun the same heads under each task's conventional preprocessing (full image, aspect ratio preserved, no center crop). If the unified-pipeline versions drop by more than a point or two of mAP or mIoU, especially on small objects and boundary pixels, the paper's claim that the unified pipeline costs no accuracy is falsified. A simpler probe: feed Ladon's detection head a standard 224×224 center-cropped input and compare with the same head on full-resolution input; any large drop isolates the preprocessing assumption rather than the shared encoder.","tokens_in":13051,"feed_emoji":"⚡","tokens_out":14597,"duration_ms":117283,"temperature":0.7,"pith_summary":"Split computing distributes a neural network across a weak mobile device and a stronger edge server: the device runs the early layers, compresses the intermediate representation, and transmits it, while the server finishes the job. The paper proposes Ladon, which it presents as the first end-to-end multi-task supervised compression model for that setting. Ladon shares one encoder and one image preprocessing pipeline across image classification, object detection, and semantic segmentation, so a single inference produces all three outputs. The authors report that Ladon matches or outperforms lightweight single-task baselines on ILSVRC 2012, COCO 2017, and PASCAL VOC 2012, while cutting end-to-end latency by up to 95.4% and mobile-device energy by up to 88.2% compared with baselines. If correct, this removes a barrier to running several vision tasks simultaneously on battery-powered, bandwidth-limited edge systems.","feed_headline":"One shared encoder cuts multi-task edge latency 95.4%","feed_subtitle":"One tiny encoder serves all three tasks in a single inference, cutting latency and energy.","key_machinery":"Supervised compression—learning a compressed representation that carries task-relevant information rather than reconstructing the input—is implemented here by an early-layer encoder $f_\\theta$ that maps an input image to a quantized representation $\\hat{z}$, and a decoder-plus-heads module $g_\\phi$ that maps $\\hat{z}$ to predictions. Training minimizes a rate-distortion loss, $$L_{\\mathrm{pre}}(x) = \\sum_{i \\in I} \\|h^t_i(x) - h^s_i(x)\\|$_2^{2}$ - \\$\\beta$ \\log p_\\psi(f_\\$\\theta$(x) + \\epsilon),$$ where the first term distills teacher embeddings into student layers, the second term charges for the bit cost of the quantized representation via an entropy model, and $\\epsilon \\sim \\mathrm{Unif}(-1/2, 1/2)$ simulates quantization noise. The structural novelty is that one encoder, one shared backbone, and one unified preprocessing pipeline feed three heads—a classification head, a Faster R-CNN with FPN detection head, and a DeepLabv3 segmentation head—trained in three steps: pretrain the encoder-decoder by distillation, fine-tune for classification, then freeze and attach the extra heads.","core_discovery":"On the authors' own terms, the central claim is that supervised compression need not be trained per task: a single encoder trained to emit small, task-relevant representations at the early layers can feed shared backbone parameters and three task-specific heads, eliminating the duplicate local computation, duplicate transmission, and separate preprocessing of prior split-computing approaches. Concretely, with ResNet-50 and ResNeSt-269e backbones and a rate-distortion loss, Ladon's encoders are only 0.543--0.935 MB, roughly 0.268--2.29% of the total multi-task local model size, and the paper reports accuracy on ILSVRC 2012 and COCO 2017 that is competitive with or better than mobile-friendly baselines, with semantic segmentation comparable. On low-power mobile devices with a 100 Kbps or LoRa-rate wireless link, the measured end-to-end latency falls by up to 95.4% and mobile-device energy by 65.0--88.2% relative to baselines. The paper's conclusion is that a deliberate multi-task supervised compression design, not a collection of task-specific models sharing a generic encoder, is the efficient way to serve multi-task inference in split computing.","pith_inferences":["The same one-encoder recipe could plausibly extend to related dense prediction heads such as depth estimation or panoptic segmentation, since those tasks also consume shared backbone features; the paper itself only demonstrates classification, detection, and segmentation.","Because semantic segmentation is the weakest of the three reported results, the number of tasks one encoder can serve is likely bounded by the most resolution-sensitive task in the mix, not by the average task.","A component-wise ablation—shared encoder only versus shared encoder plus unified preprocessing—would separate how much of the 95.4% latency saving comes from avoiding duplicate backbones and how much from avoiding duplicate preprocessing.","The unified preprocessing pipeline is underspecified in the paper; a reader reproducing the work should first pin down the exact resizing and cropping rule, since it determines whether the accuracy-parity claim holds on small objects."],"forward_implications":["A multi-task query costs about the same communication as a single-task query: one compressed representation $\\hat{z}$ is transmitted once and decoded into classification labels, bounding boxes, and segmentations.","Edge-server memory and compute shrink because one shared backbone replaces three independently trained backbones, letting a fixed server host more tasks or more concurrent users.","On mobile devices, the energy savings of 65.0--88.2% directly extend battery life, and the paper's LoRa-rate experiments show the latency advantage persists even at 37.5 Kbps.","The hyperparameter $\\beta$ in the rate-distortion loss becomes a system-level knob, letting operators trade compressed data size against task accuracy while keeping the encoder small.","The three-step training recipe—pretrain with distillation, fine-tune for classification, then freeze and attach extra heads—is a reusable path for adding new task heads to an existing compressed encoder."],"supporting_citations":[{"why":"Supplies the SC2 benchmark's evaluation metrics (encoder size, compressed data size, accuracy) and the empirical finding that supervised compression beats input and feature compression for single-task split computing.","marker":"[24]"},{"why":"Defines the entropic student supervised compression method that Ladon extends and serves as the strongest accuracy, latency, and energy baseline.","marker":"[23]"},{"why":"Provides the ResNet-50 backbone that is modified into the Ladon encoder-decoder and used as the ImageNet teacher model.","marker":"[10]"},{"why":"Supplies the entropy model and prior for quantized representations used in the rate-distortion loss that trains the compressed encoder.","marker":"[3]"},{"why":"Provides the Faster R-CNN architecture used as the object detection head attached to the shared backbone.","marker":"[30]"},{"why":"Provides the feature pyramid network combined with Faster R-CNN for the detection head.","marker":"[15]"},{"why":"Provides the DeepLabv3 architecture used as the semantic segmentation head.","marker":"[4]"},{"why":"Defines the ILSVRC 2012 classification dataset and pretraining task that anchors the encoders and the first task.","marker":"[31]"},{"why":"Defines the COCO 2017 object detection dataset used to train and evaluate the detection head.","marker":"[16]"},{"why":"Defines the PASCAL VOC 2012 semantic segmentation dataset used for the segmentation head.","marker":"[8]"}],"fun_headline_variants":["Single encoder for multi-task split computing cuts latency 95%","Shared encoder: three tasks, one pass, 95% lower latency","Multi-task split computing: one encoder, 95% latency drop","Ladon: compact shared encoder trims edge latency by 95%","One tiny encoder serves all tasks: 95% less latency on edge"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single image preprocessing pipeline—one resizing and cropping choice—can serve image classification, object detection, and semantic segmentation without meaningful accuracy loss; the paper asserts this but never specifies the pipeline, and the three tasks normally favor different preprocessing.","fun_headline_variants_meta":{"raw":{"variants":["Single encoder for multi-task split computing cuts latency 95%","Shared encoder: three tasks, one pass, 95% lower latency","Multi-task split computing: one encoder, 95% latency drop","Ladon: compact shared encoder trims edge latency by 95%","One tiny encoder serves all tasks: 95% less latency on edge"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000362,"raw_usage":{"total_tokens":1965,"prompt_tokens":971,"completion_tokens":994,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":900}},"tokens_in":587,"tokens_out":994,"duration_ms":9675,"temperature":1.0,"reasoning_tokens":900,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:27:59.858319+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Ladon's detection and segmentation heads on COCO 2017 and PASCAL VOC 2012 under the paper's unified preprocessing, then rerun the same heads under each task's conventional preprocessing (full image, aspect ratio preserved, no center crop). If the unified-pipeline versions drop by more than a point or two of mAP or mIoU, especially on small objects and boundary pixels, the paper's claim that the unified pipeline costs no accuracy is falsified. A simpler probe: feed Ladon's detection head a standard 224×224 center-cropped input and compare with the same head on full-resolution input; any large drop isolates the preprocessing assumption rather than the shared encoder.","supporting_citations":[{"cited_title":"SC2 Benchmark: Supervised Compression for Split Computing","cited_arxiv_id":null,"evidence_quote":"Supplies the SC2 benchmark's evaluation metrics (encoder size, compressed data size, accuracy) and the empirical finding that supervised compression beats input and feature compression for single-task split computing."},{"cited_title":"Supervised Compression for Resource- Constrained Edge Computing Systems","cited_arxiv_id":null,"evidence_quote":"Defines the entropic student supervised compression method that Ladon extends and serves as the strongest accuracy, latency, and energy baseline."},{"cited_title":"Deep Residual Learning for Image Recognition","cited_arxiv_id":null,"evidence_quote":"Provides the ResNet-50 backbone that is modified into the Ladon encoder-decoder and used as the ImageNet teacher model."},{"cited_title":"Variational image compression with a scale hyperprior","cited_arxiv_id":null,"evidence_quote":"Supplies the entropy model and prior for quantized representations used in the rate-distortion loss that trains the compressed encoder."},{"cited_title":"Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks","cited_arxiv_id":null,"evidence_quote":"Provides the Faster R-CNN architecture used as the object detection head attached to the shared backbone."},{"cited_title":"Feature Pyramid Networks for Object Detection","cited_arxiv_id":null,"evidence_quote":"Provides the feature pyramid network combined with Faster R-CNN for the detection head."},{"cited_title":"Berg, and Li Fei-Fei","cited_arxiv_id":null,"evidence_quote":"Defines the ILSVRC 2012 classification dataset and pretraining task that anchors the encoders and the first task."},{"cited_title":"Microsoft COCO: Common Objects in Context","cited_arxiv_id":null,"evidence_quote":"Defines the COCO 2017 object detection dataset used to train and evaluate the detection head."},{"cited_title":"The PASCAL Visual Object Classes Challenge 2012 (VOC2012)","cited_arxiv_id":null,"evidence_quote":"Defines the PASCAL VOC 2012 semantic segmentation dataset used for the segmentation head."}],"review_version":1}