{"id":"d2b45700-6373-40f7-a0c4-67590ec8f55f","arxiv_id":"2607.02959","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"RL post-training of a VLM agent with neuro-symbolic temporal-logic rewards for evidence retrieval raises Pass@1 by up to 8% and Pass@4 by 15% on long-video QA.","lead":"VSeek turns long-video QA into multi-turn natural-language search, then RL-trains the model with dense rewards from temporal-logic checklists of visual primitives. The result is higher Pass@k on long-video benchmarks with far fewer frames than uniform sampling.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"VETL reward gains over exact-match are small, inconsistent, and may not drive the headline 8%/15% improvements.","rationale":"The Reader correctly flags that VETL discards order and relies on an imperfect detector, but under-weights how little of the headline improvement is actually attributable to that reward. The systems contribution (natural-language search agent + RL) is real and the frame-efficiency numbers are attractive; the neuro-symbolic story is oversold relative to the ablations. Keeping CONDITIONAL is appropriate, but the load-bearing soft spot is the causal attribution of the 8%/15% numbers rather than the faithfulness of the TL layer alone. A multi-seed EM-vs-VETL comparison would settle it cleanly.","tokens_in":32980,"tokens_out":537,"duration_ms":5492,"concrete_test":"Re-train three independent seeds of VSeek-EM and VSeek-VETL under identical GRPO hyperparameters (Table 8) on the same 80/20 split; report mean±std of p@1 and p@4 on LongVideoBench, Video-MME and MLVU. If the VETL–EM gap is not statistically significant (or reverses) on any ID set, the claim that VETL drives the reported gains does not hold.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The abstract and strongest claim attribute Pass@1 gains of up to 8% and Pass@4 of 15% to \"explicitly optimizing for this verifiable evidence-seeking behavior\" via VETL. Table 1 shows the large jumps are almost entirely from agentic RL itself (VSeek-EM vs. Qwen3-VL-4B-Thinking: +8.8% p@1 on LongVideoBench, +8.7% on MLVU). VSeek-VETL vs. VSeek-EM is only +0.4 / +2.1 / +1.2 p@1 on the three ID sets, mixed or negative on OOD, and sometimes worse on p@4. Eq. 6 further discards the temporal operators that Algorithm 1 and the abstract emphasize, reducing the \"temporal-logic specification\" to a bag-of-primitives max-detection score with imperfect ViCLIP. Thus the causal link between the neuro-symbolic reward and the headline numbers is weak; the paper's own ablations (Tables 1, 4, 5) show most of the lift is from search + EM RL, not from verified dense rewards.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The paper presents VSeek, an agentic multi-turn framework that equips a VLM with natural-language search, subtitle search, and summary tools over a ViCLIP-indexed long video, then post-trains the policy with GRPO (and variants) so that the model jointly issues targeted queries and answers. To supply denser rewards than exact-match answer accuracy alone, the authors introduce VETL: an offline LLM+judge loop that compiles each question–answer pair into a temporal-logic formula over atomic visual primitives; the online reward (Eq. 6) is the fraction of those primitives detected by max-pooling ViCLIP scores over the retrieved clips (temporal operators are discarded). On LongVideoBench, Video-MME and MLVU the resulting VSeek-VETL model improves Pass@1 by up to ~9 % and Pass@4 by ~13 % relative to the Qwen3-VL-4B-Thinking base, with further gains on longer videos and a substantial reduction in frames processed.","tokens_in":33341,"tokens_out":734,"duration_ms":11638,"significance":"If the claimed gains hold under stricter controls, the work supplies a practical recipe for turning long-video QA into an RL-search problem and demonstrates that a neuro-symbolic checklist can serve as a dense, offline-verifiable reward. The open-sourced code, multi-algorithm ablations (GRPO/RLOO/GSPO), frame-efficiency numbers, and qualitative traces are concrete assets for the community. The core engineering insight—that natural-language search plus a bag-of-primitives reward already yields large lifts—is useful even if the temporal-logic machinery contributes only marginally.","major_comments":[{"comment":"Abstract and §1 attribute the headline “up to 8 % Pass@1 / 15 % Pass@4” gains to “explicitly optimizing for this verifiable evidence-seeking behavior” via VETL. Table 1 shows that the bulk of the lift already appears in VSeek-EM (exact-match reward only): +8.8 % / +8.7 % Pass@1 on LongVideoBench and MLVU versus the same base model. VSeek-VETL adds only +0.4 / +1.2 / +2.1 points on the three ID sets and is mixed or negative on OOD. The causal contribution of the neuro-symbolic reward is therefore overstated; the paper should either rephrase the claim or supply a controlled ablation that isolates VETL’s incremental effect after agentic search is already present.","section":null},{"comment":"§4.3 and Eq. (6) explicitly discard the temporal operators that Algorithm 1 and the abstract emphasize (“along with their temporal ordering”). The reward reduces to a bag-of-primitives max-detection score. Table 5 (VSeek-Direct) shows that even a non-TL decomposition yields nearly identical numbers, confirming that the formal temporal structure is not load-bearing. Either restore order-sensitive verification or remove the repeated claim that temporal logic is essential for the reward.","section":null},{"comment":"No error bars, statistical tests, or multiple random seeds are reported for any table. Given that the VETL–EM deltas are often <2 points, it is impossible to judge whether they are reliable. At minimum the main ID comparisons (Table 1) and the RL-algorithm sweep (Table 4) need standard deviations or bootstrap intervals.","section":null}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The punchline is simple: turning a 4B thinking VLM into a multi-turn natural-language search agent (visual clips, subtitles, summary) and post-training it with GRPO already delivers most of the reported 8–15 point Pass@k jumps on LongVideoBench/MLVU/Video-MME. VETL, the temporal-logic primitive-coverage reward, is a clean idea that helps a little more on in-domain long videos and encourages more targeted queries, but the tables show only +0.4 to +2.1 p@1 over pure exact-match RL, mixed or negative OOD, and sometimes worse p@4. The abstract’s causal claim is therefore a stretch.\n\nWhat is actually new is the combination. Prior video agents either prompt frozen models over captions or train rigid timestamp tools with heavy SFT. VSeek keeps the interface purely linguistic, indexes 8 s ViCLIP clips + subtitles, and lets the policy interleave <think>/<search>/<answer>. VETL then offline-compiles each Q+A into a TL formula via an LLM generate–judge loop, extracts the atomic visual primitives, and scores the trajectory by the fraction of those primitives that ever appear in the retrieved frames (Eq. 6). That is denser than answer-only and, as the authors note, the first use of TL as a training signal rather than post-hoc verification. The ablations (different RL algorithms, instruct vs thinking bases, direct vs TL decomposition) and the open-sourced code make the recipe reproducible.\n\nSoft spots are real but proportionate. Eq. 6 deliberately throws away the temporal operators that Algorithm 1 and the abstract advertise, reducing the “specification” to a bag-of-primitives max-detection score with an imperfect frozen detector. OOD specialization is admitted. No error bars or significance tests appear. Still, the ID gains, frame-efficiency numbers (≈32 frames vs 64 or hundreds), and qualitative traces showing better query specificity are credible. The paper is honest about the remaining bottlenecks (retriever quality, sequential search, non-chronological retrieval).\n\nThis is for people building video agents or RL post-training pipelines who need a practical dense-reward recipe. It will not reorganize the field, but it is a useful, carefully executed systems paper. I would send it to peer review; the core engineering is solid enough to deserve referee time even if the claims need tightening.","headline":"Agentic NL search + RL is the real lift; VETL is a sensible but modest dense-reward add-on whose causal role the abstract overstates.","tokens_in":33908,"tokens_out":596,"would_cite":true,"duration_ms":11907,"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":"Vision-language models answer long-video questions better when RL rewards them for retrieving a checklist of verifiable visual primitives, not just for the final answer.","keywords":["long video question answering","vision-language models","agentic search","reinforcement learning","temporal logic","neuro-symbolic rewards","visual evidence retrieval"],"falsifier":"Train an otherwise identical agent with only exact-match answer reward on the same long-video suites; if Pass@1 and Pass@4 gains disappear (or reverse) on videos longer than ten minutes, the claim that the VETL evidence reward is what drives the improvement collapses.","tokens_in":33924,"feed_emoji":"🎬","tokens_out":869,"duration_ms":7968,"temperature":0.7,"pith_summary":"Long-video question answering usually fails because models see only a fixed, uniform sample of frames and therefore miss the sparse moments that actually matter. This paper turns the problem into multi-turn natural-language search: an agent issues search queries, receives short clips, reasons, and searches again until it answers. The hard part is teaching the agent what a good search trajectory looks like. Outcome-only rewards (final answer correct or not) are too sparse and invite shortcuts. The authors therefore compile each question into a temporal-logic specification that yields a finite checklist of atomic visual primitives—objects, actions, relations—then give the agent dense credit whenever those primitives appear in the clips it retrieved. Optimizing for this evidence-seeking behavior raises Pass@1 by up to 8% and Pass@4 by 15% over the same base model on long-video benchmarks, while using far fewer frames than uniform sampling or caption-heavy pipelines.","feed_headline":"RL rewards that check visual evidence lift long-video QA","feed_subtitle":"Agents learn to search for the right clips, not just guess the final answer, gaining up to 15% Pass@4.","key_machinery":"VETL (Visual Evidence via Temporal Logic): an offline generate–judge loop that turns a question–answer pair into a temporal-logic formula over atomic visual primitives; the RL reward is then simply the fraction of those primitives that appear at least once in the agent’s retrieved clips (order ignored).","core_discovery":"Post-training a vision-language agent with dense rewards derived from temporal-logic decompositions of the question (rather than answer accuracy alone) produces a policy that actively searches for the right visual evidence and therefore answers long-video questions more accurately and more sample-efficiently than outcome-only RL or passive frame sampling.","pith_inferences":["If the primitive detector itself is weak, the reward becomes noisy; stronger open-vocabulary detectors or multi-modal verifiers would tighten the training signal without changing the overall architecture.","Discarding temporal order inside the reward may be fine for non-chronological search, but questions that truly hinge on sequence could benefit from a later stage that re-introduces order constraints.","The same checklist idea could be applied to other open-ended multimodal tools (audio, 3-D, long documents) wherever unit-test-style verification is currently missing."],"forward_implications":["Agentic long-video systems can be improved by shaping intermediate search steps rather than waiting for final-answer supervision.","Natural-language search interfaces become learnable once a verifiable primitive checklist supplies dense feedback.","Frame budgets can drop well below the usual fixed 64-frame sampling while accuracy still rises.","The same neuro-symbolic reward recipe can be reused across GRPO, RLOO and related RL algorithms with only modest hyper-parameter retuning."],"fun_headline_variants":["Temporal-logic rewards train VLMs to hunt visual evidence in long videos","Verified clip-search RL lifts long-video QA Pass@4 by 15%","Neuro-symbolic rewards turn passive VLMs into active long-video search agents","Dense formal-spec rewards beat answer-only RL on long-video questions","Agents learn to retrieve the right clips before answering long videos"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That an LLM-written temporal-logic checklist plus imperfect visual detection of its atomic primitives is a faithful enough stand-in for “correct context” even though the reward discards ordering and the detector can be wrong.","fun_headline_variants_meta":{"raw":{"variants":["Temporal-logic rewards train VLMs to hunt visual evidence in long videos","Verified clip-search RL lifts long-video QA Pass@4 by 15%","Neuro-symbolic rewards turn passive VLMs into active long-video search agents","Dense formal-spec rewards beat answer-only RL on long-video questions","Agents learn to retrieve the right clips before answering long videos"]},"model":"grok-4.5","effort":"low","cost_usd":0.005948,"raw_usage":{"total_tokens":1596,"prompt_tokens":811,"num_sources_used":0,"completion_tokens":99,"cost_in_usd_ticks":59480000,"prompt_tokens_details":{"text_tokens":811,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":686,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":811,"tokens_out":99,"duration_ms":5503,"temperature":1.0,"reasoning_tokens":686,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T05:49:18.156609+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train an otherwise identical agent with only exact-match answer reward on the same long-video suites; if Pass@1 and Pass@4 gains disappear (or reverse) on videos longer than ten minutes, the claim that the VETL evidence reward is what drives the improvement collapses.","supporting_citations":[],"review_version":1}