{"id":"c72012d7-87ce-48c4-8396-926fbf133d19","arxiv_id":"2607.17482","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"GenTrans reduces video transmission bitrate by reusing cross-clip memory and runtime state at a generative receiver, claiming 25% bitrate savings over the prior GVC system at similar LPIPS quality.","lead":"This paper proposes Generative Transmission (GenTrans), a framework for sending video over very low-bandwidth, unreliable networks by transmitting compact generative conditions and reusing memory and compute state at the receiver. It claims roughly 25% bitrate savings over the authors' prior generative codec at similar perceptual quality, plus faster decoding and graceful degradation under packet loss.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 25% bitrate reduction from GenMem is confounded: Table 1 compares GenTrans (memory + state reuse + robust training + AFT) against GVC at different operating points, with no ablation isolating GenMem.","rationale":"The Reader's verdict (REJECT) is correct, and the reader's rationale already notes the lack of a controlled ablation of GenMem. However, the reader's weakest_assumption focuses on memory desynchronization under packet loss, which is an important but secondary issue: it affects robustness in lossy scenarios, while the missing ablation undermines the paper's headline bandwidth-saving claim even in ideal conditions. My concern is therefore more fundamental and partially overlaps with the reader's stated rationale but not with their stated weakest_assumption. The concrete test of running GVC or GenTrans-without-memory at 0.006 bpp would directly settle whether the observed bitrate savings are causal. The paper provides no code/data and no statistical significance, reinforcing the rejection, but the single most load-bearing point is the confounded experimental design that fails to isolate the core mechanism.","tokens_in":8957,"tokens_out":5659,"duration_ms":54176,"concrete_test":"Run GVC at 0.006 bpp on the full MCL-JCV dataset (adjusting its rate control to this bitrate) and compute LPIPS. Separately, run GenTrans with GenMem disabled (e.g., setting m_t to a constant or skipping the memory branch) at 0.006 bpp. If either achieves LPIPS ≈ 0.185 or better, the 25% bitrate reduction is not attributable to GenMem, and the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 (Table 1) claims 'additional bitrate reduction of approximately 25%' (0.008 → 0.006 bpp) at comparable LPIPS (0.180 vs 0.185), and attributes this to GenMem. But GenTrans differs from the GVC baseline in multiple components: cross-clip memory, runtime state reuse, packet-loss-robust training, and AFT. The comparison is at different bitrate operating points, not at matched LPIPS. There is no GVC run at 0.006 bpp, and no GenTrans-without-GenMem ablation. Table 2 is a single selected low-motion video ('representative example'), so it cannot establish a general effect and is subject to selection bias. Without a controlled comparison that isolates GenMem, the observed bitrate difference could be due to any other component or to rate-control calibration, making the central claim — that cross-clip memory reuse cuts transmission cost by ~25% at equal perceptual quality — unsupported. The reader's weakest_assumption about memory desync under packet loss is also valid, but the ablation gap is more immediately load-bearing because it undermines the headline efficiency claim even under ideal network conditions.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GenTrans, a system-level framework for video transmission under ultra-low bandwidth and weak-network conditions. It builds on Generative Video Compression (GVC) and adds three mechanisms: cross-clip generative memory (GenMem) to reduce bitrate, runtime state reuse to reduce decoding latency, and AI Flow Transport (AFT) plus packet-loss-robust training for robustness. Experiments on the MCL-JCV 720p dataset report bitrate-LPIPS comparisons, a decoding-latency speedup, and packet-loss robustness. The central claim is that GenMem yields an additional ~25% bitrate reduction at comparable perceptual quality, state reuse gives a 1.5× decoding speedup, and the system degrades gracefully up to 50% packet loss.","tokens_in":9367,"tokens_out":6410,"duration_ms":57263,"significance":"If the claimed 25% bitrate reduction from cross-clip memory reuse and the 1.5× decoding speedup were properly validated, the paper would make a useful contribution to generative video communication in ultra-low-bitrate and weak-network regimes. The idea of treating historical generative context and runtime states as first-class resources is timely, and the use of a public dataset (MCL-JCV) for the main comparison is a strength. However, the current experimental evidence is too sparse and too confounded to support the claims, and a major system component (AFT) is never evaluated.","major_comments":[{"comment":"The headline claim of an 'additional bitrate reduction of approximately 25%' from GenMem is not supported by the presented comparison. GenTrans differs from GVC not only by the presence of GenMem but also by runtime state reuse, robust training, and the (unimplemented) AFT. No ablation removes GenMem, no GVC operating point at 0.006 bpp is provided, and no error bars or statistical tests accompany the LPIPS/Bpp numbers. Table 2 is a single hand-picked low-motion clip (VideoSRC30) and cannot establish a general effect. The comparison is between different operating points (0.008 vs 0.006 bpp, LPIPS 0.180 vs 0.185), so the 25% savings claim is confounded.","section":"§4.1, Table 1"},{"comment":"AFT is described as a core component ('the foundation of robust transmission'), but it is never implemented or evaluated. Section 4.3 only tests random dropping of condition tokens in the decoder; there is no network simulation, no evaluation of AFT's packetization, scheduling, congestion control, retransmission, or forward redundancy. Thus one of the three central pillars of the system is unvalidated. The abstract and introduction promise 'weak-network-aware transport', but the experiments only cover the model-side robust-training subcomponent.","section":"§3.4 vs §4.3"},{"comment":"The GenMem mechanism assumes the sender and receiver 'can operate with a shared memory context over time'. Under packet loss, the memory update m_t = Φ(m_{t−1}, v̂_t, z_t) depends on received conditions and reconstructed frames, so the receiver's memory will diverge from the sender's when tokens are dropped. The robustness experiments in Table 4 simulate loss only on condition tokens, never on the memory-update stream, and no resynchronization or consistency protocol is described. The robustness claim therefore does not cover the very mechanism that produces the bitrate savings, which is a load-bearing gap.","section":"§3.2 and §4.3"},{"comment":"The method is described only symbolically. Equations (1)–(5) introduce m_t, s_t, R, Φ, Ψ, and G without concrete definitions, architectures, loss functions, or training algorithms. The joint objective in Eq. (2) is not operationalized: the costs B_t, C_t, M_t, and U_t are never instantiated, and the weights λ_b, λ_c, λ_m, λ_u are never assigned or tuned. As a result, the experiments in Section 4 are not reproducible, and the asserted 'joint optimization' is not demonstrated to be carried out anywhere in the pipeline.","section":"§3.1–§3.5"},{"comment":"The 1.5× decoding speedup from state reuse is based on an unspecified 'representative setting' with 100 instances from an internal platform (TeleStudio2). No hardware, model, or bitrate information is given; there are no confidence intervals; and the 'Baseline (w/o reuse)' configuration is not described. The reuse rate of 37% is also unexplained. The evidence for the decoding-efficiency contribution is therefore insufficient for a quantitative claim.","section":"§4.2, Table 3"}],"minor_comments":[{"comment":"The text cites bitrate-LPIPS curves in Figure 2, but the figure is not described in detail and Table 1 contains only two operating points per codec. It is unclear how many points are plotted and whether the curves are runs or fitted trends.","section":"§4.1, Figure 2"},{"comment":"The notation B_t, C_t, M_t, and U_t is introduced without units or definitions. It is also unclear what the minimization is over: the encoder parameters, the transport policy, or the memory/state update rules.","section":"§3.1, Eq. (2)"},{"comment":"The sentence 'We compare GenTrans with HEVC and with GVC ... to isolate the contribution of GenMem' is misleading, since the comparison does not isolate GenMem. The wording should be corrected.","section":"§4.1"},{"comment":"The paper relies heavily on the authors' own prior work (AI Flow [1,8], GVC [3]) for motivation and as the main baseline. This is not inherently problematic, but the manuscripts' key axioms (the three AI Flow laws, the decomposition h_t ≈ R(m_{t−1},π_t)+Δ_t) are taken as given rather than justified. A reader outside this research line would benefit from a more neutral grounding.","section":"§2, References"},{"comment":"No sample sizes, confidence intervals, or per-sequence statistics are reported. Given the low number of operating points, the reported numbers should be treated as anecdotal.","section":"Table 1 and Table 4"}],"recommendation":"reject","confidential_remarks":"The evaluation is partly self-referential: the main baseline GVC is the authors' own prior work, and the conceptual motivation rests on their own 'AI Flow' laws. More importantly, the central claim of a 25% bitrate saving from GenMem is not supported by any controlled ablation, no error bars are given, and the AFT component—one of the three claimed contributions—is never implemented or tested. The memory-desynchronization issue under packet loss is a conceptual gap that would need a redesign or at least a resynchronization protocol. These are not minor presentation issues; they undermine the paper's central findings. The paper has interesting ideas, but as a full system paper it is far from publishable. A resubmission with proper ablations, a network-level AFT evaluation, memory-sync handling, and reproducibility details could be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Yue,\n\nQuick take: this is a system proposal, not a validated codec. The authors frame generative video transmission as joint optimization of bandwidth, computation, and memory, and that framing has some legs. The idea of reusing cross-clip generative memory and cached runtime states to cut both bitrate and decoding cost is a sensible extension of their earlier GVC work. The paper is clearly written and does not oversell the theory: the memory decomposition in Sec. 3.2 is explicitly called \"not a strict signal decomposition,\" and the equations are placeholders for a design philosophy rather than derivations. If you want to see how an active group thinks about system-level tradeoffs in this space, the first three sections are worth a skim.\n\nThe problems are in the evidence. Table 1 claims a 25% bitrate reduction from GenMem by comparing GVC at 0.008 bpp/LPIPS 0.180 to GenTrans at 0.006 bpp/LPIPS 0.185. That is not a matched comparison, and GenTrans differs from GVC in at least four ways: memory, state reuse, robust training, and the AFT transport design. There is no GenTrans-without-GenMem run and no GVC run at 0.006 bpp, so the saving could come from any of those components or from rate-control calibration. The text in Sec. 4.1 says the comparison \"isolates the contribution of GenMem,\" which is not what the table does. Table 2 is a single selected low-motion clip with a 66% saving; as a representative example it is illustrative, not evidence.\n\nThe other two result claims are thinner. AFT is described in Sec. 3.4 but never implemented or evaluated; Section 4.3 tests packet loss only on condition tokens, not on the memory-update stream, so the desynchronization risk the reader flagged is real and untested. Table 3 reports a 1.5x speedup on \"real-world decoding instances\" from a TeleStudio dataset with no description of the baseline or the workload, and no code or data is provided. For a paper whose contribution is a system, the absence of any artifact is a serious gap.\n\nBottom line: the conceptual re-packaging is not nothing, and there is a plausible research direction here. But the central efficiency claim is unsupported by the current experiments, and the transport half of the system is never built or tested. I would not cite this yet. If you are reviewing it, my recommendation would be major revision at best: implement and ablate GenMem, run AFT or drop it, add error bars and multiple clips, and release code. A serious editor could send it out because the area is active and the framing may provoke useful debate, but the paper as it stands is a proposal with preliminary numbers, not a validated result.\n\nBest.","headline":"A clean system-level framing for generative video transmission, but the headline 25% bitrate saving is an ablation-free comparison and the transport half is never built.","tokens_in":9767,"tokens_out":3020,"would_cite":false,"duration_ms":28446,"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":"This paper claims that a sender and receiver sharing a generative memory can transmit continuous video at about 25% lower bitrate than prior generative compression at equal perceptual quality, while keeping video usable through 50% packet l","keywords":["generative video compression","ultra-low-bitrate video transmission","cross-clip memory reuse","runtime state reuse","packet-loss tolerance","perceptual quality","weak-network video communication"],"falsifier":"Measure GenMem's bitrate saving on a video dataset composed of rapid scene cuts and high camera motion, where the reusable memory component should contribute little; if the saving does not shrink, the paper's explanation is wrong. Separately, apply random packet loss to the memory-update stream rather than only to condition tokens and check whether reconstruction quality or the bitrate saving collapses when sender and receiver memories diverge.","tokens_in":8884,"feed_emoji":"🎥","tokens_out":5412,"duration_ms":54836,"temperature":0.7,"pith_summary":"The paper argues that continuous video communication under ultra-low bandwidth should be treated as a joint problem of bandwidth, computation, and memory, not just as a signal-coding task. It proposes GenTrans, a generative transmission system that lets sender and receiver share a reusable visual memory across video clips, so only changed content and memory references need to be transmitted. If correct, this cross-clip memory reuse cuts bitrate by roughly 25% relative to the strongest prior generative codec at the same perceptual quality, and much more on stable, low-motion scenes. The system also reuses prior decoding states to speed reconstruction by 1.5x, and is trained to tolerate dropped condition tokens so quality degrades smoothly up to 50% packet loss. The broader point is that communication networks could coordinate computation, storage, and bandwidth instead of pushing raw pixel fidelity.","feed_headline":"Shared memory cuts video bitrate by 25%","feed_subtitle":"Generative transmission reuses visuals across clips, survives 50% packet loss, and speeds decoding by 1.5x.","key_machinery":"The load-bearing mechanism is the generative memory (GenMem): after each clip is decoded, a memory update function refreshes a shared pool of high-level visual representations that the sender and receiver both maintain. The next clip is approximated as a reusable historical component retrieved from memory plus an incremental change, so the sender transmits only a compact memory reference and the residual increment. A second mechanism, runtime state reuse, preserves the decoder's intermediate representations so that a compatible prior state initializes the next clip instead of starting full inference from scratch. Together they turn continuous video transmission into a sequence of state-conne","core_discovery":"The central claim is that generative video transmission can be reorganized as a resource-allocation problem, and GenTrans realizes this by maintaining a shared generative memory, reusing runtime decoder states, and training the decoder to work with incomplete conditions. Concretely, on the MCL-JCV 720p dataset GenTrans reaches LPIPS 0.185 at 0.006 bpp, while a prior generative codec reaches LPIPS 0.180 at 0.008 bpp—an added bitrate saving of about 25% with marginal perceptual change. In a low-motion clip, the saving grows to about 66%. Reusing compatible decoder states reduces average decoding latency from 62.4s to 41.2s, a 1.5x speedup with a 37% reuse rate. Under random packet loss on cond","pith_inferences":["A testable consequence not claimed in the paper: the 25% bitrate saving should shrink or vanish on videos with frequent hard cuts or rapid camera motion, because cross-clip memory reuse depends on stable content persisting across clips.","The packet-loss experiments drop condition tokens but never the memory-update stream; a direct stress test would drop memory-update packets and observe whether sender/receiver memory desynchronization breaks the bitrate saving.","The paper evaluates perceptual similarity only; task-oriented tests—such as running object detection or scene recognition on the reconstructed video—would directly probe whether the 'useful information' the framework claims to preserve actually survives transmission.","The 1.5x decoding speedup is measured on one sampled workload; workloads with frequent scene changes likely reduce the reuse rate, so the speedup is not a fixed property of the system."],"forward_implications":["At the same perceptual quality as a prior generative codec, GenTrans reports roughly 25% lower bitrate on a standard benchmark, with even larger savings on low-motion content.","At the same bitrate, GenTrans reports substantially lower LPIPS than conventional coding throughout the ultra-low-bitrate operating range.","Reusing compatible runtime decoding states yields a 1.5x decoding speedup and a 37% reuse rate on a sample of real decoding instances.","Training with randomly dropped condition tokens lets reconstruction quality degrade smoothly up to 50% packet loss instead of collapsing abruptly.","The explicit joint objective over transmission, computation, memory, and utility makes it possible to tune the system toward a desired operating point—e.g., more memory reuse in stable scenes, more computation in latency-sensitive settings."],"fun_headline_variants":["Generative memory reuses video bits, cuts bandwidth 25%","GenTrans: 25% less bitrate, 1.5x faster decoding","Generative codec reuses bits, cuts bitrate 25%, speeds up 1.5x","Shared memory codec: 25% bitrate cut, 1.5x faster decode","GenTrans: reuses visual memory, cuts bitrate 25%, survives packet loss"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The system assumes the sender and receiver can keep a shared memory context synchronized over time; if packet loss or drift desynchronizes that memory, the receiver would use stale references and the reported bitrate savings would vanish.","fun_headline_variants_meta":{"raw":{"variants":["Generative memory reuses video bits, cuts bandwidth 25%","GenTrans: 25% less bitrate, 1.5x faster decoding","Generative codec reuses bits, cuts bitrate 25%, speeds up 1.5x","Shared memory codec: 25% bitrate cut, 1.5x faster decode","GenTrans: reuses visual memory, cuts bitrate 25%, survives packet loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000759,"raw_usage":{"total_tokens":3226,"prompt_tokens":783,"completion_tokens":2443,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":2330}},"tokens_in":527,"tokens_out":2443,"duration_ms":16343,"temperature":1.0,"reasoning_tokens":2330,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T17:49:16.457592+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure GenMem's bitrate saving on a video dataset composed of rapid scene cuts and high camera motion, where the reusable memory component should contribute little; if the saving does not shrink, the paper's explanation is wrong. Separately, apply random packet loss to the memory-update stream rather than only to condition tokens and check whether reconstruction quality or the bitrate saving collapses when sender and receiver memories diverge.","supporting_citations":[],"review_version":1}