REACT: Evaluating Structural Distortions in Videos

• REACT is a frame-level reward model designed to evaluate structural distortions in generative videos by assigning point-wise scores and attribution labels.
• It employs a two-stage training framework with supervised fine-tuning for domain knowledge injection followed by reinforcement learning using pairwise rewards for aligning with human preferences.
• The model uses a dynamic inference mechanism to focus on diagnostically informative frames, providing interpretable evaluations that complement traditional visual quality and motion analyses.

REACT, in the context of generative video evaluation, denotes a frame-level reward model introduced in “Thinking with Frames: Generative Video Distortion Evaluation via Frame Reward Model” for structural distortions evaluation in generative videos. It is positioned against a specific limitation in contemporary text-to-video evaluation: existing video reward models typically assess visual quality, motion quality, and text alignment, but often overlook structural distortions such as abnormal object appearances and interactions, even though such distortions can degrade the overall quality of a generated video. REACT is described as assigning point-wise scores and attribution labels by reasoning over video frames, and as being supported by a human preference dataset, an additional data synthesis pipeline, a two-stage training framework, a dynamic inference mechanism, and a dedicated benchmark called REACT-Bench (Wang et al., 7 Jan 2026).

1. Problem scope and evaluative target

REACT is defined around a narrower evaluative target than generic video reward modeling. The motivating claim is that recent advances in video reward models and post-training strategies have improved text-to-video generation, but that the standard evaluative axes—visual quality, motion quality, and text alignment—do not adequately capture structural distortions (Wang et al., 7 Jan 2026). In the available description, these distortions are exemplified by abnormal object appearances and interactions.

This placement is important because it distinguishes REACT from reward models that are primarily optimized for perceptual fidelity or prompt compliance. The system is not introduced as a universal replacement for all video-quality assessment, but as a model designed specifically to recognize distortions that are structural in character. A plausible implication is that REACT is meant to operate as a complementary evaluator, augmenting reward models whose strengths lie elsewhere.

The paper abstract also makes explicit that REACT reasons over video frames rather than only over global video summaries (Wang et al., 7 Jan 2026). This frame-centric formulation suggests that structural failure modes are treated as temporally localized phenomena that may appear intermittently rather than uniformly across an entire clip.

2. Outputs and interpretability

REACT is described as assigning point-wise scores and attribution labels by reasoning over video frames, with a focus on recognizing distortions (Wang et al., 7 Jan 2026). Two aspects are central here.

First, the model is framed as a reward model, so its outputs are evaluative rather than generative. The reported output of point-wise scores indicates that assessment is not restricted to a single scalar judgment for the whole video. Second, the use of attribution labels indicates that REACT is intended to provide an interpretable account of where or how distortion is manifested.

The emphasis on attribution is notable because the abstract explicitly claims not only accurate quantitative evaluations but also interpretable attribution analysis (Wang et al., 7 Jan 2026). This suggests a dual role: REACT serves both as a scoring mechanism and as an error-analysis instrument for generative video systems. In that sense, the model is presented as evaluating failure structure, not merely ranking outputs.

3. Data construction and supervision

To support the model, the abstract states that the authors construct a large-scale human preference dataset, annotated based on a proposed taxonomy of structural distortions (Wang et al., 7 Jan 2026). It further states that they generate additional data using an efficient Chain-of-Thought (CoT) synthesis pipeline.

These claims establish three elements of the supervision strategy. The first is human preference annotation, indicating that the target behavior of the reward model is aligned with human judgments rather than defined only by heuristic rules. The second is a taxonomy of structural distortions, implying that the annotation space is organized around a structured error ontology rather than a single undifferentiated quality label. The third is the use of a CoT synthesis pipeline to expand the available data.

Because no further technical description is available in the supplied source beyond the abstract, the taxonomy itself, the annotation protocol, and the exact form of the synthesized data cannot be described more specifically without exceeding the evidence provided. The abstract nevertheless makes clear that REACT’s training signal is intended to combine human preference supervision with synthetic reasoning-oriented data augmentation (Wang et al., 7 Jan 2026).

4. Training framework and inference procedure

The abstract specifies a two-stage framework for training REACT. The first stage is supervised fine-tuning with masked loss for domain knowledge injection. The second stage is reinforcement learning with Group Relative Policy Optimization (GRPO) and pairwise rewards to enhance reasoning capability and align output scores with human preferences (Wang et al., 7 Jan 2026).

This description places REACT within a post-training regime that separates initial specialization from subsequent preference alignment. The stated role of the supervised stage is domain knowledge injection, while the reinforcement-learning stage is explicitly tied both to reasoning capability and to alignment with human preferences. The use of pairwise rewards indicates that comparative supervision remains important even after the supervised stage.

At inference time, the abstract reports a dynamic sampling mechanism introduced to focus on frames most likely to exhibit distortion (Wang et al., 7 Jan 2026). This design choice is consistent with the frame-level formulation of the problem: if structural distortions are sparse or localized, inference efficiency and evaluative sensitivity may both benefit from concentrating attention on likely failure frames. This suggests that REACT is intended not only to score videos but to do so selectively, prioritizing diagnostically informative temporal regions.

5. Benchmarking and reported experimental role

The paper introduces REACT-Bench, described as a benchmark for generative video distortion evaluation (Wang et al., 7 Jan 2026). In the abstract, REACT-Bench is presented alongside the model itself, implying that the work includes not only a method but also an evaluation substrate tailored to the same class of structural distortions.

The reported empirical claim is that REACT complements existing reward models in assessing structural distortion and achieves both accurate quantitative evaluations and interpretable attribution analysis (Wang et al., 7 Jan 2026). The language of complementarity is significant. It indicates that REACT is not framed as superseding reward models centered on visual quality, motion quality, or text alignment, but as addressing a missing evaluative axis.

No further experimental metrics, ablation results, benchmark composition, or quantitative tables are available in the supplied source text beyond this summary. Accordingly, one can state the claimed role of REACT in the evaluation ecosystem, but not the detailed empirical substantiation that would ordinarily accompany an encyclopedia treatment of a fully available paper.

6. Documentation limits and acronym ambiguity

A defining complication in documenting this REACT is that the supplied record explicitly states that the provided document is not the REACT paper described in the query, but instead a LaTeX rebuttal/template document containing formatting instructions and no technical content about the method, datasets, benchmark, architecture, or results beyond the abstract metadata (Wang et al., 7 Jan 2026). As a consequence, several items that would normally be central to an encyclopedia entry—such as the structural distortion taxonomy, the exact dataset design, the benchmark protocol, or the detailed architecture—cannot be reconstructed faithfully from the available source.

A second complication is terminological. REACT is an overloaded acronym across multiple research areas. In software engineering, it denotes a retrieval-augmented framework for commit message generation (Zhang et al., 2024). In out-of-distribution detection, ReAct refers to Rectified Activations (Sun et al., 2021). In language-model research, ReAct denotes Synergizing Reasoning and Acting in LLMs (Yao et al., 2022). In human-robot interaction, REACT names datasets for analyzing human reactions and evaluative feedback to robots over time (Candon et al., 2024).

This ambiguity matters because it can lead to category errors in citation and interpretation. In the present context, REACT refers specifically to a frame-level reward model for structural distortions in generative video, as introduced in “Thinking with Frames” (Wang et al., 7 Jan 2026). Any stronger technical account than that permitted by the abstract would require the actual paper text rather than the mismatched source currently associated with the record.