- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper studies long-video QA when each model call can ingest only a fixed number of pixels (16 frames at 224×224). MACF-Video splits a video into M temporal segments, each handled by a local agent (Qwen3-VL-8B with LoRA, weights shared) that compresses its segment and the question into K=32 continuous tokens through a two-layer MLP adapter. A coordinator with the same backbone reads the M×K tokens plus the question and produces the answer. Training has three stages: caption reconstruction to align the token space, image QA to make the tokens question-aware, and video QA with M=4 agents; inference uses M=6. On Video-MME, LongVideoBench, LVBench and MLVU-Test, the method improves over Qwen3-VL-8B by 4.5–7.7 points, over a re-implemented MapReduce by 9.7–18.6 points and over a re-implemented LatentMAS by 4.7–8.3 points (Table 1). The paper ablates the curriculum stages (Appendix D), the token count K (Table 2) and the agent count (Figure 4, Appendix G), and Appendix E compares the three communication protocols under a fixed 192-token message budget.
Critical strengths
- A controlled comparison of communication protocols. Appendix E fixes the backbone, partitioning, perception budget and a 192-token message budget for all three protocols. The learned latent tokens reach 60.4/56.8/40.2/49.2 on the four benchmarks, against 51.1/42.7/29.7/38.9 for LatentMAS and 44.5/36.9/30.3/32.4 for MapReduce. Both baselines are training-free and MACF-Video's tokens are trained, so the gap measures the trained message format together with its training. Within that scope it is the best-controlled experiment in the paper: once the work is split across agents, trained latent messages carry more task-relevant information than captions or truncated KV states of the same size. Figures 3 and 6 show examples where caption-based messages lead to a wrong answer and the latent tokens do not; in Figure 3 a caption binds the queried object to the wrong entity.
- The training recipe is ablated stage by stage. Removing stage 1 or stage 3 lowers Video-MME from 60.4 to 45.5 or 48.2 (Appendix D). This shows that the curriculum is required and makes the method reproducible as a recipe.
Critical weaknesses
- The single-model comparison matches neither total visual input nor training data, so the main claim is not established. Table 1 lists 16×224×224 for every budget-constrained row; for MACF-Video this is the per-agent budget. The MACF-Video rows use M=6 (Section 4.1; the M=6 entry of Table 7 equals the 56.8 in Table 1), so each question is answered from 96 frames, six times the input of the single-model rows. Appendix I motivates the per-call budget by deployment limits, and also states that 16×224×224 "serves as a stand-in for the effective per-forward visual context capacity rather than a hard model limit". Section 4.3(i) attributes the gain from more agents to wider temporal coverage. Figure 4 and Table 7 are consistent with this: with one agent, LongVideoBench accuracy is about 48.5 (read from Figure 4; Table 7 starts at M=2), below the 50.7 of the Qwen3-VL-8B row in Table 1, and it rises as agents, and therefore frames, are added. In addition, MACF-Video is fine-tuned with LoRA on LLaVA-Video-178K captions, Video-R1 image and video QA, and Molmo2 data, while the paper describes no fine-tuning of the Qwen3-VL-8B baseline. Two controls are needed: (a) Qwen3-VL-8B given 96 frames at 224×224, which fits well within its context; (b) Qwen3-VL-8B fine-tuned with the same LoRA setup on the stage-3 data, at 16 and at 96 frames. If these controls match MACF-Video, the contribution reduces to splitting one forward pass across several calls, which would then need memory and latency evidence. If MACF-Video is better, the claim becomes much stronger. My rating depends on these results.
- The re-implemented multi-agent baselines score below the single model they are built on. MapReduce (46.7/38.2/30.5/33.3) is below Qwen3-VL-8B (55.9/50.7/33.2/41.5) on all four benchmarks, and LatentMAS (55.7/48.5/33.2/42.3) ties or trails it on three. Both baselines see the same six segments. In the original MR. Video paper, the MapReduce design scores 60.8 on LVBench against 48.6 for Gemini-2.0-Flash, the VLM it builds on. Both baselines were modified: MR. Video's Gemini-2.0-Flash captioner and GPT-4o reasoner were both replaced by Qwen3-VL-8B, and the LatentMAS KV caches were concatenated and subsampled to 724 entries. The paper does not show that the re-implementations reproduce published results in their original setting, so the margins over these baselines cannot be attributed to the communication design. Section 4.2 also reports margins over LatentMAS of 9.3/14.1/10.5/10.3 points; these are the margins against the 192-token-capped LatentMAS of Appendix E, while against the Table 1 LatentMAS the margins are 4.7/8.3/7.0/6.9.
Questions for authors
Clarifications that could change my rating:
- What does Qwen3-VL-8B score on the four benchmarks with 96 frames at 224×224 (the total input of six agents)?
- What does Qwen3-VL-8B score after LoRA fine-tuning on the stage-3 data, at 16 and at 96 frames?
- Does the MapReduce re-implementation reproduce the MR. Video results in its original configuration? Why does it fall below the single model it wraps?
- What are the exact M=1 results on all four benchmarks?
Presentation fixes: report the total number of frames per question for each row of Table 1; make the LatentMAS margins in Section 4.2 consistent with Table 1; Figure 4 and Table 7 show LongVideoBench, while Section 4.3 says the agent-count study is on MLVU-Test.
Suggestions that do not affect my rating: report peak memory and wall-clock time against a single model with the same total frames; probe what the K tokens encode (segment position, question type); evaluate a multi-camera setting, which matches the distributed-device motivation better than temporal splits of one video.
5 comments
Response to Reviewer RfnX (2/2)
W2 & Q3: Re-implementation of multi-agent baselines
In our paper, we aim to compare communication representations and mechanisms (text / KV-cache vs. latent tokens) rather than the complete multi-agent workflows. We therefore aligned the baselines with MACF-Video's configuration. We acknowledge the differences from the original published settings and, as suggested, add experiments following those settings for further explanation.
MapReduce. We first explain why the published MR. Video result exceeds that of the single Gemini-2.0-Flash model, while our implementation with the Qwen3-VL-8B backbone performs worse than the single model.
Although MR. Video uses Gemini-2.0-Flash to generate captions, it uses GPT-4o as the reasoner in an iterative agentic loop. The loop also consumes far more input frames, exceeding 512 in our reproduction statistics. These differences may contribute to its higher score (60.8 vs. 48.6).
In our re-implementation, the controlled setting isolates the effect of text-based vs. latent-token communication, rather than comparing full MapReduce implementations. We apply MapReduce* with the same six-segment partition (16 frames per segment) and the same per-agent visual budget as MACF-Video (16*224*224). We also use Qwen3-VL-8B as both captioner and reasoner and run the workflow as a single pass, fully matching the MACF-Video setting. Although we used * to distinguish it from the original MapReduce method, this notation may not be appropriate. We therefore rename this baseline Text-Communication instead of MapReduce* in the revised paper (Section 4.1 and Table 1).
For reference, Table R3-3 summarizes the differences between MR. Video paper and our implementation, though they are not the focus of our work. Replacing only the reasoner (GPT-4o → Qwen3-VL-8B) lowers LVBench accuracy by 6.9 points (60.8 → 53.9). With a single-pass workflow, the text-based communication exceeds the reasoner's context length using the frame inputs in MR. Video.
Table R3-3. Comparison of the published MR. Video configuration and our re-implementation.
| Variant | Captioner | Reasoner | Workflow | Total Input Frames | LVBench |
|---|---|---|---|---|---|
| Published paper | Gemini-2.0-Flash | GPT-4o | Loop | >512 | 60.8 |
| Ablation 1 | Gemini-2.0-Flash | Qwen3-VL-8B | Loop | >512 | 53.9 |
| Ablation 2 | Gemini-2.0-Flash | Qwen3-VL-8B | single-pass | >512 | Out of contxt |
| Matched with MACF-Video (our paper) | Qwen3-VL-8B | Qwen3-VL-8B | Single-pass | 96 | 30.5 |
LatentMAS. Similarly, our implementation of LatentMAS uses 724 KV-cache entries to match MACF-Video's communication setting. Following the reviewer's suggestion, we additionally evaluate LatentMAS with its original full KV-cache communication. As shown in Table R3-4, our method achieves higher accuracy on all long-video benchmarks while using only 0.3% of the communication payload (1.50 vs. 481.77 MiB). We will also rename the original LatentMAS* baseline KVCache-Communication for clarity.
Table R3-4. Comparison with LatentMAS using full KV-caches.
| Method | Video-MME | Video-MME-Long | LongVideoBench | LongVideoBench-Long | LVBench | MLVU-Test | communication payload |
|---|---|---|---|---|---|---|---|
| LatentMAS (full KV-cache) | 61.0 | 52.3 | 56.0 | 48.9 | 38.3 | 45.8 | 481.77 MiB (3426 KV-Cache) |
| Ours (M=6) | 60.4 | 54.6 | 56.8 | 49.5 | 40.2 | 49.2 | 1.50 MiB |
Presentation Fixes
We have fixed the presentations in the revision as follows:
- Clarified the total frame number in the caption of Table 1.
- Corrected the LatentMAS margins in Section 4.2 to match Table 1.
- Corrected the benchmark name for the agent count ablation from MLVU-Test to LongVideoBench to match Figure 4 and Table 10 (Appendix G).
Dear Reviewer RfnX,
We would like to kindly follow up and ask whether our response has adequately addressed your concerns. If any questions or concerns remain, we would greatly appreciate the opportunity to discuss them further.
Thank you again for your time and thoughtful feedback.
Response to Reviewer RfnX (1/2)
W1 & Q1+Q2+Q4: Comparison at matched total frames and training data
As suggested, we fine-tune Qwen3-VL-8B using the same Stage-3 training data and LoRA configuration as ours. We evaluate the original and fine-tuned baselines with 16 and 96 input frames at 224×224 resolution, alongside our single-agent setting (M=1) in Table R3-1 (Table 4 in revised paper).
At 16 frames, our single-agent setting (M=1) is comparable to the baselines but generally slightly below them. This gap comes from compressing the whole video into only 32 communication tokens, a design built for efficient inter-agent communication rather than single-agent inference.
At 96 total frames, our multi-agent setting (M=6) achieves the best accuracy on all long-video benchmarks (Video-MME-Long, LongVideoBench-Long, LVBench, and MLVU-Test), This shows that distributed processing with inter-agent communication remains effective despite compression. As shown in Table R3-2, it also cuts inference latency by 45.6% (0.988 → 0.537 s) and peak memory excluding the backbone by 82.0% (2.95 → 0.53 GiB).