{"id":"49b67bb5-da93-4936-9d0b-d58f73d99076","arxiv_id":"2507.18100","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A supervised fine-tuning plus difficulty-filtered reinforcement learning recipe improves video temporal grounding on three benchmarks, with datasets and models released.","lead":"The paper describes a two-stage training recipe, supervised fine-tuning followed by reinforcement learning, that improves how open video-language models locate the moment in a video described by a text query. It also releases the intermediate datasets and models used to produce these gains.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that VTG-R1 'consistently outperforms previous models' is not yet supported because the closest RL-based temporal-grounding baselines (Time-R1 and Video-R1) are cited but never compared; on NExT-GQA, TVG-R1's R@0.5 of 20.78 even trails VideoChat-TPO's 23.40.","rationale":"The reader's verdict is CONDITIONAL and I agree with that overall assessment. The reader's stated weakest_assumption concerns the accuracy of source annotations and the reliability of Gemini-2.5-Pro CoT filtering. That is a legitimate concern about training-signal noise, and it is worth a manual audit. However, I find the absence of the closest RL-based baselines more load-bearing for the paper's headline claim of consistent SOTA. The claim is comparative; if the most similar methods outperform VTG-R1, the claim collapses regardless of annotation quality. The reader's rationale does list the missing baselines as one of the conditions for acceptance, so our positions overlap. I would keep the verdict CONDITIONAL: the proposed check on Time-R1 and Video-R1 is a necessary condition for the claim, and should be added before acceptance. I do not see internal inconsistency in the method; the ablations (Tables 3 and 4) are informative and consistent with the recipe's benefits. The recommendation to add error bars is also important but secondary to the baseline comparison.","tokens_in":10587,"tokens_out":8727,"duration_ms":89412,"concrete_test":"Run the released checkpoints of Time-R1 and Video-R1 through the same VideoMind evaluation harness used for VTG-R1 on NExT-GQA, ReXTime, and Charades-STA, using identical prompts (including the TVG-R1 think/time template) and decoding settings, and report all metrics for at least three seeds. If either baseline matches or exceeds VTG-R1 on the majority of the nine reported R@m/mIoU measures across the three benchmarks, the 'consistently outperforms' claim fails. This check directly tests whether the omitted comparisons would overturn the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest claim, 'VTG-R1 consistently outperforms previous models on most benchmarks' (Section 4.2, Tables 2 and 5), depends on the comparison set being representative of the state of the art. Yet the two most relevant baselines for this claim are absent from all experiments: Time-R1 [42] and Video-R1 [12], both of which apply RL-style post-training to LVLMs for video reasoning and temporal grounding. These works are cited in the related work ('recent methodologies have applied GRPO to multimodal image reasoning tasks... and video understanding [12,42]'), so the authors are aware of them. Without these comparisons, 'consistently outperforms' is too broad. The issue is not merely stylistic: on NExT-GQA, TVG-R1's R@0.5 (20.78) is below VideoChat-TPO (23.40) and Qwen2.5-VL-32B (22.26), so even within the reported set the claim requires qualification. Because all results are single-run with no error bars, we cannot tell whether the gains on Charades-STA (e.g., R@0.3 70.75 vs base 68.98) are within noise. The load-bearing assumption is therefore that the omitted baselines do not change the ranking; this is currently untested.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage training recipe for video temporal grounding (VTG) with an open 7B LVLM (Qwen2.5-VL-7B): first, an SFT cold start on 13K Gemini-2.5-Pro-generated chain-of-thought samples filtered to IoU > 0.8 against source ground truth; second, GRPO reinforcement learning with a weighted IoU-plus-format reward on an 18K subset that excludes samples with IoU < 0.4. The paper reports results on NExT-GQA, ReXTime, and Charades-STA, ablates the cold-start and RL-data filtering choices, and releases data, code, and models.","tokens_in":10908,"tokens_out":9210,"duration_ms":89574,"significance":"If the recipe holds, the contribution is practically valuable: it provides an open, reproducible pipeline for RL-based temporal grounding on a 7B backbone, with ablations that isolate two design choices (cold-start quality and RL difficulty filtering) that are often confounded in prior work. The release of intermediate datasets and checkpoints enables direct comparison and further research. However, the significance is currently limited by missing comparisons with the closest RL-based baselines, an unexplained SFT degradation, and internal inconsistencies between the main tables and the ablation tables.","major_comments":[{"comment":"The mIoU values reported for the same model configurations are inconsistent between the main results and the ablations. For example, TVG-ColdStart is reported with NExT-GQA mIoU 15.09 and ReXTime mIoU 10.18 in Table 2, but with 26.14 and 26.26 in Table 3; TVG-R1 appears as 29.25/28.20/46.73 in Tables 2 and 5 and as 30.41/26.38/48.78 in Tables 3 and 4. Since Tables 3 and 4 are the central evidence for the filtering claims, this discrepancy must be resolved (e.g., by stating which checkpoint, split, or seed each table uses) or the conclusions drawn from those tables are unsupported.","section":"Tables 3 and 4 vs Tables 2 and 5"},{"comment":"The claim that VTG-R1 'consistently outperforms previous models on most benchmarks' is not supported relative to the closest RL-based baselines. Time-R1 [42] and Video-R1 [12] are cited in Section 2 but never evaluated, and on NExT-GQA VTG-R1's R@0.5 of 20.78 is below VideoChat-TPO (23.40) and Qwen2.5-VL-32B (22.26). Please add these comparisons or restrict the claim to the evaluated baseline set.","section":"Section 4.2, Tables 2 and 5"},{"comment":"The SFT cold start degrades performance substantially below the base model on all three benchmarks (e.g., Table 2: NExT-GQA mIoU 15.09 vs 20.87 for Qwen2.5-VL-7B; Table 5: Charades-STA R@0.3 42.23 vs 68.98). The paper's explanation ('possibly due to overfitting or limited generalization') is speculative and not backed by analysis. Because the recipe's first stage is claimed to provide 'robust multimodal alignment and structured reasoning,' the manuscript should analyze this degradation (e.g., output-format adherence, per-category errors, training loss) or reframe the cold start as a necessary initialization for RL rather than a performance-improving stage.","section":"Section 4.2, Tables 2 and 5"},{"comment":"All experiments appear to be single runs with no error bars or significance testing. Several reported gains are small (e.g., Charades-STA R@0.3: VTG-R1 70.75 vs Base 68.98; mIoU 46.73 vs 46.14). Without multiple seeds per configuration, the stability of the recipe's improvements cannot be assessed. Please report repeated runs (at least 2-3 seeds) for the main comparisons.","section":"Section 4.1 and Tables 2-5"},{"comment":"The data filtering relies on thresholds ϵ1=0.8 and ϵ2=0.4 applied to the IoU between Gemini-2.5-Pro's predicted segment and the source ground truth. This conflates difficulty with annotation noise, and the source datasets include weakly supervised collections (e.g., TACoS, DiDeMo, InternVid-VTime). The thresholds are fixed with no sensitivity analysis, and the CoT annotations are not validated (e.g., by human evaluation). Since the ablations show filtering is a key ingredient, a threshold sensitivity study or annotation-quality check is needed to support the recipe.","section":"Section 3.1, Appendix A.1"},{"comment":"The reward specification is inconsistent with the prompt template: the template requires the time span to be enclosed in ⟨time⟩...⟨/time⟩ tags, while the format reward rform checks for ⟨think⟩, ⟨/think⟩, ⟨answer⟩, and ⟨/answer⟩ tags and does not mention ⟨time⟩. Please clarify which format is actually used during training and align the description.","section":"Section 3.1 vs Section 3.3.1"}],"minor_comments":[{"comment":"The metric is written 'R1@m' but the standard notation in temporal grounding is R@m; please define it consistently and use a single notation throughout.","section":"Section 4.1"},{"comment":"References [19], [20], and [21] all cite the same DiDeMo paper, while [28] (used for NExT-GQA) is listed as a VideoChat paper; please correct and de-duplicate the bibliography.","section":"References"},{"comment":"The prompt template has formatting issues (e.g., 'system You MUST reason...' without spacing) and the placeholder {query / instance} is not defined.","section":"Section 3.1"},{"comment":"The three panels do not indicate the number of runs or seeds, and the y-axis ranges make the differences appear larger than they are; please add error bars or state explicitly that single curves are shown.","section":"Figure 2"},{"comment":"β is set to 0.0, which removes the KL regularization from the GRPO objective; the manuscript should comment on whether the reference policy is then used at all and on the implications for training stability.","section":"Appendix A.1"},{"comment":"The column headers 'NExTG.', 'RexT.', 'Charad.' should be defined (presumably mIoU); please clarify in the captions.","section":"Tables 3 and 4"}],"recommendation":"major_revision","confidential_remarks":"The internal inconsistency between Tables 2/5 and Tables 3/4 is the most serious issue and should be resolved before publication regardless of the other revisions. The authors should also verify that the released checkpoints reproduce the reported numbers. The related-work section cites a large number of the authors' own preprints; prioritizing the comparison with Time-R1 and Video-R1 will help position the contribution relative to the closest concurrent work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you do temporal grounding with LVLMs. The real contribution is the released TVG-Coldstart-13K and TVG-RL-18K sets plus the filtering recipe: Gemini-2.5-Pro CoT annotations, keep IoU > 0.8 for SFT, exclude IoU < 0.4 from RL. The two-stage SFT+RL framework is not new—Time-R1 and Video-R1 already did GRPO for this task—but the careful ablations on coldstart quality, RL data difficulty, and response length are the useful part. The finding that filtered SFT data beats three times more unfiltered data, and that RL on filtered data saves training from collapse, will be directly usable by people building on 7B open models.\n\nThe soft spots are real and mostly at the level of the central claim. The closest baselines for this exact recipe, Time-R1 and Video-R1, are cited in related work but never compared. On NExT-GQA, TVG-R1's R@0.5 of 20.78 is below VideoChat-TPO's 23.40 and Qwen2.5-VL-32B's 22.26, so even against the reported set the 'consistently outperforms' wording is too strong. All results are single-run without error bars, and the Charades-STA coldstart degradation is large and unexplained beyond a guess about overfitting. The IoU-based filtering does create a selection bias, but the ablations show it matters, so that is not a hidden flaw.\n\nThe circularity concern raised by the skeptic does not land. The RL reward is the same IoU metric used at evaluation, which is standard supervised RL; the filtering picks data by teacher IoU, not by the model's own predictions. No cheating there.\n\nThis paper is for practitioners who want a concrete data recipe and reproducible open resources, not for readers looking for a fundamentally new training paradigm. It deserves a serious referee. The revision should add the missing RL baselines, report variance across seeds, and either fix or explain the SFT degradation. With those changes, the recipe claim would be solid.","headline":"A genuinely useful data-recipe paper for temporal grounding that releases open resources and runs clean ablations, but its headline comparison omits the two closest RL baselines and the 'consistently outperforms' claim is broader than the evidence supports.","tokens_in":11501,"tokens_out":2693,"would_cite":true,"duration_ms":27383,"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":"A two-stage recipe of filtered supervised fine-tuning followed by reinforcement learning lets a 7B open video model outperform prior temporal grounding models on three benchmarks.","keywords":["video temporal grounding","reinforcement learning","GRPO","chain-of-thought","supervised fine-tuning","IoU reward","dataset curation","large vision-language models"],"falsifier":"Run the exact two-stage recipe with the teacher's chain-of-thought rationales replaced by plain timestamp answers in the cold-start set; if final R@0.5 scores do not drop on all three benchmarks, the claim that explicit reasoning drives the gains is false.","tokens_in":10379,"feed_emoji":"🎯","tokens_out":10881,"duration_ms":106501,"temperature":0.7,"pith_summary":"This paper argues that video temporal grounding in open 7B vision-language models can be made state-of-the-art by a two-stage training recipe rather than by larger models or more data. The first stage is a cold-start supervised fine-tuning on a small, high-quality set of chain-of-thought examples whose predicted segments overlap the ground truth by more than 0.8 IoU. The second stage applies reinforcement learning on a difficulty-filtered set, with samples below 0.4 IoU removed, rewarding temporal overlap plus correct reasoning format. On three benchmarks the final model beats prior instruction-tuned baselines, and the authors release the curated datasets, models, and code.","feed_headline":"Filtered SFT plus RL pushes 7B model to top grounding scores","feed_subtitle":"Cold-start reasoning data and difficulty-gated rewards push an open 7B model to top temporal grounding scores.","key_machinery":"The load-bearing mechanism is an IoU-gated data-and-reward pipeline. A high-capacity teacher model writes chain-of-thought rationales for a pooled set of moment-retrieval and query-grounding samples; only samples whose predicted interval overlaps the ground truth by more than 0.8 IoU enter the cold-start SFT set, and samples with IoU below 0.4 are excluded from RL. Training then runs Group Relative Policy Optimization, a PPO variant that normalizes rewards inside a sampled group of responses, on a reward $r_i = \\lambda_{\\text{tIoU}} \\cdot r_{\\text{tIoU}} + \\lambda_{\\text{form}} \\cdot r_{\\text{form}}$, where $r_{\\text{tIoU}}$ is the temporal Intersection-over-Union and $r_{\\text{form}}$ checks for the required reasoning tags. The cold start provides the initial formatting and reasoning behavior that RL then amplifies.","core_discovery":"On the paper's own terms, the central discovery is that explicit, verifiable reasoning can be injected into a video temporal grounding model through reinforcement learning, provided the model starts from a carefully filtered cold start. Supervised fine-tuning alone on the same data does not consistently help and can even regress; the gains appear only after RL, and they are larger when the cold-start data are filtered for high IoU and when the RL data exclude very hard or mislabeled samples. The reward that carries this is a weighted combination of Intersection-over-Union between the predicted and ground-truth temporal segment and a binary check that the response contains the required reasoning and answer tags. With this recipe, a 7B open model reaches the reported top R@0.3 and mean IoU numbers on three temporal grounding benchmarks.","pith_inferences":["A natural extension would make the difficulty filter a curriculum: instead of dropping samples below 0.4 IoU, train on them later or down-weight them, which may improve hard-sample robustness.","The same IoU-gated reward could be applied to other dense prediction tasks with a verifiable overlap metric, such as referring video segmentation or action segmentation.","Because the teacher rationales are filtered by IoU against source annotations, the recipe may silently under-represent genuinely hard cases where the teacher is right but the source segment is imprecise; auditing failures on such cases would be a useful stress test.","The released RL set is small, so a natural question is how performance scales as the filtered pool grows; the reported curves suggest diminishing returns may set in differently with and without cold start."],"forward_implications":["If the recipe generalizes, a small open model can rival much larger or specially designed grounding models without new architectures.","The finding that SFT alone can regress while RL after cold start improves suggests that RL, not instruction tuning, is the stage that should absorb scale in this task.","Because the data and code are released, the same cold-start-plus-filtered-RL pipeline can be reproduced and adapted to other video-language tasks.","The difficulty thresholds of 0.8 and 0.4 give a concrete starting point for tuning data curation in low-resource settings."],"supporting_citations":[{"why":"supplies the group-relative policy optimization that turns the rewards into policy updates.","marker":"[37]"},{"why":"defines the 7B vision-language backbone that the two-stage recipe starts from.","marker":"[2]"},{"why":"serves as one of the instruction-tuned temporal grounding baselines the method must beat.","marker":"[23]"},{"why":"serves as another instruction-tuned baseline in the benchmark comparison.","marker":"[35]"},{"why":"serves as a task-preference-optimized baseline in the benchmark comparison.","marker":"[45]"},{"why":"serves as one of the three evaluation benchmarks for temporal grounding.","marker":"[22]"},{"why":"serves as one of the three evaluation benchmarks.","marker":"[28]"},{"why":"serves as the third evaluation benchmark.","marker":"[21]"},{"why":"contributes a large set of query-grounding samples to the pooled source data.","marker":"[32]"},{"why":"contributes another query-grounding source to the pooled cold-start and RL data pools.","marker":"[41]"}],"fun_headline_variants":["RL after filtered SFT lifts 7B to top grounding scores","Cold-start data plus RL yields top video grounding","Difficulty-gated RL boosts 7B temporal grounding","Open 7B model tops grounding benchmarks via RL","Filtered SFT then RL: recipe for temporal grounding wins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole recipe assumes the ground-truth segments in the pooled source datasets are accurate enough that IoU filtering and the IoU reward are trustworthy signals; noisy annotations would be amplified by both the cold-start filter and the RL reward.","fun_headline_variants_meta":{"raw":{"variants":["RL after filtered SFT lifts 7B to top grounding scores","Cold-start data plus RL yields top video grounding","Difficulty-gated RL boosts 7B temporal grounding","Open 7B model tops grounding benchmarks via RL","Filtered SFT then RL: recipe for temporal grounding wins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000592,"raw_usage":{"total_tokens":2734,"prompt_tokens":866,"completion_tokens":1868,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":482,"completion_tokens_details":{"reasoning_tokens":1788}},"tokens_in":482,"tokens_out":1868,"duration_ms":15237,"temperature":1.0,"reasoning_tokens":1788,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:38:04.805596+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the exact two-stage recipe with the teacher's chain-of-thought rationales replaced by plain timestamp answers in the cold-start set; if final R@0.5 scores do not drop on all three benchmarks, the claim that explicit reasoning drives the gains is false.","supporting_citations":[{"cited_title":"Vtimellm: Empower llm to grasp video moments","cited_arxiv_id":null,"evidence_quote":"serves as one of the instruction-tuned temporal grounding baselines the method must beat."},{"cited_title":"Timechat: A time-sensitive multimodal large language model for long video understanding","cited_arxiv_id":null,"evidence_quote":"serves as another instruction-tuned baseline in the benchmark comparison."},{"cited_title":"Task preference optimization: Improving multimodal large language models with vision task alignment","cited_arxiv_id":null,"evidence_quote":"serves as a task-preference-optimized baseline in the benchmark comparison."},{"cited_title":"Rextime: Temporal grounding benchmark for reasoning-intensive videos","cited_arxiv_id":null,"evidence_quote":"serves as one of the three evaluation benchmarks for temporal grounding."},{"cited_title":"Rus- sell","cited_arxiv_id":null,"evidence_quote":"serves as the third evaluation benchmark."},{"cited_title":"Queryd: A video dataset with high-quality text and audio narrations","cited_arxiv_id":null,"evidence_quote":"contributes a large set of query-grounding samples to the pooled source data."}],"review_version":1}