💻 computer science

REVISOR: Beyond Textual Reflection, Towards Multimodal Introspective Reasoning in Long-Form Video Understanding

The paper introduces REVISOR, a novel framework that enhances long-form video understanding by enabling multimodal introspective reasoning across both textual and visual modalities, supported by a Dual Attribution Decoupled Reward mechanism to ensure causal alignment between reasoning and video evidence.

Original authors: Jiaze Li, Hao Yin, Wenhui Tan, Jingyang Chen, Boshen Xu, Yuxun Qu, Yijing Chen, Jianzhong Ju, Zhenbo Luo, Jian Luan

Published 2026-05-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jiaze Li, Hao Yin, Wenhui Tan, Jingyang Chen, Boshen Xu, Yuxun Qu, Yijing Chen, Jianzhong Ju, Zhenbo Luo, Jian Luan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Problem: The "Fast Thinker" Who Skips the Details

Imagine you are trying to solve a mystery in a 30-minute movie. You are a detective (the AI) who usually solves crimes by reading the script (the text).

In the past, when AI tried to solve long video puzzles, it used a method called "Textual Reflection." This is like a detective who watches the movie, then closes their eyes and just thinks, "Hmm, what did I just see? Let me re-read my notes."

The paper argues that this approach fails for long videos because:

Videos are chaotic: Unlike a still photo, a video is full of moving parts, fast actions, and changing scenes. Just "thinking about the text" isn't enough to catch a tiny detail that happened two minutes ago.
The AI gets lost: Without looking back at the actual video, the AI often hallucinates (makes things up) or repeats the same mistake, like a detective who keeps guessing the same wrong suspect because they forgot to check the evidence.

The Solution: REVISOR (The Detective with a Rewind Button)

The authors created a new framework called REVISOR. Think of REVISOR not just as a detective, but as a detective with a magic remote control.

Instead of just closing their eyes to think, REVISOR follows a two-step process:

The First Pass (Initial Inference): The AI watches the video quickly (like a fast-forward scan) and makes a guess. But crucially, it also says, "Wait, I'm not sure about the part between minute 2 and minute 4. I need to look there again."
The Second Pass (Visual Reflection): The AI uses its "magic remote" to rewind and zoom in specifically on that 2-minute segment. It grabs high-quality, slow-motion frames of just that part. Then, it combines its original guess with this fresh, detailed visual evidence to correct its answer.

The Analogy:

Old Method: You read a recipe, realize you might have missed an ingredient, and try to remember what the chef said while you were cooking. You guess.
REVISOR: You read the recipe, realize you missed an ingredient, pause the cooking show, rewind to the exact second the chef added the spice, watch it closely, and then finish cooking.

The Secret Sauce: The "Causal Reward" (DADR)

Training an AI to do this is tricky. If you just tell the AI, "Get the right answer," it might cheat. It might guess the right answer but point to the wrong part of the video as its "evidence."

To fix this, the authors invented a special training rule called DADR (Dual Attribution Decoupled Reward).

The Analogy:
Imagine a teacher grading a student's test.

Old Grading: The teacher only checks if the final answer is correct. If the student gets "42" right, they get an A, even if they wrote "Because the moon is made of cheese" as their reasoning.
DADR Grading: The teacher gives two grades:
1. Is the final answer correct?
2. Did the student actually look at the right part of the textbook to get that answer?

If the student gets the right answer but points to the wrong page, they get a bad grade. This forces the AI to learn that finding the right video segment is just as important as getting the right answer.

What They Found

The paper tested this on four major video understanding benchmarks (like VideoMME and LongVideoBench).

The Result: REVISOR consistently beat the previous best models. It improved accuracy by about 2% to 4% on difficult, long videos.
The Key Insight: The paper found that for long videos, looking back at the video (visual reflection) is much more important than thinking harder about the words (textual reflection). In fact, forcing the AI to write more text reasoning actually made it perform worse. The AI learned to stop over-thinking the words and start focusing on re-watching the critical moments.

Summary

REVISOR is a new way for AI to understand long videos. Instead of just "thinking" about what it saw, it learns to pause, rewind, and zoom in on the specific moments that matter. By training the AI with a special rule that punishes it for looking at the wrong parts of the video, the system becomes much better at solving complex visual puzzles without needing to be retrained from scratch.

Technical Summary: REVISOR

Problem Statement

Current self-reflection mechanisms in Multimodal Large Language Models (MLLMs) predominantly rely on purely textual re-thinking processes. While effective for static image understanding and general multimodal tasks, these text-only reflection strategies exhibit significant limitations when applied to long-form video understanding. The authors identify two fundamental causes for this performance degradation:

Insufficiency of Text-Only Rethinking: Long-form videos contain richer, more dynamic visual inputs than static images. Rethinking only the textual reasoning trace is inadequate for correcting errors without explicitly re-evaluating the visual information.
Lack of Cross-Modal Interaction: Purely text-based reflection lacks the capability to interact with visual modalities during the reflection phase, preventing the model from integrating visual cues to refine its reasoning.

Empirical evidence presented in the paper demonstrates that applying standard text-based reflection methods (e.g., VL-Rethinker) to long-form video tasks often leads to performance degradation rather than improvement.

Methodology

The paper proposes REVISOR (REflective VIsual Segment Oriented Reasoning), a novel two-stage, tool-augmented multimodal reasoning framework designed to enhance long-form video understanding without requiring supplementary supervised fine-tuning (SFT) or external models.

1. Tool-Augmented Multimodal Reflection

REVISOR operates through a collaborative process between an MLLM and a visual toolbox:

Stage 1: Initial Inference and Visual Review Proposal: The MLLM performs an initial reasoning step on sparsely sampled video frames. Crucially, beyond generating a reasoning trace, the model is instructed to identify and output a specific temporal segment ( $S$ ) that is most relevant or uncertain regarding the conclusion.
Tool Invocation: Upon identifying segment $S$ , the framework invokes a Visual Toolbox to perform high-density re-sampling of frames within that specific temporal window. This provides fine-grained visual evidence ( $F_{review}$ ) without processing the entire video at high resolution.
Stage 2: Reflective Reasoning and Answer Refinement: The MLLM is re-invoked with the original question, the initial reasoning trace, and the newly acquired dense visual evidence. It integrates this information to iteratively refine its reasoning, validate hypotheses, and generate a final, more accurate answer.

2. Dual Attribution Decoupled Reward (DADR)

To train REVISOR effectively using Reinforcement Learning (specifically the GRPO algorithm), the authors introduce the Dual Attribution Decoupled Reward (DADR) mechanism. Standard reinforcement learning relying solely on final-answer verification fails to teach the model how to select the correct video segments for review. DADR addresses this by decomposing the reward signal into two components:

Final Answer Verification Reward ( $R_{final}$ ): Validates the correctness of the final refined answer against the ground truth.
Causal Segment Sufficiency Reward ( $R_{causal}$ ): This novel component enforces causal alignment between the model's reasoning and the selected video evidence. It provides a positive reward only if the model can derive the correct answer exclusively from the densely sampled visual evidence of the proposed segment $S$ .

The total reward is a weighted sum: $R(\tau) = \lambda_1 R_{final} + \lambda_2 R_{causal}$ . This mechanism explicitly encourages the model to focus on informative visual cues and discourages reliance on irrelevant content or spurious correlations.

Key Contributions

Diagnosis of Text-Reflection Limitations: The paper diagnoses the root cause of the poor performance of conventional text-based self-reflection in long-form video understanding, attributing it to the inability to re-evaluate dynamic visual inputs.
REVISOR Framework: The proposal of a two-stage reasoning framework that transforms text-based self-reflection into a tool-augmented multimodal process, enabling MLLMs to collaboratively construct introspective reflection across textual and visual modalities.
DADR Mechanism: The integration of the Dual Attribution Decoupled Reward into the GRPO training strategy, enabling MLLMs to learn how to accurately review question-relevant video segments during the reinforcement learning process.

Experimental Results

The authors evaluated REVISOR on four major long-form video understanding benchmarks: VideoMME, LongVideoBench, MLVU, and LVBench.

Performance Gains: Using Qwen2.5-VL-7B as the base model, REVISOR consistently improved average accuracy by approximately 2% across these benchmarks without additional SFT. Notably, it achieved a 2.8% improvement on the long-video subset of VideoMME and a 2.5% improvement on MLVU.
Comparison: REVISOR outperformed both the base model and text-based reflection methods (including VL-Rethinker and a text-reflection variant trained on video data). It also surpassed recent text-based reasoning approaches like Video-R1.
Temporal Grounding: In temporal video grounding tasks (Charades-STA and NExT-GQA), REVISOR achieved state-of-the-art results, outperforming previous SFT-based and RL-based methods.
Ablation Studies: Experiments confirmed that the DADR mechanism is critical; removing the Causal Segment Sufficiency Reward caused performance to drop below the base model. Furthermore, analysis showed that as training progressed, the model learned to reduce the length of textual reflection while increasing the precision of the video segments it reviewed, validating the hypothesis that visual re-examination is more critical than textual re-thinking for video tasks.

Significance and Claims

The paper claims that REVISOR represents a significant step forward in multimodal reasoning by shifting the paradigm from "textual reflection" to "multimodal introspective reasoning." The authors assert that for long-form video understanding, revisiting and refining visual information is more critical than rethinking the textual reasoning process.

By integrating the DADR mechanism, the framework enables MLLMs to learn the precise recall and utilization of salient visual cues during reinforcement learning. The work demonstrates that tool-augmented multimodal reflection can significantly enhance reasoning capabilities in complex, long-duration video scenarios without the need for external models or extensive supervised fine-tuning, offering a robust solution to the limitations of current text-centric reflection strategies.