{"id":"adb900f8-03af-45b7-95b2-b4ae657dd948","arxiv_id":"2608.01169","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"NovaCov uses a bounded, recency-weighted historical reference bank and a dual-branch submodular coverage objective to select streaming video tokens, outperforming training-free baselines.","lead":"This paper introduces NovaCov, a method that compresses video tokens for streaming video LLMs by choosing sets of tokens based on what they add to content already shown. It keeps 99.6% of the uncompressed model's accuracy while cutting prefill latency by about 46%.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Historical Reference Bank fidelity is the load-bearing risk: no sensitivity analysis or exact-history oracle supports the fixed C=512, θ=0.9, α=0.10 defaults.","rationale":"The reader's weakest_assumption matches the risk I would flag: the Historical Reference Bank is the novel component that makes set-wise streaming compression possible, and its fidelity is only empirically asserted. I checked Proposition 1 and the greedy selection implementation; the submodularity proof is sound, and the dual-branch objective is correctly implemented as two facility-location functions. The theoretical guarantee is therefore not the weak point. The empirical comparison is extensive but lacks error bars and hyperparameter robustness, which matters because the method is plug-and-play and 'training-free' — users would rely on the fixed defaults. The unbounded-bank ablation does not resolve the question, since it tests capacity rather than the approximation error of the bank as a summary of the true history. For these reasons, the concern is real but not disqualifying: the paper should be CONDITIONAL pending a sensitivity/fidelity check, which is exactly the reader's verdict. No adjustment to the verdict is needed.","tokens_in":13443,"tokens_out":14543,"duration_ms":140498,"concrete_test":"Run a sensitivity sweep on OVO-Bench (FPD/EPM/REC) and StreamingBench at 25% retention: C ∈ {128,256,512,1024}, θ ∈ {0.8,0.85,0.9,0.95}, α ∈ {0.05,0.1,0.2,0.4}, keeping other hyperparameters at the paper's defaults. Report per-setting accuracy and compare with STC-Pruner (50.6 OVO, 63.7 StreamingBench). In addition, on a 20-video subset, compute an exact-history oracle: replace Eq. (6)'s bank-based h_i with h_i^exact = max over all previously forwarded tokens, rerun the same greedy selection, and compare accuracy. If the oracle is >2 points higher, the bank is a poor proxy; if NovaCov falls below the best baseline at any plausible setting, the central claim lacks robustness.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism for cross-frame redundancy is the Historical Reference Bank's h_i (Eq. 6), which directly controls both the historical-novelty branch and the novelty weights ρ_i in Eq. (8). If h_i is miscalibrated, the dual-branch objective either re-selects content already forwarded to the LLM or suppresses genuinely new tokens, and the claimed 4.7-point StreamingBench gain over STC-Pruner could erode. The paper's only fidelity evidence is the unbounded-bank ablation (Table 4d), but that still uses EMA/threshold prototypes and does not test the bank against the exact history R*_t in Eq. (3). With seven hand-set hyperparameters (C=512, λ=16, θ=0.9, α=0.10, λ_c=0.1, λ_r=0.9, ρ_min=0.20) and no sensitivity analysis or error bars, the SOTA claim rests on a narrow point in hyperparameter space.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NovaCov, a training-free, plug-and-play token compressor for streaming video. It maintains a capacity-bounded, recency-weighted Historical Reference Bank of prototypes, updated by a match-or-insert EMA rule and recency-aware eviction. At each frame it greedily maximizes a dual-branch submodular coverage objective: a novelty-weighted current-frame facility-location branch plus a historical-novelty residual branch that credits only coverage beyond what the bank already supplies. The paper proves that both branches are facility-location functions, so greedy selection retains the (1-1/e) approximation guarantee. Evaluations on OVO-Bench, StreamingBench, EgoSchema, MLVU-dev, and VideoMME report that NovaCov retains 99.0-99.6% of uncompressed ReKV accuracy while reducing LLM prefilling latency by ~46%, outperforming existing training-free compressors.","tokens_in":13698,"tokens_out":5447,"duration_ms":57512,"significance":"If the empirical claims hold, NovaCov is a useful contribution: it is, to my knowledge, the first set-wise streaming token compressor, and it cleanly reduces the problem to a weighted facility-location objective with a classical approximation guarantee. The proof in Proposition 1 is correct and the benchmark coverage is broad, spanning streaming and offline long-video understanding. The main reservation is that the central mechanism—the Historical Reference Bank—is validated only indirectly, and all headline numbers are single point estimates without sensitivity analysis or error bars. Since the residual branch and the novelty weights both depend on the bank's coverage scores h_i, the strength of the empirical SOTA claim is not yet established.","major_comments":[{"comment":"The Bank's fidelity is load-bearing: h_i in Eq. (6) directly controls the residual branch in Eq. (8) and the novelty weights in Eq. (7). The only ablation aimed at the bank, Table 4d, is labeled 'Unbounded', but it still uses EMA/threshold prototypes and simply never evicts; it is not the exact history R*_t of Eq. (3). Thus the paper never tests whether the bounded bank faithfully represents the content already forwarded to the LLM. With defaults C=512, λ=16, θ=0.9, α=0.10, λc=0.1, λr=0.9, ρmin=0.20 and no sensitivity analysis, the claimed 4.7-point StreamingBench gain over STC-Pruner rests on a single hand-set point in a seven-dimensional space. Please add sensitivity curves for the bank hyperparameters, and an oracle variant that computes h_i against the true history (for short streams) to validate the bank design.","section":"Historical Reference Bank, Eq. (6), Table 4d"},{"comment":"All tables report single point estimates. The headline improvements over STC-Pruner are 4.7 on StreamingBench, 2.1 on OVO-Bench, and 0.5 on the offline average; these are modest enough that run-to-run variance could change the ranking on some benchmarks. Please report means and standard deviations over multiple runs or seeds, or otherwise provide statistical evidence that the observed differences are not noise. Without this, the SOTA claim is under-supported.","section":"Experiments, Tables 1–3"},{"comment":"The text states that the unbounded variant's cost 'rises without bound', but no latency or memory measurements are reported for variant (d). More importantly, because (d) still uses the same EMA/threshold update, it does not isolate the effect of boundedness on bank fidelity. The accuracy drop from 68.4 to 67.4 is attributed to saturation, but an alternative explanation is that unbounded growth changes the h_i distribution without approaching the true history. Please directly compare against the exact-history oracle and measure cost for the unbounded variant.","section":"Table 4d and 'Bounded vs. unbounded'"}],"minor_comments":[{"comment":"The claim that NovaCov 'outperforms every compression method on all five benchmarks' is too strong: on VideoMME Short, NovaCov ties with VidCom2 and STC-Pruner at 68.7, and several task-level numbers in Tables 1–2 are below a baseline. Please qualify the claim to aggregate or average performance.","section":"Main Comparisons, Table 3"},{"comment":"The rank normalization RankNorm(1-h_i) is not specified in detail. Please define how ties are handled and how ranks are mapped to [0,1].","section":"Eq. (7)"},{"comment":"No code release or reproducibility statement is mentioned. Given the number of hand-set hyperparameters, releasing code or at least detailed per-task settings would substantially improve reproducibility.","section":"Implementation details"},{"comment":"The y-axis label 'Accuracy Performance (%)' is redundant; 'Accuracy (%)' would be clearer. The legend is also hard to read at small sizes.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The theoretical part is sound and the idea is timely. The main risk is that the empirical SOTA claim may be tied to a narrow hyperparameter configuration. I would be willing to accept after the authors add sensitivity analysis, an exact-history oracle comparison, and error bars. No citation or novelty-disclosure concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a serious referee. The core idea is genuinely new: framing streaming token compression as set-wise submodular coverage where the reference is a bounded, recency-weighted bank of prototypes. The dual-branch objective is a clean facility-location combination, and the proof that it stays monotone submodular is correct—the residual branch really is just a shifted similarity matrix. That's not a perfunctory theoretical section; it does real work.\n\nThe empirical story is consistent. The gains over STC-Pruner (4.7 on StreamingBench, 2.1 on OVO-Bench) are meaningful, and the unbounded-bank ablation is a nice touch: it shows forgetting is not a bug but a feature, and it partially answers the worry that the bank is just a lossy crutch. Retaining 99.6% of ReKV accuracy at 25% tokens is a practical win for deployment.\n\nThe soft spots are reproducibility, not logic. No code, no error bars, seven hand-set hyperparameters (C, λ, θ, α, λ_c, λ_r, ρ_min) with no sensitivity analysis. That's a real gap for a paper that claims SOTA—you can't tell if the numbers sit on a narrow ridge or a plateau. The historical-bank fidelity concern is legitimate but not fatal: the paper's own ablations show the bounded bank outperforms the unbounded one, which suggests the mechanism is not just memorizing. Still, an oracle comparison against exact history R*_t would have settled it.\n\nWho's this for? Anyone building streaming VideoLLM inference. The math is accessible, the method is plug-and-play. I would send it to review as-is—it deserves referee time, and the main revision would be code release plus sensitivity analysis, not a change of argument.","headline":"A genuinely new formulation—streaming token compression as set selection with a bounded historical reference—with a clean submodularity proof and solid gains, but reproducibility gaps (no code, no error bars, seven hand-set hyperparameters) keep it at conditional accept.","tokens_in":14138,"tokens_out":1504,"would_cite":true,"duration_ms":14507,"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":"NovaCov reframes streaming video token compression as set selection: a token is worth keeping only for what it adds beyond the content the model already holds.","keywords":["streaming video understanding","token compression","set selection","submodular maximization","facility-location coverage","training-free compression","historical reference bank","greedy approximation"],"falsifier":"Construct a stream in which an object appears, disappears for longer than the bank's half-life (16 frames), then reappears; if eviction removes its prototype, the reappearing tokens are treated as novel rather than already-conveyed, so a repetition-counting task should lose accuracy relative to a bank with a much longer half-life or larger capacity. This directly tests whether the recency-weighted bank faithfully represents the historical reference.","tokens_in":13346,"feed_emoji":"🎬","tokens_out":8587,"duration_ms":70930,"temperature":0.7,"pith_summary":"Streaming video models must drop most visual tokens of every arriving frame to keep latency and memory bounded, but existing training-free compressors score each token in isolation and can waste the budget on near-duplicate content. The paper argues that compression should be a set-selection problem: a token is worth keeping only for what it adds beyond the tokens already retained. To do this in a causally constrained stream, NovaCov maintains a fixed-capacity, recency-weighted bank of historical prototypes and greedily selects tokens under a dual facility-location objective that covers the current frame while favoring content the history has not yet covered. On streaming benchmarks it keeps 99.6% of the uncompressed ReKV accuracy while cutting LLM prefilling latency by 46%, and it also improves over earlier training-free methods on offline long-video benchmarks. The payoff is a plug-in selection rule that makes a streaming VideoLLM nearly as accurate as an uncompressed one at a fraction of the compute.","feed_headline":"Set-wise token pruning keeps 99.6% of streaming video accuracy","feed_subtitle":"NovaCov values each token by what it adds over a compact history, cutting LLM prefilling time by 46%.","key_machinery":"The Historical Reference Bank: a capacity-bounded set of prototypes, each an EMA-refined vector with a last-update frame and match count, updated by match-or-insert and pruned by recency-aware eviction with utility $u_j = 2^{-(t-\\tau_j)/\\lambda}\\log(1+n_j)$. It supplies the historical term of the reference set. The dual-branch objective $F_t(S) = \\lambda_c \\sum_i \\rho_i c_i(S) + \\lambda_r \\sum_i \\max(c_i(S)-h_i,0)$ is a weighted sum of two facility-location functions, one over the current frame and one over the residual coverage above history, preserving monotone submodularity.","core_discovery":"NovaCov's central claim is that streaming video token compression is a subset-selection problem whose reference set must include the content already forwarded to the model, and that a bounded, recency-weighted Historical Reference Bank is enough to serve as that reference. Each incoming token is scored by its historical coverage $h_i$, the similarity to the most similar bank prototype, decayed by recency. The selection objective combines a current-frame facility-location branch weighted by novelty with a residual branch that credits a token only for coverage above $h_i$; both branches are monotone submodular, so greedy selection keeps the $(1-1/e)$ approximation guarantee. The paper reports","pith_inferences":["The reference-set principle suggests that other streaming components, such as KV-cache eviction or retrieval indices, could benefit from the same 'what has the model already seen' criterion, potentially unifying token and cache compression.","Because the bank's recency decay explicitly trades memory for recency, NovaCov's design quantifies a general forgetting strategy; one could tune the half-life to match a domain's typical reappearance timescales, which the current fixed $\\lambda=16$ leaves unexamined.","The residual branch formulation—shifting the similarity by $h_i$—may extend to non-facility-location objectives (e.g., coverage with determinantal point processes) while preserving a similar approximation analysis.","A testable extension is to apply the same historical-reference idea to audio or LiDAR tokens in streaming multimodal models, where the same cross-frame redundancy pattern should appear."],"forward_implications":["At 25% retention, NovaCov exceeds the uncompressed ReKV accuracy on five of ten StreamingBench tasks, showing that removing redundant tokens can reduce noise and improve cross-frame reasoning.","The method is training-free and plug-and-play, so it can be dropped into any streaming VideoLLM pipeline that already uses per-frame KV-cache retrieval, without retraining.","The bounded bank keeps per-frame cost independent of stream length, so the latency advantage grows with stream duration.","The dual-branch objective and greedy selection carry a $(1-1/e)$ worst-case guarantee, which token-wise ranking cannot offer.","The same compressor transfers to offline long-video understanding, improving all three offline benchmarks over previous training-free methods."],"supporting_citations":[{"why":"Supplies the facility-location coverage objective used for both branches of the selection function.","marker":"(Cornuéjols, Fisher, and Nemhauser 1977)"},{"why":"Places facility-location functions in the submodular summarization class that NovaCov builds on.","marker":"(Lin and Bilmes 2011)"},{"why":"Establishes that weighted facility-location functions are monotone submodular, justifying the greedy selection.","marker":"(Krause and Golovin 2014)"},{"why":"Proves the (1-1/e) approximation guarantee for greedy maximization of monotone submodular functions.","marker":"(Nemhauser, Wolsey, and Fisher 1978)"},{"why":"Shows set-wise facility-location token compression for offline video, the approach NovaCov adapts to streaming.","marker":"(Cho et al. 2026)"},{"why":"Provides the ReKV streaming backbone and the uncompressed accuracy upper bound that NovaCov retains 99.6% of.","marker":"(Di et al. 2025)"},{"why":"STC-Pruner is the strongest token-wise streaming baseline NovaCov outperforms on both streaming benchmarks.","marker":"(Wang et al. 2026b)"},{"why":"OVO-Bench is the streaming benchmark where the headline 99.6% retention and 46% latency reduction are measured.","marker":"(Niu et al. 2025)"},{"why":"StreamingBench is the second main streaming benchmark used for the comprehensive comparison.","marker":"(Lin et al. 2026)"},{"why":"The Triton kernel fuses both branches so that the set-wise objective runs at latency comparable to token-wise baselines.","marker":"(Tillet, Kung, and Cox 2019)"}],"fun_headline_variants":["Set-wise video token compression: 99.6% ReKV accuracy, 46% less prefill","Streaming video token pruning: 99.6% acc, 46% faster prefill","Rethink video tokens: set-wise compression cuts latency 46%","NovaCov: set-wise token compression keeps 99.6% ReKV accuracy","Video token compression as set selection: 46% less prefill, same ReKV accuracy"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The Historical Reference Bank, with its fixed capacity, match threshold, EMA rate, and decay half-life, faithfully summarizes everything the model has already been shown, so the historical coverage $h_i$ correctly identifies redundant tokens; the paper checks this only empirically and gives no sensitivity analysis over these hyperparameters.","fun_headline_variants_meta":{"raw":{"variants":["Set-wise video token compression: 99.6% ReKV accuracy, 46% less prefill","Streaming video token pruning: 99.6% acc, 46% faster prefill","Rethink video tokens: set-wise compression cuts latency 46%","NovaCov: set-wise token compression keeps 99.6% ReKV accuracy","Video token compression as set selection: 46% less prefill, same ReKV accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001301,"raw_usage":{"total_tokens":5147,"prompt_tokens":751,"completion_tokens":4396,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":495,"completion_tokens_details":{"reasoning_tokens":4280}},"tokens_in":495,"tokens_out":4396,"duration_ms":30885,"temperature":1.0,"reasoning_tokens":4280,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T00:25:55.696968+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a stream in which an object appears, disappears for longer than the bank's half-life (16 frames), then reappears; if eviction removes its prototype, the reappearing tokens are treated as novel rather than already-conveyed, so a repetition-counting task should lose accuracy relative to a bank with a much longer half-life or larger capacity. This directly tests whether the recency-weighted bank faithfully represents the historical reference.","supporting_citations":[],"review_version":1}