Organizations: Institute of Computing Technology, Chinese Academy of Sciences, China · University of Chinese Academy of Sciences, China · College of Computer Science and Electronic Engineering, Hunan University, China · Peng Cheng Laboratory, China
The dense, temporal nature of video presents a profound challenge for automated analysis. Despite the use of powerful Vision-Language Models, prevailing methods for video understanding are limited by the inherent disconnect between reasoning and perception: they rely on static, pre-processed information and cannot actively seek raw evidence from video as their understanding evolves. To address this, we introduce LensWalk, a flexible agentic framework that empowers a Large Language Model reasoner to control its own visual observation actively. LensWalk establishes a tight reason-plan-observe loop where the agent dynamically specifies, at each step, the temporal scope and sampling density of the video it observes. Using a suite of versatile, Vision-Language Model based tools parameterized by these specifications, the agent can perform broad scans for cues, focus on specific segments for fact extraction, and stitch evidence from multiple moments for holistic verification. This design allows for progressive, on-demand evidence gathering that directly serves the agent's evolving chain of thought. Without requiring any model fine-tuning, LensWalk delivers substantial, plug-and-play performance gains on multiple model recipes, boosting their accuracy by over 5% on challenging long-video benchmarks like LVBench and Video-MME. Our analysis reveals that enabling an agent to control how it sees is key to unlocking more accurate, robust, and interpretable video reasoning.
Figures & tables
Figure 1: Comparison of video understanding paradigms: (a) Single Forward of model with one-off context selection. (b) Retrieval-Based Video Agent. (c) LensWalk , an agent framework that actively plans observations to fulfill subsequent reasoning, explores iteratively, and self-regulates its observation budget. (d) By actively planning observations only on video segments essential to its reasoning process, LensWalk simultaneously achieves high accuracy and exceptional efficiency.
Figure 2: Illustration of LensWalk’s reasoning-scheduled active observation on a real trace. The LLM reasoner alternates Scan Search and Segment Focus to seek question-relevant evidence, initially misclassifies a mole as the queried water vole and records conflicting memories in the subject table (red text), then invokes Stitch Verify to inspect key segments, resolve the contradiction, and arrive at the correct answer.
Methods (Reasoner/Observer)
LVBench
LongVideoBench
Video MME
EgoSchema
Overall
(Val) Long
Long (w/o sub)
Val
Commercial VLMs
Gemini-1.5-Pro [ 42 ]
33.1
58.6
67.4
-
Gemini-2.0-Flash [ 14 ]
48.3
45.7
63.0
71.2
GPT-4o [ 18 ]
48.9
60.9
65.3
70.4
GPT-4.1 [ 30 ]
51.9
64.6
63.1
72.2
Table 1: Comparison on long video benchmarks. For LensWalk, “(·/·)” denotes Rasoner/Observer; “( M )” means “( M / M )”.
Table 4
Observation Tools
Reasoning Coherence
VideoMME Long
Scan Search
Segment Focus
Stitch Verify
Timestamp Anchor
Subject Memory
✓
✓
65.4
✓
✓
68.1
✓
✓
66.8
✓
✓
✓
✓
69.7
✓
✓
✓
✓
69.4
Table 4: Ablation on the Observation Tools and Reasoning coherence modules of ours(o3/GPT-4.1) on Video-MME (long split).
Figure 3: Comparison of video understanding methods on accuracy and token efficiency. Abbrev. of Gemini : Gem.
Figure 4: Tool-call behavior of LensWalk across model recipes, visualized following [ 61 ] . We group behaviour into six strategy types and show their ratio (sector angle), average frame cost (dashed polygon), and accuracy (noted as score; sector radius), revealing adaptive allocation of more frames to harder queries and diverse exploration under uncertainty.
Appendix figures & tables21 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: System prompt used for the LensWalk Reasoner .
Figure 6: User prompt used for LensWalk Reasoner .
Figure 7: Prompt for the Observer in Scan Search tool.
Figure 8: Prompt for the Observer in Segment Focus tool.
Figure 9: Prompt for the Observer in Stitched Verify tool ( stitched_observer ).
Figure 10: Subject Table Memory–Update prompt and required JSON output.
Organization
Model
Release
Version
OpenAI
GPT-4.1 [ 30 ]
2025-04
gpt-4.1-2025-04-14
GPT-4.1-mini [ 30 ]
2025-04
gpt-4.1-mini-2025-04-14
o3 [ 31 ]
2025-04
o3-2025-04-16
GPT-5 [ 29 ]
2025-08
gpt-5-2025-08-07
Alibaba
Qwen3-235B-A22B [ 55 ]
2025-04
qwen3-235b-a22b-instruct
Qwen2.5-VL-72B [ 2 ]
2025-02
qwen2.5-vl-72b-instruct
Appendix
Table 5: Models used in our experiments.
Figure 11: JSON Tool Schema used for Reasoner of LensWalk (part 1).
Figure 12: JSON Tool Schema used for Reasoner of LensWalk (continued).
Method
Acc. (%)
Online Inference (s)
Offline Preprocessing (s)
Avg. Frames / Query
o3
57.1
38.9
0
256
LensWalk
68.6
190.35
0
290.3
DVD
74.2
153.3
2180.4
8202 (2 fps)
MR. Video
65.5
326.2
4135.2
9227 (2.25 fps)
VideoAgent
64.1
200.5
1131.3
4101 (1 fps)
Appendix
Table 6: Shared-backbone comparison on LVBench using o3 as the reasoner. We report accuracy, online inference time, offline preprocessing time, and average frames consumed per query. Retrieval-based agents require dense pre-processing of the full video, while LensWalk does not.
Benchmark
Split
Acc. (%)
Gain
Avg. Video Length (s)
Reasoning-Intensive
Avg. Steps
Avg. Frames Used
Video-MME
Short
82.3
+1.7
80.7
No
2.8
89.7
Medium
79.9
+8.3
515.9
No
4.2
233.0
Long
70.0
+6.9
2466.7
No
6.8
387.1
EgoSchema
—
74.8
+2.6
180.0
No
2.6
89.7
Video-MMMU
—
77.1
+9.7
506.2
Yes
4.8
178.4
MMVU
—
80.9
+2.0
51.4
Yes
3.1
58.5
Appendix
Table 7: Adaptive scaling of LensWalk (o3/GPT-4.1) across benchmarks with different video lengths and reasoning demands. Gains are relative to the single-pass GPT-4.1 baseline.
GPT-4.1 Observer (baseline: 63.1)
GPT-4.1-mini Observer (baseline: 59.4)
o3 Observer (baseline: 64.7)
Reasoner
Acc. (%)
Reasoner
Acc. (%)
Reasoner
Acc. (%)
GLM-4.7
68.3 (+5.2)
GLM-4.7
66.0 (+6.6)
o3
71.4 (+6.7)
DeepSeek-V3
66.4 (+3.3)
DeepSeek-V3
63.3 (+3.9)
—
—
Extracted Frames
65.0 (+1.9)
Extracted Frames
62.0 (+2.6)
Extracted Frames
65.5 (+0.8)
Appendix
Table 8: Additional model recipes and a static extracted-frames baseline on Video-MME (long split). Each group header shows the corresponding single-pass observer baseline. Extracted Frames feeds the union of frames visited by LensWalk under the same 256-frame budget into a single-pass VLM.
Figure 13: Example of Direct Inquiry behavior.
Figure 14: Example of Progressive Zoom-in behavior.
Figure 15: Example of Integrative Verify behavior.
Figure 16: Example of Strategic Reflection behavior.
Figure 17: Example of Scope Partitioning behavior.
Figure 18: Example of Static Repetition behavior. The notation “zoom in × n” indicates that several fine-grained but low-impact intermediate inspections are omitted from the visualization for brevity.
Figure 19: Example of Failure mode: Premature Conclusion .
Figure 20: Example of Failure mode: Evidence Dilution . The notation “zoom in × n” indicates that several fine-grained but low-impact intermediate inspections are omitted from the visualization for brevity.
Figure 21: Example of Failure mode: Persistent Ambiguity .
Video agentic models have advanced challenging video-language tasks. However, most agentic approaches still heavily rely on greedy parsing over densely sampled video frames, resulting in high computational cost. We present VideoSeek, a long-horizon video agent that leverages video logic flow to actively seek answer-critical evidence instead of exhaustively parsing the full video. This insight allows the model to use far fewer frames while maintaining, or even improving, its video understanding capability. VideoSeek operates in a think-act-observe loop with a well-designed toolkit for collecting multi-granular video observations. This design enables query-aware exploration over accumulated observations and supports practical video understanding and reasoning. Experiments on four challenging video understanding and reasoning benchmarks demonstrate that VideoSeek achieves strong accuracy while using far fewer frames than prior video agents and standalone LMMs. Notably, VideoSeek achieves a 10.2 absolute points improvement on LVBench over its base model, GPT-5, while using 93% fewer frames. Further analysis highlights the significance of leveraging video logic flow, strong reasoning capability, and the complementary roles of toolkit design.
Vision-Language Models (VLMs) typically rely on static initial frames for video reasoning, restricting their ability to incorporate essential dynamic information as the reasoning process evolves. Existing methods that augment Chain-of-Thought (CoT) with additional frame information often exhibit suboptimal CoT quality and lack the crucial ability to synthesize visual information for hypothetical or counterfactual scenarios. We introduce Act-to-See (Act2See), a novel framework that enables active visual perception by empowering VLMs to actively interleave video frames within text CoTs. Act2See is developed via Supervised Fine-Tuning (SFT) on a high-quality dataset of reasoning traces generated by a frontier VLM. These traces integrate active calls to either retrieve existing frames or generate new ones, and are rigorously verified against human-annotated CoTs to ensure quality. This approach cultivates an emergent capability: at inference time, the model actively determines when to search for or synthesize the necessary visual evidence. Act2See establishes new state-of-the-art results on challenging benchmarks, including VideoEspresso and ViTIB, and outperforms comparable or larger models on Video-MME, EgoNormia, and VCR-Bench, demonstrating an advancement in enabling VLMs with active visual perception for video reasoning.
Multimodal Large Language Models (MLLMs) have achieved remarkable progress on short video understanding yet remain limited on long videos due to the limited visual context window. Prevailing approaches rely on uniform frame sampling or recent coarse-to-fine agentic zooming, both of which struggle to localize sparse, decisive evidence in sufficiently long videos. We formulate long video understanding as a Sequential Evidence Acquisition (SEA) problem, in which an agent reads the video turn by turn along the temporal axis, deciding at each turn how fast to watch, what evidence to retain, when to revisit uncertain segments, and when to stop and answer. Inspired by this view, we propose VideoScout, a multi-turn reasoning agent that instantiates the SEA paradigm through adaptive reasoning pacing. Specifically, by dynamically controlling the viewing pace, VideoScout enables efficient traversal of long videos within a bounded visual context window, allowing the agent to access more video content while balancing content analysis depth with reading efficiency. To train VideoScout, we construct VideoScout-66K, a set of over 66K high-quality exploration turns from 10K answer-verified trajectories, and adopt a two-stage pipeline: cold-start supervised fine-tuning teaches the agent per-turn output format, while the Decoupled Clip and Dynamic sAmpling Policy Optimization (DAPO) algorithm performs trajectory-level reinforcement learning with a composite reward that jointly considers answer accuracy, output format compliance, and the temporal alignment between the agent's viewing progress and the teacher's answer timing measured by intersection-over-union (IoU). Extensive experiments on long video understanding and reasoning benchmarks demonstrate that our 7B model achieves strong performance compared with existing trained 7B agentic models.