{"id":"bc6386f1-830b-452b-ab7b-1c9451b54d5c","arxiv_id":"2607.04872","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":11,"one_line_summary":"Event-centric tokenization plus two-step embedding matching lets a video-LLM jointly answer and timestamp RTL queries while using under 20% of LITA’s visual tokens.","lead":"EventCoT turns a video into a short list of event tokens, picks the events that answer a question, then grounds start/end times by matching special tokens to frame embeddings. It reports state-of-the-art reasoning temporal localization on ActivityNet-RTL with far fewer visual tokens than prior video-LLMs.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"SOTA claim rests on a training-inference mismatch for Step-2 inputs that is only partially mitigated by random unselected-event sampling.","rationale":"The reader's weakest assumption (fixed N=16 / T=100 DPC-KNN boundaries) is real and is already flagged in the paper's Limitations, but it is secondary: Tables 3 and 7 show that event partitions beat uniform partitions and that the detector itself has high boundary recall (94.3%). The more load-bearing issue for the SOTA claim is the train/test discrepancy in how S is constructed for Step 2. Without an ablation that isolates oracle vs. predicted S (with and without R), it is hard to know whether the reported gains come from the event-centric CoT mechanism or from a training regime that still leaks GT temporal structure. That does not overturn the contribution, but it keeps the verdict CONDITIONAL for the same practical reason the reader gave (need stronger evidence / code) while shifting the primary soft spot from boundary detection to selection fidelity. Agreement is therefore only partial: same overall risk level, different root cause.","tokens_in":28311,"tokens_out":678,"duration_ms":6131,"concrete_test":"On the ActivityNet-RTL validation set, re-run Step 2 under three controlled inputs: (i) oracle GT-overlapping events only, (ii) pure Step-1 predicted S with no R, (iii) Step-1 S + 30% R (current protocol). Report mIoU / P@0.5 / GPT-4 for each. If (ii) drops >10 points relative to the published 58.5 mIoU, or if (iii) is statistically indistinguishable from a pure-random-event baseline of matched token count, the load-bearing role of event selection is not established.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (SOTA mIoU 58.5 / GPT-4 48.7 with ~55 tokens) depends on Step 2 producing accurate ⟨start⟩/⟨end⟩ via embedding matching over V = frames from S ∪ R (Eq. 8). During training, S is the set of events that overlap the ground-truth interval (Sec. 3.3.1), so the model always sees GT-overlapping events plus 30% random unselected events. At inference, S is the noisy output of Step 1 (thresholded relevance scores). The paper acknowledges the shortcut risk of pure GT-overlap and adds R, but never measures how often Step-1 S recovers the GT-overlapping set, nor how much of the reported mIoU is carried by the random R rather than by genuine selection. Table S6 shows that when Step-1 IoU is Low (26% of samples), Step-2 mIoU is only 16.8; the headline aggregate therefore mixes a strong high-IoU regime with a weak recovery regime. If the random-R augmentation is doing most of the work on hard cases, the event-centric CoT story is weaker than claimed and the efficiency advantage is partly an artifact of training-time oracle selection.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The paper introduces EventCoT, an event-centric chain-of-thought framework for reasoning temporal localization (RTL), where a model must jointly produce a free-form answer and the supporting time interval in one response. The method (i) segments a video into N non-overlapping events via a temporal-proximity-aware, power-scaled DPC-KNN boundary detector and an event-context attention tokenizer, (ii) selects question-relevant events by embedding matching of a ⟨segment_key⟩ placeholder against event tokens (Step 1), and (iii) generates the answer while grounding ⟨start⟩/⟨end⟩ by matching placeholder embeddings to frame features within selected (plus randomly sampled unselected) events (Step 2). On ActivityNet-RTL the Vicuna-7B model reports SOTA mIoU 58.5 and GPT-4 score 48.7 with far fewer visual tokens than LITA; zero-shot ReXTime results are also strong, especially with a Qwen2.5-7B backbone. Ablations support events over uniform intervals, embedding matching over discrete time tokens, and the main loss and component choices.","tokens_in":28796,"tokens_out":1003,"duration_ms":15128,"significance":"If the results hold under fair scrutiny, the work is a clear and useful contribution to efficient video-LLM temporal reasoning. RTL is harder than post-hoc grounded VQA because localization must be embedded in the generated answer; showing that event-level intermediate reasoning plus disentangled embedding matching can beat both dense uniform tokenization (LITA) and costly agentic retrieval, while cutting visual tokens by roughly an order of magnitude, is practically and scientifically valuable. Strengths include systematic ablations (events vs uniform, matching vs discrete tokens, losses, boundary detectors, components), explicit efficiency measurements (Fig. 1, Fig. S1), zero-shot transfer to ReXTime, and an honest Limitations section on fixed N/T. The design is reproducible in outline (hyperparameters fixed across datasets; training mixture and prompts documented).","major_comments":[{"comment":"Sec. 3.3.1 and Eq. (8): Step 2 is trained with S = events that overlap the ground-truth interval (oracle selection), plus a random 30% of unselected events R, while inference uses noisy Step-1 S with the same R sampling. This train–inference gap is load-bearing for the event-centric CoT claim. Table S6 shows that when Step-1 IoU is Low (26.3% of samples), Step-2 mIoU falls to 16.8, so the headline 58.5 mIoU mixes a strong high-IoU regime with weak recovery. The main text should report (i) how often predicted S recovers the GT-overlapping set (precision/recall of event selection under the inference threshold), (ii) Step-2 performance when trained and tested with predicted S only (no GT-overlap oracle), and (iii) an ablation that removes R at inference while keeping it at training (or vice versa). Without these, it remains unclear how much of the SOTA gain is genuine selection versus train","section":null},{"comment":"Implementation Details and Limitations: N=16 events and T=100 frames are fixed for all videos, with hand-set DPC-KNN parameters (α=0.3, k=5, p=2). Table S3 shows N strongly affects localization (mIoU 39.5 at N=8 vs 53.0 at N=32), and the Limitations section correctly notes specialization to relatively short videos. The SOTA and efficiency claims on ActivityNet-RTL are therefore conditioned on a regime where a fixed coarse partition is adequate. Either restrict the claim explicitly to short/medium ActivityNet-scale videos, or provide at least one experiment with adaptive N (or longer videos / streaming) so that the event unit is not an untested free parameter of the central result.","section":null},{"comment":"Table 1 comparison protocol: several strong baselines are zero-shot foundation models or agentic systems with different visual budgets and interaction loops, while EventCoT and LITA are fine-tuned on the same RTL-oriented mixture. Fine-tuned Qwen3.5-9B / TimeLens improve localization but hurt GPT-4 score, which supports the disentangling argument, yet the paper should state more carefully which comparisons are same-data fine-tunes versus zero-shot, and report EventCoT’s token count and wall-clock under a matched frame budget (T=100) against the fine-tuned open models only, so the efficiency–accuracy Pareto claim is not inflated by heterogeneous protocols.","section":null}],"minor_comments":[],"recommendation":"minor_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean engineering paper that actually moves the RTL numbers. The new piece is not any single ingredient (event tokens, CoT selection, or placeholder matching already exist) but the two-step event-centric CoT: DPC-KNN boundary detection with temporal proximity and power scaling, event-context attention into compact tokens, Step-1 selection via ⟨segment_key⟩ matching, then Step-2 answer generation with ⟨start⟩/⟨end⟩ embedding matching over selected-plus-random frames. On ActivityNet-RTL they report mIoU 58.5 and GPT-4 48.7 with ~55 visual tokens versus LITA’s 356, and they beat fine-tuned Qwen3.5/TimeLens and agentic systems on localization while keeping answer quality. Zero-shot ReXTime is also strong, especially with the Qwen2.5 backbone. Ablations are thorough: events beat uniform partitions, matching beats discrete time tokens, and the boundary detector beats shot-change baselines on recall with almost no compute.\n\nThe stress-test concern about training-inference mismatch is real but overstated. Training feeds GT-overlapping events plus 30% random unselected events; inference uses Step-1 predictions plus the same random R. Table S6 shows the dependence: high Step-1 IoU (48% of samples) yields strong Step-2, low Step-1 IoU (26%) only recovers to 16.8 mIoU. So the headline aggregate mixes a clean coarse-to-fine regime with a weaker recovery regime. That does not invent the SOTA, but it means part of the robustness comes from R and multi-turn context rather than pure event selection. The paper is honest about the shortcut risk and about fixed N=16 / T=100 specializing the method to shorter videos; those are the real soft spots, not circularity or fake metrics.\n\nMath and losses are conventional (BCE, CE+DIoU, LM). Citations cover LITA, TRACE, Chat-UniVi, Temporal-CoT, ET-Bench, LISA fairly. No code is a practical annoyance for a methods paper this detailed.\n\nWho it is for: anyone building video-LLMs that must emit grounded answers under token budgets. Worth a serious referee. I would engage, cite the architecture and the RTL numbers, and push for adaptive N and released code.","headline":"Solid, compositional SOTA for RTL with real token savings; the train/infer gap on Step-2 is real but not fatal, and the fixed-N event model is the bigger long-term limit.","tokens_in":29375,"tokens_out":622,"would_cite":true,"duration_ms":6026,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Treating video as events and matching embeddings to placeholders lets a model answer “when” questions with timestamps and fewer tokens.","keywords":["reasoning temporal localization","event-centric tokenization","video chain-of-thought","embedding matching","ActivityNet-RTL","ReXTime","grounded video QA","visual token efficiency"],"falsifier":"On ActivityNet-RTL, replace the learned event boundaries with uniform intervals of the same count and retrain; if mIoU and GPT-4 score collapse to LITA-level or below, the event-unit claim is falsified.","tokens_in":29258,"feed_emoji":"🎬","tokens_out":587,"duration_ms":4933,"temperature":0.7,"pith_summary":"Reasoning temporal localization asks a video model to give a free-form answer that already includes the start and end times that support it, so reasoning and grounding cannot be separated. EventCoT does this by first turning the video into a small set of event tokens, selecting the events that matter for the question, and only then generating the answer while grounding start and end times by matching special placeholder tokens to frame embeddings. The paper shows this two-step chain of thought reaches the best reported scores on ActivityNet-RTL while using far fewer visual tokens than prior work, and transfers zero-shot to the grounded video QA benchmark ReXTime. A sympathetic reader cares because the same design both cuts token cost and disentangles temporal grounding from text generation, which the paper argues is why joint optimization fails in ordinary video LLMs.","feed_headline":"Event tokens cut visual cost and lift RTL scores","feed_subtitle":"Select question-relevant events, then match placeholders to frames for timestamps.","key_machinery":"Event-centric chain-of-thought: DPC-KNN boundary detection plus event-context attention produces N event tokens; Step 1 matches a ⟨segment_key⟩ embedding to those tokens; Step 2 matches ⟨start⟩/⟨end⟩ embeddings to frame features of the selected events to insert timestamps into a single generated answer.","core_discovery":"EventCoT claims that representing a video as compact event tokens, selecting question-relevant events as an intermediate reasoning step, and grounding the supporting interval by embedding matching of placeholder tokens yields state-of-the-art reasoning temporal localization with substantially fewer visual tokens than frame-uniform or agentic baselines.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Event tokens cut visual cost for SOTA reasoning temporal localization","Select question-relevant events then match placeholders to timestamps","EventCoT: fewer visual tokens lift RTL scores via event-centric CoT","Compact event tokens enable joint reasoning and temporal grounding","Event focus plus embedding match delivers efficient RTL results"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That a fixed split of every video into sixteen events from one hundred uniformly sampled frames is a good enough semantic unit for the question-relevant selection step to work across the videos the method is tested on.","fun_headline_variants_meta":{"raw":{"variants":["Event tokens cut visual cost for SOTA reasoning temporal localization","Select question-relevant events then match placeholders to timestamps","EventCoT: fewer visual tokens lift RTL scores via event-centric CoT","Compact event tokens enable joint reasoning and temporal grounding","Event focus plus embedding match delivers efficient RTL results"]},"model":"grok-4.5","effort":"low","cost_usd":0.004518,"raw_usage":{"total_tokens":1245,"prompt_tokens":690,"num_sources_used":0,"completion_tokens":83,"cost_in_usd_ticks":45180000,"prompt_tokens_details":{"text_tokens":690,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":472,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":690,"tokens_out":83,"duration_ms":4932,"temperature":1.0,"reasoning_tokens":472,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T12:09:56.967935+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On ActivityNet-RTL, replace the learned event boundaries with uniform intervals of the same count and retrain; if mIoU and GPT-4 score collapse to LITA-level or below, the event-unit claim is falsified.","supporting_citations":[],"review_version":1}