{"id":"3659cd5c-31e2-4f98-ba9b-d005f5ca630c","arxiv_id":"2608.11201","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"VidForensics-M1 uses temporal grounding evidence, injected through a reward-conserving redistribution rule, to improve AI-generated video detection accuracy and cross-generator generalization.","lead":"This paper trains a video detector to both say whether a video is real or AI-made and point to the exact seconds where the forgery happens. It introduces a reward redistribution scheme that shifts credit toward correct answers with accurate time localization, reporting gains across many video generators.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Temporal-vs-textual comparison is confounded: Section 5.1 filters the textual training set with Gemini, so Label-Exp-RL sees fewer samples than Label-Tem-RL; reported gains may reflect data volume, not evidence quality.","rationale":"The reader's verdict is CONDITIONAL, and my analysis supports keeping it there. The theoretical appendix is internally sound: Lemma 1's reward conservation, Theorem 2's orthogonal decomposition, and Theorem 3's expected evidence-gradient direction all follow from the stated assumptions and appear parameter-free. The empirical weakness lies in the controlled comparison between evidence types. The paper goes to lengths to binarize temporal evidence for fairness but does not state that both RL variants see identical videos; if anything, the pipeline description implies Label-Exp-RL uses a Gemini-filtered subset while Label-Tem-RL uses the full 100K set. Since the main quantitative evidence that temporal grounding is superior is the delta over Label-Exp-RL, this is the load-bearing comparison. The distribution-shift concern (segment-replacement fakes versus full-generation fakes in ViF-Bench and GenBuster-Bench) is also important and deserves a separate control, but it does not need to be resolved to see that the current temporal-vs-textual numbers are not yet a clean test. I therefore keep the verdict at CONDITIONAL, i.e., unchanged from the reader, and would require the matched-data ablation, plus ideally a distribution-shift analysis, before accepting the generalization claim.","tokens_in":20576,"tokens_out":10184,"duration_ms":94541,"concrete_test":"Retrain Label-Tem-RL and Label-Exp-RL on exactly the same filtered training set (the real/fake pairs that survive the Gemini-3.1-Pro filter), using the same base model, DAPO hyperparameters, number of steps, and group size, and changing only the evidence reward: binarized temporal IoU versus Qwen3.5-4B semantic consistency. If Label-Tem-RL no longer beats Label-Exp-RL by the reported roughly 6-point margins on ViF-Bench, the temporal-over-textual conclusion is an artifact of training-set size or filtering. If the gap persists, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.1 describes a 100K-sample balanced set, then says GPT-5.5 generates textual explanations and Gemini-3.1-Pro filters out samples with unreliable explanations, removing their paired real videos to maintain a 1:1 ratio. Label-Exp-RL is therefore trained on a strict subset of the videos available to Label-Tem-RL. The only equalization the paper explicitly makes is binarizing temporal IoU at 0.7 to match the binary textual consistency score; it never equalizes the training distribution. The headline comparisons in Table 1 (+6.25% accuracy, +6.96% F1 for Label-Tem-RL over Label-Exp-RL) and the GenBuster-Bench gains in Table 2 are thus potentially attributable to more training data rather than to temporal grounding being a better evidence source. Because the central claim is precisely that rule-based temporal grounding is more suitable than model-based textual rationales, this confound must be controlled before the claim can be accepted. A secondary distribution-shift risk (segment-replacement fakes versus full-generation test videos) is real, but it would not by itself invalidate the temporal-vs-textual comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces VidForensics-M1, a reinforcement-learning framework for AI-generated video detection that jointly optimizes label correctness and supporting evidence. It proposes an automated data construction pipeline that creates paired real/fake videos by removing a random temporal segment from a real video, reconstructing the missing content with boundary-frame-conditioned generative models, and reinserting the aligned segment. This process yields ground-truth manipulated temporal intervals. The authors also propose Evidence-Guided Reward Redistribution (EGRR), which redistributes the label-level reward among label-correct responses according to evidence quality while conserving total reward mass; the theoretical analysis in Appendix A proves conservation, correctness dominance, orthogonality of label and evidence signals, and an expected evidence-improving gradient. The central empirical claim is that rule-based temporal grounding provides more reliable and verifiable feedback than model-based textual rationales for meta-detection, supported by comparisons of Label-Tem-RL versus Label-Exp-RL on ViF-Bench and GenBuster-Bench.","tokens_in":20798,"tokens_out":3465,"duration_ms":30033,"significance":"If the central claim holds, the paper contributes a practical mechanism for incorporating verifiable, rule-based evidence into RL training of video forensics detectors, with potential to improve generalization to unseen generators. The proposed data construction pipeline is a valuable and reusable resource, and the EGRR theoretical analysis is mathematically sound and clearly presented. However, the main empirical comparison is currently confounded by unequal training data between the temporal and textual branches, and the generalization claim rests on an unexamined distribution shift between constructed segment-replacement fakes and fully generated test videos. These issues must be resolved before the paper's central claim can be accepted.","major_comments":[{"comment":"The comparison between Label-Tem-RL and Label-Exp-RL is confounded by the Gemini-3.1-Pro filtering step applied only to the textual branch. The paper states that Gemini-3.1-Pro is used to \"filter out samples with unreliable textual explanations and remove their corresponding real videos to maintain a 1:1 ratio\" (Section 5.1), so Label-Exp-RL is trained on a strict subset of the 100K balanced set available to Label-Tem-RL. The only equalization mentioned is binarizing temporal IoU at 0.7 to match the binary textual consistency score; no effort is made to equalize the training distribution. The headline gains in Table 1 (+6.25% accuracy, +6.96% F1) and the GenBuster-Bench gains in Table 2 (+6.4% OOD, +11.4% Wild recall) could therefore be attributable to more training data rather than to the superiority of temporal grounding as an evidence source. To support the central claim, the paper must train Label-Exp-RL on an equal-size random subset of the retained samples or otherwise control for data volume.","section":"§5.1, Tables 1 and 2"},{"comment":"All reported metrics are point estimates without error bars, confidence intervals, or significance tests. MLLM training and evaluation are known to be noisy, and the claimed improvements over Label-Exp-RL (e.g., 6.25% accuracy and 6.96% F1 on ViF-Bench) are small enough that they could fall within run-to-run variability. The paper should report results from multiple seeds with means and standard deviations, or at least provide a significance test, to substantiate the empirical claims.","section":"§5.2, Tables 1 and 2"},{"comment":"The training data for VidForensics-M1 consists of segment-replacement fakes created by removing a random interval from a real video and reinserting a generated segment, while the evaluation benchmarks include fully generated videos from models such as Sora-2 and Kling-V1. The paper provides no analysis of the distribution shift between these two forgery types, and no ablation to determine whether the detector relies on boundary seams, content inconsistency, or other artifacts unique to the construction process. Without such analysis, the reported GenBuster-Bench gains (Table 2: +10.1% OOD and +19.2% Wild recall for Label-Tem-RL over Label-RL) may not reflect genuine generalization to unseen generative models. The authors should include experiments that vary the forged-segment ratio, add post-processing to remove seams, or evaluate on an additional fully generated held-out set to verify that the learned signal transfers.","section":"§3.3 and §5.2"}],"minor_comments":[{"comment":"The figure caption contains typographical errors: \"start tmie end tmie\" should read \"start time end time\".","section":"Figure 2"},{"comment":"The row label \"+ Label-Tem-RL\" appears after the numeric values in the row, which is confusing; the label should precede the metrics for clarity, as in the other method rows.","section":"Table 1"},{"comment":"Reference [37] appears to be a duplicate of reference [11]; please consolidate or distinguish them.","section":"References"},{"comment":"The notation for interval intersection and union in the IoU definition is not explicitly defined; please state that ∩ and ∪ denote interval intersection and union, respectively.","section":"Equation (2)"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid theoretical contribution and a useful data construction pipeline, but the central empirical claim is currently undermined by the training-set confound between the temporal and textual branches. The lack of error bars or significance tests further weakens the headline quantitative comparisons. I recommend major revision; the authors should be able to address the confound by re-running Label-Exp-RL on a matched subset and adding multi-seed results. The distribution-shift concern regarding segment-replacement fakes versus fully generated videos is also important for the generalization claim and should be addressed with ablations or additional evaluations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this one. First, the theoretical core is genuinely good: the EGRR scheme, with its exact reward-conservation identity and the expected-gradient ascent result, is correct and is a nice contribution to RL-based forensics. Second, the headline comparison fails as written. Section 5.1 says they filter the textual-explanations set with Gemini-3.1-Pro, removing samples with unreliable explanations and their paired real videos, so Label-Exp-RL is trained on a strict subset of the videos used for Label-Tem-RL. The reported gains of +6.25 accuracy and +6.96 F1 over Label-Exp-RL, and the bigger GenBuster-Bench numbers, could just be a data-volume effect. The paper never equalizes the training distribution; it only binarizes the IoU threshold. That is a real confound and it sits exactly on the central claim.\n\nWhat the paper does well: EGRR is worth taking seriously. Lemma 1 shows the reward adjustment is zero-sum, Theorem 2 gives an orthogonal decomposition, and Theorem 3 proves the expected evidence gradient is an ascent direction. The proof is clean and the fallback to plain label-RL when evidence has zero variance is a nice property. The automated segment-replacement pipeline is also a sensible way to produce paired real-fake videos with ground-truth temporal intervals, and the idea of using rule-based supervision to avoid reward hacking is reasonable.\n\nSoft spots beyond the confound: all tables are point estimates with no error bars or significance tests; no code or data are provided, so the large OOD recall gains are not yet independently checkable. The secondary distribution-shift concern is real but not fatal to the whole approach: constructed fakes with boundary-conditioned segment replacement might carry artifacts distinct from fully generated test videos, but that would not invalidate the method, just the generality of the numbers.\n\nWho this is for: people working on MLLM-based synthetic-video detection and anyone interested in reward shaping for RLVR. The theoretical appendix alone justifies a look. My recommendation: send it to peer review, but the authors should be required to control the training-set confound (e.g., train temporal grounding on the same filtered subset), provide error bars or significance tests, and release artifacts. The theory is solid enough that this deserves a serious referee, but the empirical claim should not be accepted in its current form.","headline":"The EGRR reward-redistribution theory is clean and genuinely useful, but the paper's central empirical claim that temporal grounding beats textual rationales is undermined by a training-set confound in Section 5.1.","tokens_in":21370,"tokens_out":1754,"would_cite":true,"duration_ms":17818,"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":"AI-generated video detection improves sharply when reinforcement learning rewards not just the predicted label but the exact manipulated time interval, and this rule-based temporal evidence beats model-written text rationales as training…","keywords":["AI-generated video detection","video forensics","meta-detection","evidence-guided reward redistribution","temporal grounding","reinforcement learning","multimodal large language models","GRPO"],"falsifier":"Evaluate the detector on fully AI-generated videos that contain no spliced real footage and hence no boundary seam or content-mismatch cue, and compare recall against the same model trained with label-only reinforcement learning; if the roughly ten-point advantage disappears, the temporal evidence was teaching construction fingerprints rather than generalizable forgery detection. A companion check: fully generated videos have no ground-truth manipulated interval, so the detector's \"verifiable\" localization cannot be scored by $\\mathrm{IoU}$ — asking humans whether the flagged seconds contain visible artifacts, and testing whether the detector fails on segment-replacement fakes whose regenerated segment is unusually well matched to its boundaries, would settle whether temporal grounding is genuinely the source of the gain.","tokens_in":20377,"feed_emoji":"🎬","tokens_out":11895,"duration_ms":135149,"temperature":0.7,"pith_summary":"The paper argues that the main obstacle to detecting AI-generated video is the supervision signal, not model size: label-only rewards let a detector exploit shortcuts, and text explanations of artifacts inherit the biases of the model that wrote them. Its proposal, VidForensics-M1, is a \"meta-detection\" training loop that rewards the predicted label and the supporting evidence together, using the manipulated time interval as verifiable evidence. Training fakes are built by cutting a random segment out of a real clip, regenerating that segment from its boundary frames, and splicing it back, so the forged interval is known exactly by construction. The paper reports that training with temporal grounding lifts accuracy, recall, and F1 by roughly ten points over label-only reinforcement learning on ViF-Bench and raises recall on fully generated videos in GenBuster-Bench, indicating that evidence-based feedback carries over to unseen generators.","feed_headline":"Reward the exact tampered seconds; fake-video detection jumps 10%","feed_subtitle":"Temporal grounding beats text rationales by up to 11 points on unseen video generators.","key_machinery":"Two coupled mechanisms carry the argument. First, the automated data construction pipeline removes a random temporal segment from each real 5-second video, regenerates the missing content from its boundary frames with video generation models, aligns duration, resolution, and frame rate, and reinserts it, producing paired real-fake videos whose manipulated interval is ground truth by construction. Second, Evidence-Guided Reward Redistribution (EGRR) modifies the reinforcement-learning reward so that among label-correct responses for a fake video, reward flows from responses whose predicted interval has below-group-average temporal IoU to those with above-average IoU, via the additive term $\\ell_{ij}(e_{ij}-q_i)$ in the semantic reward, where $\\ell_{ij}$ is label correctness, $e_{ij}$ is the IoU between predicted and true forged intervals, and $q_i$ is the mean evidence quality among label-correct responses in the sampled group. The zero-sum identity $\\sum_j \\ell_{ij}(e_{ij}-q_i)=0$ guarantees the label-reward mass is conserved exactly, so EGRR only reallocates credit within the correct-label subset; when all sampled evidence has equal quality it degenerates to plain label-level RL. A theorem shows that the expected evidence gradient is an ascent direction for evidence quality conditioned on a correct label, so the scheme provably teaches localization without disturbing label learning.","core_discovery":"The central claim is that rule-based temporal grounding is a more reliable and verifiable meta-detection feedback signal than model-based textual rationales. Detectors trained with the Evidence-Guided Reward Redistribution (EGRR) scheme, in which reward among label-correct responses is shifted toward those whose predicted forged interval overlaps the ground-truth interval (temporal $\\mathrm{IoU} \\geq 0.7$), outperform label-only reinforcement learning by 10.30% accuracy, 9.74% recall, and 11.21% F1 on ViF-Bench, and outperform text-rationale supervision by 6.25%, 6.38%, and 6.96%; on the fake-only subset of GenBuster-Bench, recall improves by 10.1% (OOD) and 19.2% (Wild) over label-only RL. The paper attributes the difference to objectivity: temporal intervals are recorded by the construction process as ground truth, whereas textual rationales are generated by a reference model and verified by another model, so they remain vulnerable to hallucination, semantic bias, and reward hacking. A side observation is that temporal grounding also teaches the detector to localize the forged segment, and its free-form explanations become more accurate as well.","pith_inferences":["If the central claim is right, the cheapest lever for improving detector generalization is controlling how training forgeries are made so that ground-truth evidence is mechanically extractable, rather than scaling models or accumulating more generators; the same construct-then-verify recipe should transfer to image splicing and audio deepfakes.","An implicit corollary the paper does not test: the gap between temporal grounding and text rationales should widen as generators improve, because fixed artifact categories age out while interval-localization pressure keeps forcing evidence-based decisions.","A clean check of the causal story would train on fully generated fake videos (no real context, no splice seams) while rewarding localization through a rule-based proxy, for instance consistency between two independently sampled generations, and see whether the accuracy gain persists; the paper's construction leaves this variant untested."],"forward_implications":["Detectors trained with temporal-grounding reward generalize to generators unseen in training: the gains on ViF-Bench are spread across 18 models including Sora-2 and Kling-V1, and GenBuster-Bench fake-only recall rises 10.1% on OOD and 19.2% on Wild data over label-only RL.","The same evidence signal teaches temporal localization itself: interval-trained detectors output forged intervals that overlap ground truth (tIoU 0.81 in the reported example), whereas text-trained detectors report the whole clip as suspicious.","Evidence-guided training can be layered onto label-level RL without changing the label objective: EGRR conserves label reward mass exactly and reduces to plain label-level RL whenever sampled evidence carries no discriminative information.","Model-written textual rationales are the weaker evidence source on every reported metric, consistent with the claim that their generation and verification depend on external models and are prone to hallucination and semantic bias."],"supporting_citations":[{"why":"Provides the supervised-fine-tuning baseline (teacher-generated visual chains of thought) that the paper contrasts with meta-detection.","marker":"[10]"},{"why":"DeepTraceReward, the reward-model baseline trained from human rationales and temporal intervals, whose evidence is model-mediated rather than rule-based.","marker":"[11]"},{"why":"Skyra, the grounded-artifact-reasoning method whose ViF-Bench evaluation the paper reports on, plus the CoT-annotation paradigm it extends.","marker":"[12]"},{"why":"VidGuard-R1, the label-level GRPO training recipe that serves as the direct label-RL baseline for EGRR.","marker":"[16]"},{"why":"BusterX++, the pure label-level RL detector used as a comparison point in both benchmarks.","marker":"[17]"},{"why":"BusterX, source of the GenBuster-200K training data and the GenBuster-Bench fake-only evaluation.","marker":"[32]"},{"why":"Grounded-VideoLLM, which establishes precise temporal grounding in video LLMs and backs the choice of interval evidence.","marker":"[23]"},{"why":"The Qwen3.5-9B base detector whose policy is optimized by the proposed method.","marker":"[42]"},{"why":"DAPO, the reinforcement-learning algorithm used to optimize the detector.","marker":"[50]"}],"fun_headline_variants":["Reward tampered timestamps in RL; fake-video detection improves 10%","Temporal grounding beats text rationales; fake-video detection gains 10%","Use tampered intervals as RL evidence; forgery detection up 10%","Exact tamper times as RL reward lift fake-video detection by 10%","Timing, not text, drives stronger AI-video forgery detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a fake video made by deleting a random segment of a real clip and regenerating that segment from its boundary frames is representative of the AI-generated videos the detector is tested on, including fully generated clips from Sora-2, Kling-V1, and others; the paper offers no analysis of how much of its reported gain depends on splice seams and construction artifacts unique to its own pipeline.","fun_headline_variants_meta":{"raw":{"variants":["Reward tampered timestamps in RL; fake-video detection improves 10%","Temporal grounding beats text rationales; fake-video detection gains 10%","Use tampered intervals as RL evidence; forgery detection up 10%","Exact tamper times as RL reward lift fake-video detection by 10%","Timing, not text, drives stronger AI-video forgery detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001025,"raw_usage":{"total_tokens":4385,"prompt_tokens":1075,"completion_tokens":3310,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":691,"completion_tokens_details":{"reasoning_tokens":3205}},"tokens_in":691,"tokens_out":3310,"duration_ms":20817,"temperature":1.0,"reasoning_tokens":3205,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:12:31.181269+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the detector on fully AI-generated videos that contain no spliced real footage and hence no boundary seam or content-mismatch cue, and compare recall against the same model trained with label-only reinforcement learning; if the roughly ten-point advantage disappears, the temporal evidence was teaching construction fingerprints rather than generalizable forgery detection. A companion check: fully generated videos have no ground-truth manipulated interval, so the detector's \"verifiable\" localization cannot be scored by $\\mathrm{IoU}$ — asking humans whether the flagged seconds contain visible artifacts, and testing whether the detector fails on segment-replacement fakes whose regenerated segment is unusually well matched to its boundaries, would settle whether temporal grounding is genuinely the source of the gain.","supporting_citations":[{"cited_title":"Qwen3.5: Towards native multimodal agents.https://qwen.ai/blog?id= qwen3.5, February 2026","cited_arxiv_id":null,"evidence_quote":"The Qwen3.5-9B base detector whose policy is optimized by the proposed method."},{"cited_title":"Dapo: An open-source llm reinforcement learning system at scale.Advances in Neural Information Processing Systems, 38:113222– 113244, 2026","cited_arxiv_id":null,"evidence_quote":"DAPO, the reinforcement-learning algorithm used to optimize the detector."}],"review_version":2}