-
Limited technical novelty; overstated claims. (a) The framework is a standard ReAct loop (Yao et al., 2022) over a toolbox of off-the-shelf model APIs; Eq. (1)–(2) are generic agent equations. Sec. 3.1 frames the problem as sequential decision making that "maximize[s] the information gain," yet there is no learned policy, the state s_t is never defined, and no information-gain quantity is computed or used—the formulation is decorative. (b) The "novel audio-based event localization strategy/algorithm" is, as far as Sec. 3.2 and App. H.5/H.7 show, two prompts asking Gemini-2.5-Flash to list audio segments or return timestamps for a query. There is no algorithmic component: no proposal generation, no scoring, no cross-modal verification, no handling of audio-visual asynchrony. Using audio or ASR timestamps to localize events for video QA has been done before (e.g., the Daily-Omni agent, XGC-AVis, and audio-guided methods the paper itself cites). (c) The claim of being "the first fully active perception agent for omnimodal understanding" is unconvincing: several cited works (XGC-AVis, the Daily-Omni agent, MAGNET; Active Video Perception and DeepVideoDiscovery for video) already use agents with tool calls and iterative evidence collection. The "static/fixed" vs. "active" distinction is never rigorously defined, and no experiment isolates it under controlled conditions (see point 3). (d) Claims that the method "solves the cross-modal alignment problem" (Contributions, Fig. 2, Sec. 3.1) are unsupported. The agent turns each modality into text and lets an LLM reconcile the texts—that sidesteps alignment rather than solving it. No experiment measures alignment (e.g., audio-visual synchrony, sound-source grounding, AV event localization accuracy). Phrases such as "paradigm shift," "genuine fine-grained understanding," and "mirroring human cognitive processes" are likewise unsupported.
-
The main comparisons are not apples-to-apples, so the reported gains cannot be credited to the proposed design. (a) OmniAgent combines several frontier systems—o3 + Gemini-2.5-Flash + Qwen3-VL + Qwen3-Omni—with up to 30 tool calls per question, yet it is compared with single forward passes of Qwen3-Omni-30B and Gemini-2.5-Flash. Gemini-2.5-Flash is itself one of OmniAgent's tools, so the system strictly contains its strongest baseline. "Without training" makes the result sound cheaper than it is: the method is training-free only because it uses much stronger and much more expensive proprietary models at inference time. (b) Table 4(a) shows that replacing all tools with Gemini-2.5 gives 83.3, above the default configuration, and Fig. 6 shows that with GPT-4o as the brain OmniAgent scores 68.7—below its own tool Gemini-2.5-Flash used alone (72.7) and below Qwen3-Omni-30B (72.08); with o4-mini the score is 75.9. Most of the gain therefore appears to come from o3's reasoning ability plus repeated calls to Gemini (an ensembling or test-time-compute effect), not from the proposed agent design. (c) Key controlled baselines are missing, so the contribution of "active perception" cannot be separated from "a strong reasoning LLM + strong perception APIs": o3 (or the same brain) given uniformly sampled frames plus the full ASR transcript and audio caption in one pass (a static pipeline with the same models); Gemini-2.5-Flash / Gemini-2.5-Pro with CoT and self-consistency or majority voting at matched API cost / number of calls; and the competing agents (XGC-AVis, Daily-Omni agent, DVD) re-run with the same brain (o3) and the same tool models, or OmniAgent run with their backbones. (d) Stronger and more recent end-to-end baselines are absent: Sec. 4.1 says the paper compares with "OpenAI o3," but no o3 row appears in Tables 1–3; Gemini-2.5-Pro is missing; Qwen3.5-Omni and Qwen3.8-Omni are cited in the Introduction (Team 2026a,b) but not compared. (e) Almost all baseline numbers are copied from prior reports (App. A), coming from different frame budgets, prompts, and possibly different evaluation protocols. Many entries are incomplete: Gemini-2.5-Flash* in Table 1 and Qwen3-Omni in Table 3 have only an average, and Table 2 has no agent baselines at all. DVD is a video-only agent evaluated only on Daily-Omni, with a modified FPS and a swapped captioner; App. A says DVD was tested only on Daily-Omni because of cost, while App. C says it was also run on "a randomly sampled subset" of the other benchmarks—yet those accuracies are not reported. (f) The headline claim of a "10%–20%" margin over leading open- and closed-source models (Abstract, Intro) is inconsistent with the tables: on OmniVideoBench the gain over Gemini-2.5-Flash is 6.7 points (59.10 vs. 52.40), and on WorldSense it is 4.7 points over video-SALMONN 2+ (61.2 vs. 56.5) and 7.2 points over Qwen3-Omni (61.2 vs. 54.0). The paper also never says whether the margin is absolute or relative.
-
The ablations are insufficient and partly contradict the narrative. (a) Reflection is presented as a core component—a "cross-modal consistency check" in Fig. 2, Fig. 3, and Sec. 3.3—but it is never ablated, and its implementation is described only through the system prompt in App. G.1; the same holds for the memory and re-planning modules in Fig. 3. (b) Table 4(b) covers only 3 of the 7 tools and only 4 configurations; ASR, Global Caption, Event List, and Global QA are never ablated. The text calls the drop from removing Audio QA "substantial," but it is 2.5 points (82.7 → 80.2). (c) The "Fixed-order" (66.7) and "Random-choose" (65.2) baselines in Fig. 6 are not specified—which order, how many calls, what answer synthesis—and both score below a single Gemini-2.5-Flash call (72.7), suggesting poorly constructed straw men rather than meaningful controls. (d) All ablations use only Daily-Omni, whose clips are 30s/60s, whereas the paper's main argument concerns localization in long videos; OmniVideoBench (up to 30 min) is the natural place for these ablations and none appear there. The effect of the step budget (max 30, average 6.7) is not studied either. (e) Event localization accuracy—the core claimed contribution—is never evaluated directly (e.g., temporal IoU / recall@k on an audio-visual event grounding benchmark), and there is no analysis of how localization errors propagate to final accuracy, even though App. E names error propagation as a key issue.
-
The behavioral analyses are partly circular. Finding 1 and Fig. 5/8 present the "audio first, then fine-grained video" pattern as evidence that "empirically validates the efficacy of our algorithmic design," but App. G.1 and G.2 explicitly instruct the brain to "Prefer to listen first, then look" and to "Use audio to find time and content first whenever possible, then inspect the corresponding visuals: from listen to look." The observed pattern is induced by the prompt, not emergent, which also contradicts the claim (Intro) that the decision process is "completely autonomous, decided by the LLM itself." In addition: Finding 2 contradicts itself—ELO is first described as returning "precise timestamps" from "low-cost" audio (Sec. 3.1–3.2), but Finding 2 calls ELO and VGA "computationally intensive tools [that] provide macro-level insights … unable to provide granular, fine-grained details." Fig. 7 shows that 41% of correctly answered WorldSense questions are solved with audio tools alone, suggesting a large share of the benchmark can be solved from transcripts or text shortcuts—this weakens rather than supports the claim of fine-grained cross-modal reasoning; evaluation on benchmarks designed against text shortcuts (e.g., AVUT, Yang et al. 2025c, cited in the paper) or on long audio-video benchmarks (e.g., LVOmniBench, Tao et al. 2026, also cited) would be more convincing. Finding 3 (GPT-4o "modality bias") is qualitative and not quantified.
-
The interpretability claims rest on post-hoc summaries. The caption of Fig. 1 admits: "Due to the unavailability of the thinking process, we use reflection in the final summary," and the system prompt (G.1) tells the model "Do NOT expose raw tool-call traces or long chain-of-thought." The "transparent, explainable" reasoning in Fig. 1, 4, and 9–12 is therefore a self-reported summary written after the fact by o3 and may not be faithful to the real decision process. All case studies are success cases (and arguably cherry-picked), with no failure or error analysis.
-
Statistical rigor and evaluation protocol are weak. There are no repeated runs, variance, or confidence intervals, although the authors acknowledge the agent is stochastic (App. E); temperature and o3 reasoning-effort settings are not reported. "Queries interrupted by connectivity issues were re-evaluated in a final testing phase" (App. A)—how many queries were affected, and were re-runs applied uniformly? Selective re-evaluation can bias results. When the 30-step limit is reached, o3 "synthesizes" an answer: how often does this happen, and how accurate are these forced answers? The answer parsing and failure-handling protocol for the multiple-choice format is not described.
No rebuttal for this review.
This article is a relatively large topic, and I raised some detailed questions, nice responce, I will increase his score.