- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes a small–large model collaboration framework for long-video understanding. A shared 4B model performs routing, temporal search, and local evidence verification, while a frozen larger model integrates the retained visual evidence in one final answering call. Search supervision is derived from teacher-assessed priorities under the same sparse observations available to the student, rather than directly targeting evidence boundaries that may be unobservable. The framework supports different inference budgets. With a frozen 27B reasoner, the highest-effort configuration reports 74.8% on VideoMME, 79.2% on MLVU, and 58.7% on LVBench, with ablations examining candidate priorities, verification, and search supervision.
Critical strengths
-
The strongest evidence is the improvement over the corresponding frozen reasoner, rather than comparisons across unrelated systems. The 4B+27B configuration improves VideoMME, MLVU, and LVBench by 6.0, 13.6, and 15.3 points, respectively. Improvements with a frozen 4B reasoner further support the value of the acquired evidence. These results demonstrate that a small model can contribute meaningfully beyond its standalone answering capability.
-
The method appropriately separates proposing promising intervals from establishing whether those intervals contain useful evidence. Top-ranked candidates outperform lower-ranked ones, and verification improves performance by 1.1–2.2 points on the evaluated deferred subsets. Preserving partial evidence without requiring the verifier to solve the entire question is also a sensible design for multi-evidence reasoning.
Critical weaknesses
-
The efficiency analysis uses parameter-weighted video-input sessions, excluding actual input lengths, continuation tokens, output generation, caching, and preprocessing. The appendix appropriately acknowledges these limitations. Nevertheless, even this proxy reaches 2.28×–2.77× the standalone baseline for Extra-High inference, while Medium costs 1.10×, 1.33×, and 2.19× across the three benchmarks. The paper therefore demonstrates a useful accuracy–resource trade-off, but does not establish superiority over alternative ways of allocating the same total computation, such as additional large-model observations or multi-window inference.
-
Appendix E.2 defines “evidence-missing” examples using the same teacher-derived priority profile that supplies Seeker supervision. A low teacher relevance score is not an independent measurement of whether decisive evidence is visually absent. Furthermore, Seeker retains multi-turn supervision while the localization model predicts a single interval, despite both being evaluated with one interval. Consequently, Figure 7 may partly reflect teacher-defined partitioning and differences in supervision, rather than solely the advantage of observation-aligned search targets.
Questions for authors
- How does the framework compare with dense or multi-window 27B inference under matched end-to-end latency or GPU-time budgets, including all search, verification, and generation costs?
- Can evidence visibility be independently annotated from the actual sparse frames, and can the search/localization comparison match supervised-token budgets and the number of target intervals?
- What happens when the Verifier rejects every candidate, and how frequently does it discard evidence that becomes useful only in combination with another interval?
4 comments
Response to W2
W2. Evidence visibility and matched search/localization supervision. We reran the ablation with visibility labels derived from the actual sparse frames, independently of the teacher priorities used to train Seeker. We first sampled each video globally at 2 fps with a 64-frame cap, then asked Qwen3.8-27B to answer using only the sampled frames near the annotated evidence interval. We labeled an example as evidence-missing when the model answered incorrectly from these sparse frames and correctly from denser local frames around the same interval.
For each training mixture, EL and Seeker used the same source QA examples, initialization, and training settings. We matched the supervision budget at one target interval per example for both models. At evaluation, both predicted one interval, and we matched the interval durations before passing them to the same answering pipeline.
| Method | Evidence-missing share | VideoMME | MLVU |
|---|---|---|---|
| EL | 0% | 54.1 | 52.8 |
| EL | 50% | 52.7 (−1.4) | 50.3 (−2.5) |
| Seeker | 0% | 53.6 | 53.4 |
| Seeker | 50% | 55.6 (+2.0) | 55.1 (+1.7) |
EL and Seeker perform similarly with 0% evidence-missing examples. Raising the share to 50% lowers EL accuracy on both benchmarks and improves Seeker accuracy. The original trend therefore holds with independently derived visibility labels and the same number of supervised intervals. We have replaced the original ablation in the main text with this more closely matched comparison and updated the appendix with the experimental protocol.
Response to W3
W3. What happens when the Verifier rejects every candidate, and how frequently does it discard evidence that becomes useful only in combination with another interval?
If the Verifier rejects every candidate, the frozen Reasoner answers from frames sampled uniformly across the full video under the same 64-frame budget.
Besides, we agree that verifying intervals independently can discard evidence that is useful only in combination. In the revised appendix, we analyze 692 Router-deferred VideoMME questions with at least one retained and one rejected interval. For the 247 questions answered incorrectly using the retained intervals alone, we test each rejected interval in isolation and then add it back to the retained set.
In 79 of the 692 questions (11.42%), at least one rejected interval led to a correct answer when combined with the retained evidence, even though Reasoner answered incorrectly with either input alone. This identifies a limitation of interval-wise verification. Evaluating candidate intervals in the context of the evidence already retained is a valuable direction for future work.
Thank you again for your valuable suggestions. Your suggestions helped us make the comparison between evidence search and evidence localization more rigorous and better isolate the effect of the supervision target under sparse observations. We would be appreciate to clarify any remaining points during the discussion period.
Follow-up on Our Responses
Dear Reviewer VscU,
Thank you once again for your thoughtful comments on our submission. As the discussion phase is approaching its end, we would appreciate your feedback on whether our responses have sufficiently addressed your concerns.
We have responded to each point by adding a GPU-time comparison with a denser 27B baseline, rerunning the search–localization experiment with independently assessed evidence visibility and matched interval supervision, and clarifying the fallback strategy while quantifying complementary evidence rejected by the Verifier. These additions are included in the revised manuscript.
If any concerns remain, we would be grateful for the opportunity to clarify them.
Thank you for your time and consideration.
Response to W1
We appreciate the reviewer’s valuable suggestion.
W1. How does the framework compare with dense or multi-window 27B inference under matched end-to-end latency or GPU-time budgets, including all search, verification, and generation costs?
To test under equivalent compute budget, we doubled the standalone 27B baseline’s uniform frame budget from 64 to 128 and compared it with our framework using the same frozen 27B Reasoner. We used concurrent inference in vLLM to keep both H200 GPUs fully utilized. The GPU time accounts for routing, search, verification, and final reasoning where applicable. The revised appendix reports the stage-wise breakdown.
Each cell shows accuracy (%) / GPU time (s per question).
Extra-High stays within 5% of the 128-frame baseline’s GPU time on every benchmark and improves accuracy by 2.0, 5.9, and 13.2 percentage points, respectively. Moreover, visual tokens are a major contributor to GPU memory use in video QA. Obviously, Increasing the baseline to 128 frames can raise peak GPU memory demand, whereas our framework scales test-time computation without requiring a larger peak GPU memory budget.
Relative to the 64-frame baseline, Extra-High uses 1.13–1.22× the measured GPU time, compared with 2.28–2.77× under the original parameter-weighted session proxy. One reason for the lower ratios is output generation. The 27B Reasoner generates a reasoning trace, and the Router, Seeker, and Verifier emit only short, structured outputs. The original proxy omitted this decoding cost, which increased the apparent cost of the small-model stages relative to standalone 27B inference.