{"id":"3aee7950-6b54-4c1f-b7be-7e3d3345f75f","arxiv_id":"2608.08698","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"Shuffling video tokens by importance and scheduling them over fading channel blocks, then reconstructing from surviving tokens, improves perceptual video quality under packet loss compared with DCVC, VVC, and H.265.","lead":"This paper proposes a wireless video system that packs video tokens into shuffled packets, prioritizes them by importance and channel quality, and lets a trained decoder fill in lost pieces. It reports better perceptual quality and slower degradation under packet loss than standard video codecs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Detokenizer is trained on i.i.d. packet erasures up to 20%, but the block-fading test produces correlated whole-block losses and reaches PER 30%; this distribution shift is the least-secure support for the claimed graceful degradation.","rationale":"I read the paper as an empirical systems claim: the proposed token packetization, scheduling, and fine-tuned reconstruction beat four codecs perceptually at comparable channel bandwidth and degrade more gracefully under loss. To be true, all three components must transfer to the block-fading testbed. The packetization and scheduler are deterministic and can be inspected; the trained detokenizer is the only learned component whose behavior under the deployment channel is not directly verified. The paper explicitly trains with random packet masks up to 20% and evaluates under channel-dependent erasure up to 30%, but does not show that the test mask distribution is covered by training. This is a distribution-shift risk, not an inconsistency with external consensus. The proposed check, retraining on block-structured masks generated by the actual scheduler, directly tests whether the mismatch is consequential. If the curve is unchanged, the concern is resolved and the reader's conditional verdict can be upgraded; if the curve worsens, the central claim needs revision. This matches the reader's weakest_assumption, so agreement is 'agree' and the verdict stays CONDITIONAL rather than being hardened or relaxed.","tokens_in":7977,"tokens_out":12510,"duration_ms":138668,"concrete_test":"Fine-tune a second detokenizer with exactly the same loss and hyperparameters, but generate training masks by simulating the deployed block-fading scheduler: sample h_b, schedule packets as in Sec. III-C, and erase every packet in blocks whose realized BLER/outage condition is triggered, sweeping the PER target to 0.3. Then re-plot Fig. 4. If the block-trained LPIPS curve lies within 0.02 of the reported i.i.d.-trained curve at PER >= 0.2, the training mismatch is not load-bearing; if it deviates by more than 0.05 or WVTC's margin over DCVC-RT disappears, the reported graceful degradation rests on an untested distribution shift.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Sec. V-B is that WVTC degrades more gracefully than DCVC-RT, DCVC-FM, VVC, and H.265 as PER increases. For this claim to hold, the fine-tuned detokenizer must reconstruct well from the particular missing-token patterns produced by block fading after CSI-aware scheduling. In Sec. IV the detokenizer is trained by erasing whole packets independently with ratio r_m ~ U(0,0.2). In the test channel, packets in the same fading block share the same channel coefficient h_b; a deep fade raises the erasure probability of every packet in that block together, so losses arrive in block-correlated groups rather than as independent packet drops. The supplied Fig. 4 goes to PER=0.30, beyond the 0.20 maximum seen in training. Because the receiver is given no explicit mask and must infer missing tokens from zero-filled positions, a correlated erasure of an entire scheduled block is structurally different from a random 20% packet mask; the network may not have learned to inpaint that pattern. No block-correlated ablation, error bars, or code is provided, so the reported graceful-degradation curve is not yet tied to the actual test-time loss distribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a loss-resilient wireless video token communication (WVTC) framework for block fading channels. It combines content-intrinsic token importance scoring, a shuffled mixed I/P-token packetization scheme, and an online CSI-aware packet scheduler to disperse important and correlated tokens across packets and fading blocks. At the receiver, a fine-tuned erasure-aware detokenizer reconstructs the video from surviving tokens without retransmission. The authors report improvements in LPIPS and FID over DCVC-RT, DCVC-FM, VVC, and H.265 across channel bandwidth ratios, and claim more graceful perceptual degradation as the packet error rate increases.","tokens_in":8269,"tokens_out":4896,"duration_ms":50784,"significance":"If the results hold, WVTC offers a practical no-retransmission design for robust video transmission over block fading channels, leveraging the structure of learned discrete video tokens. The problem is well motivated and the proposed components—importance scoring via temporal novelty, packetization to disperse structural and temporal information, and CSI-aware scheduling with an importance-concentration penalty—are novel and technically coherent. The paper also makes a useful contribution by fine-tuning a pretrained VidTok detokenizer to reconstruct from erased token tensors. However, the central performance claims currently rest on thin evidence: single point estimates, a training/evaluation distribution mismatch, and an ambiguous zero-fill sentinel. Strengths of the manuscript include its clear system model and the plausible mechanism for graceful degradation; the main weaknesses are in the empirical validation and one reconstruction-ambiguity issue.","major_comments":[{"comment":"The erasure-aware detokenizer is fine-tuned under i.i.d. whole-packet erasures with masking ratio rm ~ U(0, 0.2), but the test channel is block fading, which produces correlated erasures of all packets within a fading block, and the PER axis in Fig. 4 extends to 0.30, beyond the training range. The claim of more graceful degradation in Section V-B therefore depends on the unverified assumption that the detokenizer generalizes to block-correlated loss patterns. Please provide either block-correlated fine-tuning, an ablation with block-correlated erasures, or error bars over channel realizations to support this central claim.","section":"Section IV and Section V-B, Fig. 4"},{"comment":"The receiver fills erased token positions with zeros and provides no explicit mask to the detokenizer. Since the FSQ codebook has cardinality |K| = 4096 and typically contains the index 0 as a valid token, the zero-fill sentinel may collide with a legitimate token value; the detokenizer would then be unable to distinguish an erased position from a genuine token 0. The paper does not specify the token index range or otherwise address this ambiguity. Please clarify that 0 is not in the codebook, or introduce a dedicated out-of-codebook mask token.","section":"Section II-C and Section IV"},{"comment":"The performance comparisons are single point estimates with no error bars, no repeated seeds, and no confidence intervals, despite the stochastic channel model and random masking during training. The statements that WVTC achieves the 'lowest or near-lowest' LPIPS/FID and exhibits 'more graceful perceptual degradation' are not supported with statistical evidence. Report means and variances over multiple independent runs and channel realizations, and release the code and evaluation protocol for reproducibility.","section":"Section V-B, Figs. 3 and 4"},{"comment":"The importance score (Eq. 8) and the scheduler (Eq. 14) depend on hand-set constants alpha_1, alpha, M, kappa, beta, lambda_h, w_adv, and Q. No sensitivity analysis or ablation is provided to show that the reported gains are robust to these choices. At minimum, vary lambda_h and M and report the resulting changes in Fig. 3 and Fig. 4, so that the reader can assess the sensitivity of the claimed improvements to these hyperparameters.","section":"Section III-A and Section V-A"}],"minor_comments":[{"comment":"Several author names contain stray spaces ('Y ongjeong Oh', 'Y ongpeng Wu'); please fix these typographical issues.","section":"Author list"},{"comment":"Figure 1 appears to contain raw integer text ('31 9 14 45 ...') that looks like a rendering artifact; please check and replace with a properly rendered schematic.","section":"Fig. 1"},{"comment":"The text states 'average gamma_b = 8 dB', but gamma_b is defined in Eq. (3) as the instantaneous SNR. Please clarify that the average is E[gamma_b] or phrase it as 'average SNR of 8 dB'.","section":"Section V-A"},{"comment":"The number of Monte Carlo channel realizations used in Figs. 3 and 4 is not specified; please add this to the experimental setup.","section":"Section V-A"},{"comment":"Figure 4 reports only LPIPS under PER, while the abstract and conclusion claim improvements in perceptual quality more broadly. Consider also plotting FID under PER to match the stated claims.","section":"Fig. 4"},{"comment":"The UVG and HEVC dataset sources are not cited; please add the appropriate references for reproducibility.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper proposes an interesting framework and the system design is coherent, but the empirical validation is currently too weak to support the central claims. The zero-fill sentinel issue and the training/evaluation distribution mismatch are directly addressable in a revision, and I do not see them as irreparable. I recommend major revision rather than rejection, with the expectation that the authors provide stronger statistical evidence and address the reconstruction ambiguity."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real system paper with a novel joint design for video TokenCom, but the empirical case has a training/test mismatch the authors don't address. Worth a serious look, not ready as-is.\n\nWhat's actually new: the importance metric uses I-token structural anchors and P-token temporal novelty, avoiding text prompts or an extra vision-language model. The shuffled mixed I/P packetization and the online CSI-aware scheduler with an importance-concentration penalty are coherent and not off-the-shelf. The erasure-aware detokenizer fine-tuning is a sensible way to let the receiver inpaint missing tokens without retransmission. The paper compares against DCVC-RT, DCVC-FM, VVC, and H.265 under the same MCS setup, which is more than most TokenCom papers do, and the reported gains are plausible.\n\nSoft spots, in proportion. The main one: the detokenizer is trained by erasing whole packets with a uniform random masking ratio up to 0.2, but the test channel is block fading, so losses are correlated within fading blocks, and the PER curve reaches 0.30. A deep fade can wipe out most of a block. That is structurally different from a random 20% mask, and the receiver gets no mask, so it must infer missing tokens from zero-filled positions. The paper has no ablation with correlated erasures, so the graceful-degradation claim is unverified generalization. That is the load-bearing weakness. Next, the paper motivates itself against Morphe and Video TokenCom but never compares to them. Also no error bars, repeated seeds, code, or data, so the figures are point estimates. And using zero as the missing-token sentinel is risky because 0 may be a valid codebook index; the paper doesn't test for collision. Hand-set hyperparameters are normal and not a concern.\n\nIf the authors close the training/test erasure gap, add the TokenCom baselines, and release at least evaluation code, this becomes a solid contribution. Right now it's a plausible system with an unproven empirical claim.\n\nThis paper is for people working on token-based semantic communications or learned video transmission. It deserves peer review, not a desk reject, but reviewers should push on the distribution mismatch. I'd want to see the revision before relying on the numbers.\n\nRecommendation: send to review, with a request for a block-correlated erasure ablation and comparison to Morphe and Video TokenCom. If the receiver can't handle correlated losses, the central claim doesn't hold.","headline":"Serious video-TokenCom system paper with a genuinely new packetization/scheduler design, but its erasure-trained receiver is tested on correlated block-fading losses and the missing TokenCom baselines make the empirical claim unproven.","tokens_in":8798,"tokens_out":4119,"would_cite":false,"duration_ms":41304,"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":"Token video survives fading without retransmission","keywords":["token communication","video transmission","block fading","unequal error protection","packet scheduling","erasure-aware reconstruction","perceptual quality","wireless video"],"falsifier":"Measure LPIPS at the same average packet error rate under two loss patterns: i.i.d. random packet erasures and block-fading erasures where an entire fading block is lost at once. If WVTC's LPIPS under block-fading erasures is substantially worse than under i.i.d. erasures at the same PER, then the erasure-aware detokenizer has not generalized to correlated losses, and the reported graceful degradation is an artifact of the training-mask distribution.","tokens_in":7764,"feed_emoji":"🎥","tokens_out":5773,"duration_ms":55874,"temperature":0.7,"pith_summary":"This paper argues that wireless video encoded as discrete tokens can be made loss-resilient without retransmission, even over block fading channels where a deep fade wipes out whole blocks of packets. The proposed WVTC framework scores each token's importance from the video's own predictive structure, shuffles structural and temporally correlated tokens across packets, and allocates packets to fading blocks using only the current channel state. A receiver-side detokenizer fine-tuned on randomly erased packets reconstructs missing content from the surviving tokens. If this works, it means perceptually strong wireless video with bounded latency and no re-transmission overhead, degrading gracefully as packet errors climb.","feed_headline":"Token video survives fading without retransmission","feed_subtitle":"Shuffling structural and motion tokens across packets lets receivers rebuild lost frames from survivors.","key_machinery":"The central object is the token importance map plus the packetizer-scheduler-detokenizer pipeline built on it. I-tokens, which anchor structure, are given top importance; P-token importance is measured as temporal neighborhood novelty, a weighted cosine similarity between each token and a 3x3 neighborhood in the preceding token group. The packetizer mixes a cyclic-strided traversal of I-tokens with a pseudorandom round-robin of P-tokens so that structural anchors and correlated temporal regions are spread across packets using only a shared seed. The online scheduler greedily selects packets by importance density times a sigmoid-mapped decoding reliability, minus a penalty for concentrating important packets in the same fading block, using current-block CSI alone. At the receiver, a detokenizer fine-tuned on whole-packet erasures fills zeroed token positions from surrounding context, and a super-resolution network restores the final frames.","core_discovery":"On its own terms, the paper's central claim is that importance-aware shuffled packetization plus online channel-aware packet scheduling plus an erasure-aware detokenizer jointly make token-based video transmission perceptually better and more loss-resilient than both learned and conventional video codecs at comparable channel bandwidth ratios. Numerically, WVTC reaches the lowest or near-lowest LPIPS and FID across most operating points and exhibits more graceful perceptual degradation as packet error rate increases, because important and temporally correlated tokens are dispersed rather than clustered, and important packets are steered away from unreliable fading blocks. The reconstruction is done from surviving tokens alone, with no retransmission.","pith_inferences":["The same importance metric could be repurposed for content-aware bitrate selection or adaptive modulation choices per packet, not just packet scheduling.","Fine-tuning the detokenizer on block-fading-correlated erasure masks, where an entire fading block is erased at once, would likely close the train-test mismatch and could make the reported gains hold at higher packet error rates.","The shuffled mixed-token packetization idea transfers to other tokenized media where a few structural tokens anchor many detail tokens, such as images or point clouds.","In multi-user or broadcast settings, the scheduler might prioritize packets carrying shared structural tokens across users, improving group reconstruction efficiency."],"forward_implications":["WVTC operates without retransmission, so end-to-end latency stays bounded while packet losses occur; this suits real-time wireless video.","Because packetization and scheduling require no extra position signaling or textual prompts, the gains come without sending side information beyond the packet ID and CRC.","The learned codec baselines suffer increasingly severe error propagation as reference information is lost, while WVTC's perceptual quality falls off more slowly with packet error rate.","Increasing channel bandwidth ratio improves WVTC most in medium- to high-rate regions, where more informative packets can be delivered and dispersed."],"supporting_citations":[{"why":"Supplies the frozen discrete video tokenizer and the trainable detokenizer that WVTC adapts for erasure robustness.","marker":"[5]"},{"why":"Supplies the super-resolution network that produces the final reconstructed frames from the detokenized low-resolution video.","marker":"[6]"},{"why":"Provides the natural video clips used to fine-tune the detokenizer under random whole-packet erasures.","marker":"[7]"},{"why":"Defines the channel bandwidth ratio metric that equalizes channel-resource comparison across all benchmark schemes.","marker":"[8]"},{"why":"A learned real-time video compression baseline whose perceptual quality WVTC is compared against.","marker":"[9]"},{"why":"A feature-modulated learned video compression baseline used in the rate and packet-error comparisons.","marker":"[10]"},{"why":"The conventional video coding standard baseline used as a reference for the perceptual quality comparisons.","marker":"[11]"},{"why":"Supplies the 5G MCS lookup table, LDPC coding, QAM, and Rayleigh block-error-rate estimates used in the fading-link simulation.","marker":"[12]"}],"fun_headline_variants":["Shuffled video tokens survive block fading without resend","Loss-resilient token video: no retransmission, just shuffle","Wireless video tokens: importance-aware packing defeats fading","Token packet shuffling makes video loss-resilient over fading"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The receiver-side detokenizer is trained on independent, random whole-packet erasures with a fixed masking ratio, but in the tested channel a deep fade erases all packets scheduled in a fading block together; the claim relies on the detokenizer generalizing to those correlated, block-level loss patterns.","fun_headline_variants_meta":{"raw":{"variants":["Shuffled video tokens survive block fading without resend","Loss-resilient token video: no retransmission, just shuffle","Wireless video tokens: importance-aware packing defeats fading","Token packet shuffling makes video loss-resilient over fading"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00017,"raw_usage":{"total_tokens":1212,"prompt_tokens":836,"completion_tokens":376,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":452,"completion_tokens_details":{"reasoning_tokens":306}},"tokens_in":452,"tokens_out":376,"duration_ms":4318,"temperature":1.0,"reasoning_tokens":306,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:26:54.834618+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure LPIPS at the same average packet error rate under two loss patterns: i.i.d. random packet erasures and block-fading erasures where an entire fading block is lost at once. If WVTC's LPIPS under block-fading erasures is substantially worse than under i.i.d. erasures at the same PER, then the erasure-aware detokenizer has not generalized to correlated losses, and the reported graceful degradation is an artifact of the training-mask distribution.","supporting_citations":[{"cited_title":"VidTok: A versatile and open-source vid eo tokenizer,","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen discrete video tokenizer and the trainable detokenizer that WVTC adapts for erasure robustness."},{"cited_title":"Swinir: Image restorati on using swin transformer,","cited_arxiv_id":null,"evidence_quote":"Supplies the super-resolution network that produces the final reconstructed frames from the detokenized low-resolution video."},{"cited_title":"Video enh ance- ment with task-oriented flow,","cited_arxiv_id":null,"evidence_quote":"Provides the natural video clips used to fine-tune the detokenizer under random whole-packet erasures."},{"cited_title":"Towards practical real-time neural video compression,","cited_arxiv_id":null,"evidence_quote":"A learned real-time video compression baseline whose perceptual quality WVTC is compared against."},{"cited_title":"Neural video compression with featu re modulation,","cited_arxiv_id":null,"evidence_quote":"A feature-modulated learned video compression baseline used in the rate and packet-error comparisons."},{"cited_title":"Overview of the versatile video codin g (VVC) standard and its applications,","cited_arxiv_id":null,"evidence_quote":"The conventional video coding standard baseline used as a reference for the perceptual quality comparisons."}],"review_version":1}