{"id":"1186227a-e861-492f-97df-d1afdc9eb2b6","arxiv_id":"2608.06691","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A new efficient transformer backbone, CoDAT, achieves competitive action recognition accuracy at a fraction of the latency and energy of existing models on edge hardware.","lead":"CoDAT is a lightweight transformer that combines strided single-head attention, convolutional attention, and temporal shifts to recognize human actions on low-power edge devices. It reports competitive accuracy with much lower latency and energy than prior models on standard image and video benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central accuracy-efficiency claim is plausible, but the 6x speedup over TokShift/LAPS rests on unverified cross-model benchmark parity; ONNX Runtime precision/execution-provider settings are not disclosed and published latency numbers appear inconsistent with strict FP32 throughput.","rationale":"I read the paper in good faith and found the architecture plausible: the SSHA complexity analysis, TShift placement study, and the spatial/temporal ablations are internally consistent, and the FLOPs reductions are arithmetically coherent. The load-bearing condition for the central claim is that the reported speedups are measured under identical deployment conditions. The paper does not disclose enough about the ONNX Runtime export, precision, execution provider, and baseline measurement protocol to verify this, and the reader's weakest assumption identifies the same issue. This does not invalidate the design; it means the empirical Pareto claim is not yet independently verifiable. The reader's CONDITIONAL verdict remains the right call, so I recommend no change.","tokens_in":27151,"tokens_out":14759,"duration_ms":153782,"concrete_test":"Reproduce Table 7 with a pinned ONNX Runtime build and one execution provider for all models (e.g., CUDA EP, FP32, batch=1, num_threads=4, 50 W power cap) using the released ONNX exports for CoDAT-S384, TokShift, and LAPS; then repeat the same benchmark under FP16/TensorRT to check whether the speedup is provider-dependent. If TokShift's latency under strict FP32 differs materially from 10.15 ms, or if CoDAT's speedup over the baselines drops below the claimed 5x under the same provider, the headline claim requires revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest claim is the UCF-101 result (Table 7): CoDAT-S384 matches TokShift and LAPS at 95.4% Top-1 while running 6.3x/5.8x faster and using 13.1x/3.9x fewer FLOPs. The FLOPs arithmetic is straightforward, but the latency/energy half of the claim is load-bearing and is not protocol-fair in the text as written. Section 4.1 states only that inference uses ONNX Runtime, CUDA on Jetson AGX Orin, and a 50 W power cap; it does not state numerical precision (FP32/FP16/TF32/INT8), ONNX optimization level, or whether a TensorRT execution provider was used for CoDAT or for the baselines. This matters concretely: TokShift's reported 10.15 ms at 135 GFLOPs implies roughly 13 TFLOPS, at or beyond the AGX Orin's practical FP32 envelope, so at least some models appear to run on a non-default, possibly lower-precision path. If the baselines were optimized more aggressively than CoDAT, or vice versa, the '6x faster' ratio could shift materially. Accuracy point estimates are also single-run values from a mix of published and self-reported protocols; the zero-margin '95.4 vs 95.4' match has no variance or seed information. The architecture itself is internally coherent—the SSHA complexity bookkeeping, TShift placement ablation, and repeated spot checks in Tables 8 and 9 support the efficiency story—so the weak point is external benchmark parity, not the design.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript introduces CoDAT, a three-stage vision transformer for edge action recognition. The central module, CoDA, combines a strided single-head attention (SSHA) branch that reduces spatial tokens and value channels with a spatial convolutional attention (SCA) branch, and a parameter-free temporal shift (TShift) is inserted before the final ConvFFN of each block. The paper evaluates CoDAT on ImageNet-1K, Kinetics-400, MA-52, and UCF-101, and reports latency, throughput, and energy on Jetson AGX Orin and Raspberry Pi 5. The authors claim that CoDAT matches or approaches the accuracy of much heavier CNN and ViT baselines while reducing FLOPs, latency, and energy substantially, with the strongest claims made for UCF-101, where CoDAT-S384 is reported to match TokShift and LAPS at 95.4% top-1 accuracy while running about 6x faster.","tokens_in":27547,"tokens_out":10907,"duration_ms":98493,"significance":"The architecture is internally coherent and the empirical scope is broader than in many edge action-recognition papers: four datasets, two hardware platforms, an explicit energy-measurement procedure, and systematic ablations of stride ratio, channel ratio, shift placement, and stage design. The SSHA complexity bookkeeping and the TShift receptive-field argument are plausible, and the internal ablations in Tables 8 and 9 support the claim that CoDAT's efficiency comes from the strided and channel-compressed attention design rather than from simple model shrinkage. If the efficiency-accuracy claims survive a fair cross-model benchmark protocol, the paper would be a useful contribution to edge IoT perception. However, the abstract and conclusion currently contain quantitative claims that do not match the tables, and the latency/energy comparisons are not yet protocol-fair as written, so the central claim cannot be accepted without substantial revision.","major_comments":[{"comment":"The latency and energy comparisons are load-bearing but the runtime configuration is incomplete. Section 4.1 states only that inference uses ONNX Runtime with CUDA on Jetson AGX Orin at a 50 W power cap; it does not report numerical precision (FP32/FP16/TF32/INT8), ONNX optimization level, or whether a TensorRT execution provider was used. This matters concretely: in Table 5, the TokShift row reports 135 GFLOPs at 10.15 ms, which implies about 13.3 TFLOPS sustained, well above the practical FP32 throughput of a 50 W AGX Orin; the baselines therefore appear to have been run on a faster non-default path while the paper does not state whether CoDAT used the same one. Unless the same ONNX export options, execution provider, and precision are documented for every model, the '6x faster', '2.9x faster', and energy ratios in the abstract, conclusion, and Section 4.3 are not established. Please provide a per-model runtime configuration table or restrict the speed claims to models measured under identical settings.","section":"Section 4.1, Table 5"},{"comment":"Several quantitative claims in the abstract and conclusion do not match the tables. On Kinetics-400, the conclusion says CoDAT is 2.9x faster than VSwin-T and nearly 2x faster than LAPS/TokShift; Table 5 gives CoDAT-L384 at 4.74 ms/F versus VSwin-T at 6.96 ms/F (1.47x), and CoDAT-M384 at 2.56 ms/F versus LAPS at 9.34 ms/F and TokShift at 10.15 ms/F (about 3.6x and 4.0x). On MA-52, the conclusion reports 63.6% fine-grained accuracy and latency reductions of more than 5x and 9x relative to VSwin-T and UniFormer-B, while Table 6 gives 63.19% and ratios of 2.72x and 5x. In the image-classification conclusion, CoDAT-L is said to have 2.5x lower latency than ViT-S, while Table 4 gives 1.71 ms / 1.33 ms = 1.29x. Please correct all abstract and conclusion numbers to agree with the tables, or rerun the measurements and update the tables.","section":"Abstract, Conclusion, Tables 4-6"},{"comment":"The cross-model benchmark parity is not sufficiently documented. Tables 4-7 mix self-reported and published baseline numbers, and Section 4.1 says multi-view configurations are applied 'exclusively for top-1 accuracy evaluation based on each model default configuration,' but it is unclear which view configuration corresponds to each baseline row and whether the baselines were run by the authors or taken from the literature. Several rows, such as S2AFormer-mini in Table 4 and I3D, STM, and MViT in Table 5, have no latency or energy entries, so the efficiency comparisons cover different model subsets. To make the central speed and energy claims benchmark-fair, please specify for every baseline the source of the accuracy number (authors' run vs published), the exact view/crop protocol, and the runtime configuration used for latency and energy.","section":"Section 4.1, Tables 4-7"},{"comment":"The accuracy results appear to be single-run point estimates, which is problematic for claims that hinge on small differences. The central UCF-101 claim rests on CoDAT-S384 exactly matching TokShift and LAPS at 95.4% (Table 7), and several Table 4 conclusions rely on 0.1-0.4% accuracy differences (e.g., CoDAT-S versus SHViT-S3, or CoDAT-L versus Swin-T). Without multiple seeds, confidence intervals, or at least a stated fixed-seed protocol, these differences are within typical run-to-run noise, especially for a small dataset like UCF-101. Please report variance or multiple seeds for the main comparisons, or soften claims that depend on sub-1% differences.","section":"Tables 4 and 7"}],"minor_comments":[{"comment":"The row labeled 'TShift1/4C' in Table 9 reports a latency of 7.2 ms while all neighboring rows report 0.89 ms and the FLOPs are unchanged; this is almost certainly a typo and should be corrected.","section":"Table 9"},{"comment":"The text says SSHA gives 'a reduction of r_s^4-times over the standard full-resolution O(N^2 C)', but since the value channels are C_v = C/4, the actual reduction is 4 r_s^4. Please correct the wording or state explicitly that the factor of 4 is absorbed into the channel reduction.","section":"Section 3.1.2"},{"comment":"There are several typos and inconsistent labels, including 'VSwim-T' for VSwin-T in Section 4.3.5, the stray 'Trans.' row in Table 5, and the phrase 'the official train/test splits 1.' Please copyedit the final version.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The architecture and internal ablations are convincing enough that I would be willing to support acceptance after a revision that (i) corrects the abstract/conclusion numbers, (ii) provides a complete per-model runtime configuration, and (iii) either adds variance information or softens the sub-1% accuracy claims. The paper fits the IoTJ scope, but for a deployment-oriented journal the latency/energy measurement protocol must be fully reproducible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CoDAT is a solid engineering paper. The architecture combines strided single-head attention with a convolutional spatial-attention branch and a temporal shift module, all standard ideas but assembled in a way that yields a genuinely light backbone. The experimental work is the strongest part: four datasets, two edge devices, energy measurements, and a thorough ablation that isolates the contributions of each design choice. I believe the central claim—that you can get near-SOTA video accuracy at a fraction of the compute and power—is plausible, and the ablations support it.\n\nThe main soft spot is external benchmark parity. The 6.3x speedup over TokShift on UCF-101 depends on latency numbers for all models measured under identical conditions. Section 4.1 says only that inference uses ONNX Runtime with CUDA on the Jetson, capped at 50W. It does not state whether the baselines ran in FP32, FP16, or with a TensorRT execution provider, and it does not say whether the same optimization level was used. That matters because a 2x difference can come from precision alone. The stress-test note's specific arithmetic about TokShift's throughput appears off—135 GFLOPs is per clip, so per frame it is ~17 GFLOPs, and 10.15 ms latency implies ~1.7 TFLOPS, which is within the Orin's FP32 envelope—so that example doesn't prove a problem. But the missing protocol documentation remains a genuine concern.\n\nThere are also minor issues: no error bars on the 95.4% matches, some abstract numbers don't line up exactly with the tables, and the baselines mix published and measured values without careful footnoting. The GitHub link is given, but the paper does not include the evaluation scripts or the exact ONNX conversion settings, so reproducing the 6x claim would be difficult.\n\nWho this is for: practitioners building edge action-recognition systems will get a useful reference point, and the ablation study is a good model for how to report efficiency work. The novelty is not high, but the contribution is real. I'd send it to a serious referee with the request to nail down the inference protocol and either release the benchmark scripts or clearly mark which latency numbers were taken from prior papers. A desk reject would be too harsh.","headline":"Competent edge action-recognition paper with strong ablations; the 6x speedup is plausible but the ONNX inference protocol is under-documented.","tokens_in":28073,"tokens_out":3834,"would_cite":false,"duration_ms":36685,"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":"A dual-attention transformer brings near-state-of-the-art action recognition to edge hardware at a fraction of the energy cost.","keywords":["action recognition","vision transformer","edge computing","temporal shift","efficient attention","IoT","single-head attention","energy efficiency"],"falsifier":"Re-measuring the baselines from Tables 4-7 under the exact same inference settings as CoDAT (same single-clip, single-crop input, same exported model format, same hardware and power metering, same warm-up and repetition protocol) would settle the claim. In particular, if TokShift and LAPS on their public checkpoints run at less than 6x the latency and less than 13x the FLOPs of CoDAT-S384 on UCF-101, the headline efficiency claim fails.","tokens_in":27004,"feed_emoji":"⚡","tokens_out":6983,"duration_ms":56643,"temperature":0.7,"pith_summary":"The paper claims that accurate video action recognition can run in real time on low-power IoT hardware if the standard multi-head self-attention is replaced by a collaborative dual-branch design: a strided single-head attention that compresses both spatial and channel dimensions for global context, paired with a convolutional spatial attention for local detail. A zero-parameter temporal shift placed after the attention block supplies inter-frame motion information at negligible cost. On ImageNet-1K, Kinetics-400, MA-52, and UCF-101, the resulting CoDAT family matches or approaches state-of-the-art accuracy while running several times faster and consuming far less energy per frame than comparable CNN, transformer, and hybrid baselines on Jetson AGX Orin and Raspberry Pi 5. If these results hold, they shift the practical deployment envelope for action recognition toward battery- and power-limited edge devices in surveillance, healthcare, and industrial monitoring.","feed_headline":"Action recognition on edge chips at up to 13x lower cost","feed_subtitle":"A dual-attention transformer with strided heads and zero-parameter temporal shifts keeps accuracy while slashing latency and energy.","key_machinery":"The load-bearing object is the Collaborative Dual-Attention (CoDA) module and its placement relative to the temporal shift. Strided Single-Head Attention (SSHA) projects the input through pointwise convolutions with a stride $r_s$, reducing the attention from $N$ tokens to $M = (H/r_s) \\times (W/r_s)$ tokens with value channels $C_v = C/4$, so the attention cost drops from $O(N^2 C)$ to $O(M^2 C_v)$. Spatial Convolutional Attention (SCA) computes a single-channel saliency mask from max and average pooling through a $3\\times3$ convolution and sigmoid, costing $O(N C)$, and multiplies it into the features. A learnable projection fuses the two branches. Around each CoDA block, a zero-parameter TShift layer permutes a small fraction of channels across adjacent frames and is placed after CoDA but before the final ConvFFN, so the attention branches always see temporally clean single-frame features while the feed-forward network integrates the shifted temporal context.","core_discovery":"CoDAT's central claim is that the accuracy–efficiency frontier for edge action recognition can be moved by jointly attacking the two dominant costs of video transformers: token count and channel redundancy. The paper shows that a single attention head over spatially strided queries, keys, and values, combined with a lightweight convolutional attention branch, recovers the global and local cues that full multi-head attention provides, while a parameter-free temporal shift captures motion between frames. Empirically, CoDAT-S384 matches TokShift and LAPS on UCF-101 at roughly 6-times lower latency and up to 13-times fewer FLOPs, CoDAT-M runs about 2-times faster than EfficientViT-384 and FastViT-S12 on ImageNet-1K at comparable accuracy, and on Kinetics-400 CoDAT is up to 2.9-times faster than Video Swin Transformer and about 2-times faster than temporal-shift ViT variants while staying within one percentage point of their Top-1 accuracy. The efficiency gains persist on the fine-grained MA-52 benchmark, where CoDAT-M384 reaches the highest fine-grained Top-1 accuracy among the compared models.","pith_inferences":["If the energy and latency ratios hold under fully standardized benchmarking, the practical meaning is that always-on camera analytics on battery-powered nodes becomes plausible; the implied operating point is several frames per second to real time on boards that previously could only run lightweight CNNs.","The fixed stride ratio $r_s$ per stage is a knob the paper leaves untouched; a content-adaptive stride that increases compression on easy frames could push the efficiency gains further than the static design the paper evaluates.","The placement principle found here—keep attention branches temporally clean and shift channels only into the feed-forward network—suggests a general recipe for shift-based video transformers that could be tested on other lightweight backbones.","A natural testable extension is streaming deployment: the paper measures per-frame cost on fixed clips, but batched frame shifting with a buffer would show whether the $2L+1$ temporal receptive field behaves as claimed under continuous input."],"forward_implications":["Real-time action recognition becomes practical on low-power hardware: on Jetson AGX Orin, CoDAT-S runs UCF-101 inference at 0.89 ms/frame with 13.79 mJ/frame energy.","The efficiency gains transfer from image to video: CoDAT-L matches ViT-S/DeiT-S on ImageNet-1K with about 3x fewer parameters and 2.8x higher throughput.","On Kinetics-400, CoDAT-M384 outperforms or matches CNN baselines such as SlowFast and transformer baselines such as TokShift while requiring over 2x fewer GFLOPs and running 2.75x to 3.9x faster.","On the fine-grained MA-52 benchmark, CoDAT-M reaches the best fine-grained Top-1 accuracy among the compared models while running 2.7x faster than SlowFast and 5x faster than UniFormer-B.","The approach also works on compressed video: on MPEG-4-compressed UCF-101, CoDAT-S384 surpasses MTRFN by 0.8 percentage points at 8.8x fewer FLOPs."],"supporting_citations":[{"why":"Supplies the cascaded group attention baseline and the training recipe and Local Aggregation design that CoDAT builds on.","marker":"[48]"},{"why":"Motivates single-head attention by showing multi-head redundancy, and provides a key comparison baseline (SHViT).","marker":"[50]"},{"why":"Contributes the temporal shift operation that CoDAT adopts as the parameter-free TShift module.","marker":"[40]"},{"why":"Provides the structural reparameterization approach and the FastViT baseline used for efficiency comparisons.","marker":"[47]"},{"why":"The shift-based ViT baseline (TokShift) that CoDAT matches on UCF-101 with lower latency and FLOPs.","marker":"[28]"},{"why":"The long-range attention plus periodic shift baseline LAPS that CoDAT compares against on Kinetics-400 and UCF-101.","marker":"[29]"},{"why":"Video Swin Transformer, the transformer baseline CoDAT claims to beat by up to 2.9x in speed.","marker":"[26]"},{"why":"UniFormer, the unified convolution-attention baseline for Kinetics-400 and MA-52 comparisons.","marker":"[27]"},{"why":"Supplies the Kinetics-400 dataset on which CoDAT's video accuracy is measured.","marker":"[17]"},{"why":"Supplies the UCF-101 dataset used for transfer and compressed-video evaluation.","marker":"[16]"}],"fun_headline_variants":["Edge action recognition up to 13x cheaper with dual-attention transformer","Strided attention and temporal shift make edge video AI fast","CoDAT: low-cost dual-attention for real-time edge action recognition","Efficient edge transformer: 2x faster than EfficientViT and FastViT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the published baseline numbers were produced under conditions close enough to CoDAT's own training and single-clip inference settings to be directly comparable; if the baselines used different training recipes, resolutions, or multi-view protocols, the claimed accuracy and speed advantages could shrink or vanish.","fun_headline_variants_meta":{"raw":{"variants":["Edge action recognition up to 13x cheaper with dual-attention transformer","Strided attention and temporal shift make edge video AI fast","CoDAT: low-cost dual-attention for real-time edge action recognition","Efficient edge transformer: 2x faster than EfficientViT and FastViT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000344,"raw_usage":{"total_tokens":1996,"prompt_tokens":1157,"completion_tokens":839,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":773,"completion_tokens_details":{"reasoning_tokens":759}},"tokens_in":773,"tokens_out":839,"duration_ms":8286,"temperature":1.0,"reasoning_tokens":759,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:21:34.352277+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-measuring the baselines from Tables 4-7 under the exact same inference settings as CoDAT (same single-clip, single-crop input, same exported model format, same hardware and power metering, same warm-up and repetition protocol) would settle the claim. In particular, if TokShift and LAPS on their public checkpoints run at less than 6x the latency and less than 13x the FLOPs of CoDAT-S384 on UCF-101, the headline efficiency claim fails.","supporting_citations":[{"cited_title":"Effi- cientvit: Memory efficient vision transformer with cascaded group attention,","cited_arxiv_id":null,"evidence_quote":"Supplies the cascaded group attention baseline and the training recipe and Local Aggregation design that CoDAT builds on."},{"cited_title":"Shvit: Single-head vision transformer with memory efficient macro design,","cited_arxiv_id":null,"evidence_quote":"Motivates single-head attention by showing multi-head redundancy, and provides a key comparison baseline (SHViT)."},{"cited_title":"Tsm: Temporal shift module for efficient video understanding,","cited_arxiv_id":null,"evidence_quote":"Contributes the temporal shift operation that CoDAT adopts as the parameter-free TShift module."},{"cited_title":"Fastvit: A fast hybrid vision transformer using structural reparameteriza- tion,","cited_arxiv_id":null,"evidence_quote":"Provides the structural reparameterization approach and the FastViT baseline used for efficiency comparisons."},{"cited_title":"Token shift transformer for video classification,","cited_arxiv_id":null,"evidence_quote":"The shift-based ViT baseline (TokShift) that CoDAT matches on UCF-101 with lower latency and FLOPs."},{"cited_title":"Long-term leap attention, short-term periodic shift for video classification,","cited_arxiv_id":null,"evidence_quote":"The long-range attention plus periodic shift baseline LAPS that CoDAT compares against on Kinetics-400 and UCF-101."},{"cited_title":"Video swin transformer,","cited_arxiv_id":null,"evidence_quote":"Video Swin Transformer, the transformer baseline CoDAT claims to beat by up to 2.9x in speed."},{"cited_title":"Uniformer: Unifying convolution and self-attention for visual recognition,","cited_arxiv_id":null,"evidence_quote":"UniFormer, the unified convolution-attention baseline for Kinetics-400 and MA-52 comparisons."}],"review_version":1}