{"id":"67636cae-057b-4105-a751-2a3945674207","arxiv_id":"2509.09263","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DATE combines inference-time timestamp token injection with a caption-rewritten, temporally regularized CLIP sampling strategy to improve absolute time reasoning and event localization in long videos.","lead":"DATE improves long-video understanding by injecting explicit text timestamps into the video frame sequence and by sampling frames based on semantic relevance with temporal spacing. Tested on hour-long video benchmarks, a 7B model with DATE outperforms several 72B models on some tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"TIM ablation confounds removal of MRoPE temporal IDs with addition of timestamp tokens; observed TIM-only gain may stem from disabling absolute-time RoPE rather than from text-token time grounding.","rationale":"The reader's weakest assumption identifies exactly this confound, and I agree it is the most load-bearing concern. The alternative concerns—mixed evaluation protocols in Table 1, lack of error bars—affect the magnitude and universality of the SOTA claim but not the core mechanism. The TIM confound is more fundamental because it targets the paper's named contribution: if timestamp tokens are inert, the method reduces to TASS (a sampling strategy), and the 'absolute time enhancement' framing collapses. The paper deserves credit for re-testing Qwen2.5-VL, reporting bad cases, and ablating both components; those are good scientific practices. But the proposed four-condition test is cheap and decisive. Until it is run, the conditional verdict is appropriate; the reader's CONDITIONAL should stand.","tokens_in":15506,"tokens_out":9000,"duration_ms":88463,"concrete_test":"Run four ablations on LVBench with 256 frames and identical TASS-sampled frames using DATE-7B: (A) original MRoPE, no timestamp tokens (baseline, expected 43.7); (B) sequential temporal indexing, no timestamp tokens; (C) original MRoPE, interleaved <time_token> tokens; (D) sequential indexing, interleaved <time_token> tokens (current TIM, expected 44.9). If B ≈ D, timestamp tokens contribute nothing and TIM's gain comes from disabling absolute-time MRoPE. If C ≈ D, tokens are the active ingredient. Report per-condition accuracy and differences.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central novelty, TIM (§3.1), changes two variables simultaneously relative to the Qwen2.5-VL baseline: (i) it replaces MRoPE's absolute-time temporal position IDs with sequential indices, and (ii) it interleaves textual timestamp tokens between frame embeddings. The 'TIM only' row of Table 2 (V-MME 65.8→66.5, LongVideoBench 61.8→61.9, LVBench 43.7→44.9) therefore cannot isolate the effect of the timestamp tokens. Table 3's 'Prompt' condition—timestamps in the prompt while keeping original MRoPE—degrades LVBench to 42.5 vs. baseline 43.7, consistent with the model failing to ground textual time in video content or with prompt-based timestamps being an ineffective injection format. The paper asserts 'absolute time perception is handled independently via the explicit <time_token>s' but provides no condition that tests timestamp tokens while leaving the positional encoding unchanged. If the true cause of TIM's gain is removing the absolute-time MRoPE component, the claim that DATE improves absolute time understanding via token injection is unsupported, and the central 'Dynamic Absolute Time Enhancement' mechanism is misattributed. TASS alone accounts for the larger LVBench gain (46.7 vs 43.7), making TIM's mechanism the load-bearing part of the paper's contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DATE, a training-free inference-time enhancement for long-video multimodal LLMs. It consists of two components: TIM, which interleaves textual timestamp tokens between visual frame embeddings and replaces Qwen2.5-VL's absolute-time MRoPE temporal indices with sequential indexing; and TASS, which converts user questions into descriptive captions via DeepSeek-V3, scores 1-FPS frames with CLIP, filters negative frames, and greedily selects temporally spaced top candidates. Experiments on Video-MME, LongVideoBench, and LVBench report consistent gains over a re-tested Qwen2.5-VL-7B baseline, with the 7B model approaching or surpassing some larger models on certain benchmarks.","tokens_in":15837,"tokens_out":4842,"duration_ms":51011,"significance":"If the results hold, DATE is a practically attractive recipe: it improves temporal grounding and event localization in existing MLLMs without weight updates, and TASS's query-to-caption reformulation plus temporal regularization appears to be a genuine improvement over uniform sampling and over the recent AKS baseline. The authors provide public code, re-test all baselines under a common protocol, and include extensive qualitative examples. The main weakness is that the paper's central mechanistic claim about TIM is not supported by the current ablations, because the TIM-only condition changes two variables simultaneously. The TASS contribution is better supported and accounts for the larger share of the reported gains.","major_comments":[{"comment":"The TIM ablation is confounded. The 'TIM only' row changes two things relative to the Qwen2.5-VL baseline: (i) it replaces MRoPE's absolute-time temporal position IDs with sequential indices, and (ii) it injects textual timestamp tokens. The 'Prompt' condition in Table 3 adds timestamps to the prompt while keeping MRoPE, and it degrades performance (LVBench 42.5 vs 43.7). No condition tests timestamp tokens while leaving MRoPE unchanged, nor sequential indices without timestamp tokens. Therefore the observed TIM-only gain (e.g., LVBench 43.7→44.9) could be caused by removing the absolute-time MRoPE component rather than by the model grounding absolute time through textual tokens. The assertion in §3.1 that 'absolute time perception is handled independently via the explicit <time_token>s' needs a control experiment that isolates the token injection from the positional-encoding change.","section":"§3.1, Table 2, Table 3"},{"comment":"Hyperparameters α=4 and δ0=20 are selected using the same LVBench benchmark that produces the headline numbers, and no held-out split is used. The reported differences are often very small (e.g., LongVideoBench 61.8→61.9 for TIM-only), and no error bars or multiple trials are reported. It is therefore unclear whether the TIM-only gains are statistically meaningful. Please report means and variances over multiple runs, or fix hyperparameters on a separate validation split and then evaluate once on the test benchmarks.","section":"§4.1, §4.6, Fig. 7, Table 2"},{"comment":"The state-of-the-art comparison mixes official reported numbers (♠) with re-tested numbers (♣) under different preprocessing protocols. The table footnote states that Qwen2.5-VL's official LongVideoBench result was obtained at 224 resolution, while the re-tested baseline is at 448. Claims such as 'surpassing many 72B models' and 'outperforming the prior best model' are therefore protocol-dependent. To support the SOTA claim, report all baselines and the proposed method under exactly the same resolution, FPS, frame count, and benchmark version, and clearly separate re-tested from official numbers.","section":"Table 1 and footnotes"},{"comment":"The reported sampling times for TASS appear to omit the two most expensive stages: DeepSeek-V3 caption generation and CLIP similarity scoring over all 1-FPS frames. If these stages are excluded, the comparison with AKS under 'the same CPU' is not an end-to-end latency comparison, and the efficiency claim is misleading. If they are included, the wall-clock time would be much larger than the reported 1.2–21.2 s. Please either report end-to-end latency including caption generation and feature extraction, or explicitly state that only the greedy selection step is timed.","section":"Table 3, §4.1"}],"minor_comments":[{"comment":"Typo: 'achieves superior spatial perception' should be 'achieves superior spatial perception' (and probably 'temporal perception' is intended).","section":"§1, contribution (3)"},{"comment":"The paper says 'Our diagnostic experiments further confirm that such models do not solve problems related to absolute time reliably,' but no diagnostic experiment is presented. Either add this evidence or soften the claim.","section":"§1 and §4.1"},{"comment":"The attention-map analysis is based on a single question and a single video. This is anecdotal; quantify with several examples or aggregate statistics before claiming timestamp tokens act as temporal anchors.","section":"§4.5, Fig. 6"},{"comment":"Typo in caption: 'betweet' should be 'between'.","section":"Fig. 3"},{"comment":"The asset list duplicates 'LongVideoBench (CC-BY-NC-SA 4.0 license)' twice.","section":"Appendix F"},{"comment":"The limitation about linear inference growth is appreciated. It would be useful to also mention the dependence on an external caption-generation LLM (DeepSeek-V3) in the limitations discussion.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The TASS contribution is solid and could stand on its own, but the paper's title and central contribution hinge on TIM, whose mechanism is not isolated by the current ablations. The missing control condition (sequential MRoPE without timestamps, or timestamps with unchanged MRoPE) is a standard and feasible experiment; I am not asking for a different scope. I also recommend the editors insist that the SOTA table use fully consistent evaluation protocols, since the current mix of resolutions and re-testing conditions weakens the headline claims. No concerns about novelty disclosure or asset compliance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things before reading this one. First, the training-free recipe is real and the gains are consistent – DATE lifts Qwen2.5-VL-7B by +1.5 to +3.7 points on Video-MME, LongVideoBench, and LVBench, and the LVBench improvement is the kind of jump that matters for hour-long video QA. Second, the paper's central mechanism, Timestamp Injection (TIM), is not actually isolated by the ablation. The stress-test note is right: the TIM-only row in Table 2 changes two variables at once – it replaces MRoPE's absolute-time position IDs with sequential indexing and interleaves textual timestamp tokens. So the +1.1 LVBench gain attributed to TIM could be from disabling the absolute-time RoPE rather than from the model grounding 'at 45s' through text alone. Table 3 makes this worse: putting timestamps in the prompt while keeping MRoPE intact degrades LVBench to 42.5 vs 43.7 baseline, which is consistent with the model not truly grounding textual time in video content. The paper's claim that 'absolute time perception is handled independently via the explicit <time_token>s' needs a cleaner experiment to support it. What is genuinely new: rewriting the user's question into a CLIP-friendly declarative caption before similarity scoring, and the temporally-regularized greedy sampler that enforces a minimum interval between selected frames. These are solid improvements over AKS and BOLT, and the whole thing runs inference-only on existing checkpoints, which is practically useful. TASS alone accounts for the larger LVBench gain (46.7 vs 43.7), so even if TIM's contribution is murky, the sampling side seems to carry weight. The other soft spots are moderate but fixable: hyperparameters alpha and delta_0 are tuned on the same benchmarks that produce the headline numbers, there are no error bars, and the SOTA table mixes evaluation protocols (they re-test Qwen2.5-VL at 448 resolution while the official LongVideoBench number is at 224). The paper is honest about its bad cases and includes a limitations section, which earns some trust. Bottom line: this deserves a serious referee, not a desk reject. It's a useful training-free enhancement with a confounded mechanism check. I'd send it to review and ask for an ablation that keeps MRoPE intact while adding timestamp tokens, plus variance estimates. I'd bring it to reading group as a good case study in ablation design.","headline":"A plausible training-free recipe for long-video temporal reasoning with consistent gains, but the central timestamp-injection mechanism is confounded in the ablation: the TIM row changes both the positional encoding and the added tokens, so the paper's headline mechanism is not actually isolated.","tokens_in":738,"tokens_out":1643,"would_cite":true,"duration_ms":31645,"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":"Injecting textual timestamps between frames and selecting frames with a caption-guided, temporally regularized sampler gives pretrained video LLMs state-of-the-art absolute-time localization without retraining.","keywords":["long video understanding","absolute time localization","timestamp injection","frame sampling","multimodal large language models","temporal reasoning","event localization","CLIP similarity"],"falsifier":"Run a three-way comparison on identical frames: (a) original model with absolute-time position indices, (b) original model plus timestamp tokens but keeping those indices, and (c) DATE's sequential indices plus timestamps. If (b) is as good as (c), the sequential-index claim is unnecessary; if (c) is not better than (a), timestamp injection per se is not the source of the gain.","tokens_in":15400,"feed_emoji":"⏱️","tokens_out":8213,"duration_ms":75985,"temperature":0.7,"pith_summary":"The paper tries to establish that a multimodal large language model can understand absolute time in hour-long videos without retraining, by making time a visible part of the input rather than a hidden property of position numbers. Its core claim is that interleaving each sampled frame with a textual timestamp token, while replacing the model's time-scaled position indices with simple sequential ones, gives the model a stable temporal coordinate system. It also claims that frame selection should be treated as retrieval: turning a user's question into a descriptive caption and greedily picking high-similarity frames under a minimum-time-gap constraint captures key events without destroying continuity. On three hour-long benchmarks the method moves a 7B baseline to state-of-the-art among small models and above several 72B models, with the 72B version setting the overall state of the art. A sympathetic reader would care because this is an inference-time recipe: better absolute temporal localization without weight updates or new training data.","feed_headline":"Text timestamps help a 7B video model beat 72B rivals","feed_subtitle":"Injecting clock labels between frames and sampling by caption similarity lifts absolute-time localization on three hour-long benchmarks.","key_machinery":"Two mechanisms carry the argument. Timestamp Injection Mechanism (TIM): after each frame's visual tokens, the input inserts a textual time token (e.g., \"at 45s\"), and the temporal branch of the multimodal rotary position encoding is switched from absolute-time-scaled indices to simple sequential indices; this creates a continuous temporal reference system while preventing position-index drift. Temporal-Aware Similarity Sampling (TASS): a language model rewrites the user question into a concise declarative caption, CLIP scores every frame against that caption, below-mean frames are discarded, and a greedy algorithm picks the highest-scoring remaining frames subject to a minimum interval that","core_discovery":"The central discovery is that absolute-time reasoning in a pretrained video multimodal LLM can be decoupled from positional geometry. The paper shows that the baseline model's multimodal rotary position encoding, which embeds real seconds into position indices, degrades on long videos; when those indices are replaced with sequential indices and explicit textual timestamps are interleaved with frame embeddings, the model answers \"at what time did X happen\" questions more accurately. The authors further show that the frames fed to the model matter as much as the time signaling: by rewriting questions as short declarative captions and using CLIP similarity with a temporally regularized greedy s","pith_inferences":["Inference: If the mechanism holds, \"reading the clock\" can be moved out of position geometry entirely, which suggests a transferable design for any long-sequence model: attach explicit coordinate tokens and use plain ordinal positions to avoid drift.","Inference: The caption-rewriting trick is generalizable: any embedding trained on declarative captions will prefer declarative queries; a natural test is to apply it to other retrieval-based video or audio pipelines.","Inference: A clean experiment the paper does not run is to rescale the timestamp units (seconds vs. HH:MM:SS) and check whether answers track the new clock; that would directly test whether the model grounds language-time rather than exploiting token order.","Inference: The paper's reported bad cases suggest the added tokens consume context budget and push the model toward its capacity limit; a possible extension is to compress timestamps or make injection conditional on query difficulty."],"forward_implications":["Absolute-time localization can be improved at inference time, with no weight updates and no new training data.","A 7B model using DATE can outperform several 72B baselines on hour-long benchmarks, indicating that sampling and temporal-reference design can matter as much as model scale.","Framing frame selection as image-caption retrieval, with a rewritten declarative query, yields higher recall of key events than uniform sampling or question-based similarity.","Explicit timestamp tokens act as attention anchors, concentrating the model's attention on the visual content of the corresponding frame.","Temporally regularized greedy sampling provides better accuracy than variance-based keyframe selection, especially at very low frame budgets."],"fun_headline_variants":["Timestamps and smart sampling boost long-video AI","Injecting clock labels improves video LLM time reasoning","7B video model outdoes 72B with timestamp injection","How absolute time labels sharpen long-video understanding","Temporal cues in text beat positional encoding for video"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that a pretrained video-language model can genuinely read absolute time from plain-text timestamp tokens interleaved between frames once its time-scaled position indices are replaced with sequential ones; the paper's own ablation changes both parts at once, so this premise is not yet isolated.","fun_headline_variants_meta":{"raw":{"variants":["Timestamps and smart sampling boost long-video AI","Injecting clock labels improves video LLM time reasoning","7B video model outdoes 72B with timestamp injection","How absolute time labels sharpen long-video understanding","Temporal cues in text beat positional encoding for video"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000655,"raw_usage":{"total_tokens":2832,"prompt_tokens":736,"completion_tokens":2096,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":480,"completion_tokens_details":{"reasoning_tokens":2033}},"tokens_in":480,"tokens_out":2096,"duration_ms":14160,"temperature":1.0,"reasoning_tokens":2033,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T19:23:10.816220+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a three-way comparison on identical frames: (a) original model with absolute-time position indices, (b) original model plus timestamp tokens but keeping those indices, and (c) DATE's sequential indices plus timestamps. If (b) is as good as (c), the sequential-index claim is unnecessary; if (c) is not better than (a), timestamp injection per se is not the source of the gain.","supporting_citations":[],"review_version":1}