{"id":"5936d70f-bd2e-4e8b-9570-60cf058c6347","arxiv_id":"2608.02078","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"CAVE adds boundary evidence tokens and an attention alignment reward to RL-based video temporal grounding, improving average timestamp accuracy by about 1.9 points over TimeLens-7B.","lead":"This paper introduces CAVE, a reinforcement-learning training method for video temporal grounding that adds special 'start' and 'end' evidence tokens and rewards the model when their visual attention lands on the true boundary frames. It reports consistent accuracy gains over a strong baseline, especially on strict temporal overlap thresholds, on three public video grounding benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"VBEAR's indirect credit assignment to timestamp tokens is unverified; gains could stem from warm-up or format regularization rather than evidence alignment.","rationale":"The reader's weakest assumption is that the policy-gradient signal from VBEAR transfers to the earlier timestamp tokens. This is indeed the most load-bearing premise: the method's stated mechanism for reducing evidence–timestamp misalignment depends on this indirect credit assignment. The paper's ablations show that adding VBEAR helps, and the analysis using timestamp-token attention suggests some transfer occurs, but neither closes the gap that the gains might come from warm-up or format regularization. The concrete tests proposed—gradient analysis or a sham-VBEAR control—would directly settle whether VBEAR's effect on timestamps is causal and specific. Given the absence of these controls and the small metric differences, the existing CONDITIONAL verdict remains appropriate; no change is needed.","tokens_in":13202,"tokens_out":7027,"duration_ms":55028,"concrete_test":"Compute the expected gradient of the VBEAR reward with respect to the log-probabilities of the numeric timestamp tokens on a held-out set (e.g., via a REINFORCE estimator). If the gradient magnitude is negligible compared to that from the IoU reward, the indirect signal is too weak to drive the observed gains. Alternatively, run a sham-VBEAR control where the reward is computed on attention from a random output token instead of the evidence tokens; if the improvement over the warm-up+GRPO baseline persists, the effect is not specific to evidence alignment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that VBEAR improves temporal grounding by aligning boundary-specific visual evidence with timestamps. However, the reward is computed solely on the attention of <Start>/<End> evidence tokens, which are generated after the numeric timestamps. The policy-gradient path from VBEAR to the earlier timestamp tokens is indirect: the reward depends on timestamps only through the hidden state that conditions evidence-token generation. If this signal is weak, the observed improvements from adding VBEAR (Table 2, rows 1 vs 2) could instead arise from the supervised warm-up (which already teaches evidence-token attention to ground-truth boundaries) or from the output-format regularization. Moreover, the paper evaluates evidence–timestamp alignment using timestamp-token attention (Figure 4), not the evidence-token attention that VBEAR directly optimizes, leaving open the possibility that VBEAR only shapes the added tokens while timestamps improve for unrelated reasons. The risk of reward hacking is concrete: since the evidence tokens can attend to ground-truth boundaries regardless of the predicted interval, VBEAR may saturate to a near-constant value, providing no useful gradient for timestamp refinement.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CAVE, a method for video temporal grounding with LVLMs. It introduces two special boundary-evidence tokens, <Start> and <End>, generated after the numerical timestamp tokens. A supervised warm-up stage trains these tokens to attend near the ground-truth boundaries, and a subsequent GRPO stage adds a Visual Boundary Evidence Alignment Reward (VBEAR) computed from the evidence tokens' attention ratios. A performance-aware gate (PAGE) scales this reward according to the group mean IoU of the rollouts. Experiments on Charades/ActivityNet/QVHighlights TimeLens benchmarks are reported for 7B and 8B backbones, with ablations, a matched-IoU evidence-consistency analysis, and a case study. The central claim is that CAVE reduces evidence--timestamp misalignment and improves temporal grounding, with an average gain of about 1.9 points over TimeLens-7B across twelve metrics.","tokens_in":13562,"tokens_out":5188,"duration_ms":41886,"significance":"The paper addresses a genuine and under-explored issue: even when a VTG model's internal visual attention peaks near the ground-truth boundaries, the numeric timestamp output can still be off. Introducing dedicated evidence tokens and rewarding their boundary-focused attention is a reasonable mechanism to inject boundary-level supervision into RL. The matched-IoU analysis in Figure 4 is a good attempt to separate evidence alignment from localization accuracy, and the ablation study isolates VBEAR and PAGE. If the mechanism is confirmed, the contribution would be useful to the RL-based VTG community. The evaluation uses public benchmarks and compares directly with the TimeLens baseline under the same backbone and data, which is a strength. However, the current evidence has several load-bearing gaps, detailed below, that prevent the central claim from being fully established.","major_comments":[{"comment":"VBEAR is computed from the attention of the <Start>/<End> evidence tokens, but the evidence--timestamp consistency analysis that supports the central claim uses attention from the numeric timestamp tokens (Figure 4 caption: 'using generated timestamp-token attention'). These are different quantities. Because the warm-up stage directly supervises evidence-token attention toward ground-truth boundaries, the alignment of those tokens is partly true by construction. The more meaningful claim is that numeric timestamp-token alignment improves, yet that quantity is not optimized by VBEAR. Please report the same analysis for evidence-token attention and, ideally, a variant in which the reward is computed from timestamp-token attention, to verify the transfer.","section":"Method (Eq. 5-8) and Analyses (Figure 4)"},{"comment":"The reward r_b depends only on attention of the evidence tokens, not on the predicted interval. Since the evidence tokens are generated after the timestamps (the paper states 'the evidence tokens are conditioned on the completed temporal decision'), VBEAR can be maximized even for rollouts with inaccurate timestamps. The warm-up already teaches the evidence tokens to attend near ground-truth boundaries, so VBEAR may saturate to a near-constant value for many rollouts and provide little gradient for timestamp refinement. The improvement of row (2) over row (1) in Table 2 could then stem from output-format regularization or exploration rather than evidence alignment. Please demonstrate that VBEAR rewards are not saturated and are informative for localization quality, e.g., by plotting reward versus IoU over rollouts or reporting the reward distribution by rollout group.","section":"Method, VBEAR (Eq. 6-7)"},{"comment":"No error bars, significance tests, or multiple-seed results are reported. The central quantitative claim is an average improvement of about 1.9 points, but several individual cells change by only 0.1--0.3 points (e.g., Charades R1@0.3: 70.5->70.4; ANet R1@0.3 for 8B: 68.9->68.8; Charades R1@0.3 for the VBEAR ablation: +0.3). Without statistical confidence, the reader cannot judge whether these differences are meaningful. Please add bootstrap confidence intervals or repeated-run standard deviations, especially for the ablations and the main baseline comparison.","section":"Tables 1-2"},{"comment":"The matched-IoU analysis in Figure 4(b) is a valuable control, but the paper does not specify how many sample pairs remain after matching, how the 'nearly identical IoU' threshold was chosen, or whether the reported macro average is stable. In addition, Eq. (9) defines the PAGE gate through f_exp but the explicit normalized expression is only promised in the appendix. Since PAGE is a named contribution, the exact function should appear in the main text or the appendix should be part of the submitted version. These details affect reproducibility of the central evaluation.","section":"Figure 4(b) and PAGE (Eq. 9)"}],"minor_comments":[{"comment":"The caption contains garbled threshold notation ('0τ!=0.3τ\"=0.61.0'), which should be cleaned up.","section":"Figure 3 caption"},{"comment":"The sentence 'CAVE-8B outperforms TimeLens-8B across all three benchmarks on nearly every metric' is technically accurate, but the single regression (68.8 vs 68.9) should be mentioned or the wording adjusted.","section":"Table 1, ANet-TL R1@0.3"},{"comment":"The similarity-guided expansion that constructs W_b is described only verbally. Please provide the exact algorithm or a precise formula, including how max_neighborhood_size interacts with the similarity threshold, so the result is reproducible.","section":"Method, W_b construction"},{"comment":"The boundary-level evidence density D_num uses the mean attention over W_b and over the whole video. It would help to state how W_b is sized and whether the results are sensitive to that window size, since the main analysis uses a single window configuration.","section":"Preliminary Study, Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"The paper is promising and the core idea is worth publishing if the requested analyses confirm the mechanism. The main risk is the mismatch between the quantity optimized (evidence-token attention) and the quantity evaluated (timestamp-token attention), combined with the absence of significance testing. These are fixable within the manuscript's scope, so I am not recommending rejection. I would also gently note that the authors should ensure the appendix is included in the revised submission, as several exact formulations are deferred there."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a real, incremental-but-solid contribution to RL-based video temporal grounding. The new part is the boundary-specific <Start>/<End> evidence tokens, initialized with a supervised warm-up and then rewarded during GRPO for concentrating visual attention inside the ground-truth boundary neighborhoods. The misalignment diagnosis in Figure 2 is a fresh and useful analysis, and the matched-IoU control in Figure 4 is exactly the right kind of check. I'd send this to review.\n\nWhat's good: the ablations are clean. Adding VBEAR to the warm-up+GRPO baseline gives consistent gains across all three benchmarks (Table 2, row 1 vs 2), and removing the warm-up attention loss hurts, which shows the warm-up is doing more than just teaching format. The improvements are largest at stricter IoU thresholds, which matches the method's stated goal. The 8B backbone results also show the effect is not a 7B-only fluke. The paper mostly claims what it can test, and it does not oversell.\n\nSoft spots, in rough order of severity. First, no error bars or significance tests anywhere. Table 1 already contains one small regression (Charades R1@0.3), so the average +1.9 points could be within noise on some cells. Second, the mechanism is underdetermined. VBEAR rewards the evidence tokens' attention, but those tokens are generated after the numeric timestamps. The gradient to the earlier timestamp tokens is indirect, and the stress-test worry—that the gains could come from the warm-up or the format reward rather than from evidence alignment—is not directly refuted. The matched-IoU analysis (Figure 4b) is the best counterevidence, but it evaluates timestamp-token attention, not the evidence-token attention being rewarded, so the connection is still inferred. Third, there are many hand-set thresholds (tau_on, tau_sat, the PAGE intervals, kappa, and the lambda weights) with no sensitivity analysis. Fourth, key construction details are deferred to the appendix, and no code is released, so reproduction would be difficult.\n\nThe central claim—that boundary-level visual supervision helps and reduces evidence–timestamp misalignment—is plausible and mostly supported. It just needs a sharper mechanism test and standard statistical hygiene. This deserves a serious referee, not a desk reject. I'd ask for code, error bars, and an experiment that either rewards timestamp-token attention directly or ablates the evidence tokens at inference to separate the format effect from the alignment effect.\n\nWho it's for: people working on RL reward design for video-language grounding, and anyone interested in attention-level supervision in MLLMs. I'd bring it to a reading group and would likely cite it.","headline":"A genuinely new attention-reward idea for RL-based VTG, with a nice diagnosis and clean ablations; the main open question is whether the reward really shapes the timestamps or just the added tokens.","tokens_in":13966,"tokens_out":2588,"would_cite":true,"duration_ms":23060,"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":"CAVE claims that augmenting RL training with a boundary-evidence alignment reward turns visual boundary perception into accurate timestamps, beating TimeLens-7B by about 1.9 points averaged over twelve metrics.","keywords":["video temporal grounding","large vision-language models","reinforcement learning","visual evidence alignment","boundary evidence tokens","reward design","evidence-timestamp misalignment","attention supervision"],"falsifier":"Train CAVE with the evidence tokens emitted before the numerical timestamps. If VBEAR then yields no improvement over the same warm-up + GRPO baseline, the reward's benefit depends on the current token order's indirect credit path and the evidence-alignment explanation is falsified.","tokens_in":13167,"feed_emoji":"🎯","tokens_out":4790,"duration_ms":36215,"temperature":0.7,"pith_summary":"This paper argues that reinforcement-learned video temporal grounding models often see the right boundary—their attention peaks at the true start/end frames—yet still emit off-target timestamps. The authors propose CAVE, which adds dedicated <Start> and <End> evidence tokens, warms them up with supervised attention, then rewards each token for concentrating on the ground-truth boundary neighborhood while suppressing distracting salient regions during RL. A competence-aware gate scales this evidence reward by how well each rollout group is already localizing. If correct, CAVE closes a gap that pure outcome rewards leave unconstrained: internal boundary evidence gets translated into accurate timestamps. The paper reports CAVE-7B beating the TimeLens-7B baseline by about 1.9 points averaged over twelve metrics.","feed_headline":"Aligning evidence to boundaries lifts video grounding by ~1.9 pts","feed_subtitle":"CAVE adds start/end evidence tokens and gates the reward, turning boundary perception into accurate timestamps.","key_machinery":"Boundary-specific evidence tokens <Start>/<End> combined with the Visual Boundary Evidence Alignment Reward (VBEAR) and Performance-Aware Gating (PAGE). The tokens are autoregressively generated after the numerical interval, so they summarize the visual evidence associated with the completed temporal decision. VBEAR measures, per boundary, the ratio of the token's mean attention within the ground-truth boundary neighborhood to its mean attention on salient off-boundary frames, and rewards values in (1,2] with saturation. The burden of the argument is carried by this ratio and the gating: it provides dense, boundary-localized visual supervision that standard IoU rewards cannot.","core_discovery":"The central claim is that current VTG RL methods suffer from visual evidence–timestamp misalignment: a substantial fraction of incorrectly localized predictions nonetheless show strong model attention around both ground-truth boundaries, so boundary perception is not reliably converted into the numerical start/end tokens. CAVE attacks this by inserting two special tokens after the predicted interval, <Start> and <End>, whose cross-attention over video frames is supervised in a short warm-up to encode boundary-specific semantics. During RL, the Visual Boundary Evidence Alignment Reward compares each token's mean attention inside a similarity-expanded ground-truth window against its attention","pith_inferences":["The evidence-token ratio ρ_b could be reused as a model-free confidence score at inference, flagging predictions whose boundary evidence is weak; the paper does not propose this.","The similarity-guided window expansion and salient off-boundary reference could be adapted to audio or other modalities, where a parallel evidence–output misalignment may exist.","A testable extension: use VBEAR as a verifier in a search/reranking step rather than a policy gradient; since it is a scalar reward on attention, it could rank multiple candidate intervals at inference without further training.","The smaller gains on Charades-TimeLens hint that the value of evidence alignment depends on annotation precision; datasets with coarser boundaries may need looser window construction. This is our inference, not the paper's claim."],"forward_implications":["If CAVE is right, RL-based VTG can be improved without extra annotation: the evidence tokens are trained on the same ground-truth intervals already used for IoU rewards.","The method should generalize to stronger backbones: CAVE-8B outperforms TimeLens-8B on nearly every metric across the three datasets.","The ablation results imply that warm-up of boundary semantics is necessary; removing the attention loss L_att consistently degrades all metrics.","Because VBEAR evaluates each boundary independently, it prevents one boundary's strong evidence from masking the other's weakness, yielding gains on strict IoU thresholds (0.7)."],"fun_headline_variants":["CAVE aligns boundary evidence to timestamps, lifting VTG by 1.9 pts","Special start/end tokens fix evidence-timestamp mismatch in video grounding","Gated evidence reward sharpens video temporal grounding accuracy","CAVE: turn boundary attention into precise predictions with +1.9 pts","Fix VTG misalignment with CAVE's competence-aware boundary tokens"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The policy-gradient signal from VBEAR must transfer backward through the whole autoregressive completion to the earlier timestamp tokens; if this indirect credit assignment is too weak, the reported gains would come from the format warm-up or output regularization rather than from evidence alignment.","fun_headline_variants_meta":{"raw":{"variants":["CAVE aligns boundary evidence to timestamps, lifting VTG by 1.9 pts","Special start/end tokens fix evidence-timestamp mismatch in video grounding","Gated evidence reward sharpens video temporal grounding accuracy","CAVE: turn boundary attention into precise predictions with +1.9 pts","Fix VTG misalignment with CAVE's competence-aware boundary tokens"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000259,"raw_usage":{"total_tokens":1422,"prompt_tokens":740,"completion_tokens":682,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":587}},"tokens_in":484,"tokens_out":682,"duration_ms":5806,"temperature":1.0,"reasoning_tokens":587,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T15:40:05.323913+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train CAVE with the evidence tokens emitted before the numerical timestamps. If VBEAR then yields no improvement over the same warm-up + GRPO baseline, the reward's benefit depends on the current token order's indirect credit path and the evidence-alignment explanation is falsified.","supporting_citations":[],"review_version":1}