DAR-R1: Dynamic Affective Reasoning Framework

• DAR-R1 is a two-stage multimodal framework that segments video emotions into structured phases with precise temporal boundaries and causal rationales.
• It combines cold-start supervised fine-tuning with Group Relative Policy Optimization to enhance segmentation accuracy, emotion classification, and explanation quality.
• The framework predicts when an emotion changes, what new emotion is induced, and why, outperforming similar models on dynamic affective reasoning benchmarks.

Searching arXiv for papers on DAR-R1 and related dynamic affective reasoning benchmarks. {"query": "\"DAR-R1\" OR \"Dynamic Affective Reasoning\" video emotion benchmark", "max_results": 10, "sort_by": "submittedDate", "sort_order": "descending"} {"query": "\"viewer-centric\" video emotion reasoning benchmark arXiv", "max_results": 10, "sort_by": "relevance", "sort_order": "descending"} DAR-R1 is a two-stage multimodal framework introduced for the Dynamic Affective Reasoning (DAR) benchmark, a viewer-centric formulation of video emotion analysis in which a model must determine when a viewer’s dominant emotion changes, what the new emotion is, and why the event sequence induces that emotion (Zhang et al., 11 Jul 2026). Rather than treating each video as a single labeled clip, DAR-R1 operates on a temporally structured target consisting of affective segments with boundaries, emotion labels, and causal rationales. Its design combines cold-start supervised fine-tuning with Group Relative Policy Optimization (GRPO), and its reported purpose is to improve emotional localization, segment-level emotion prediction, and causal reasoning over consecutive video events (Zhang et al., 11 Jul 2026).

1. Conceptual definition and problem setting

DAR-R1 is defined in the paper “Benchmarking Dynamic Affective Reasoning: A Viewer-Centric Video Emotion Dataset” as the benchmark’s principal two-stage training framework, built on top of an existing multimodal LLM rather than a new architecture (Zhang et al., 11 Jul 2026). The motivating claim is that video emotion analysis is usually formulated too coarsely: conventional systems assign one static label per clip, whereas DAR treats emotion as a temporally evolving, viewer-induced trajectory over causally meaningful events.

The benchmark’s conceptual basis is explicitly linked to Affective Events Theory (AET), under which current affect depends on prior stimuli and accumulates over time. In this formulation, the target is not the on-screen character’s internal state, but the observer’s induced emotion. The distinction is operationally important: a calm character can still induce anxiety in the viewer, and an injury on screen can induce empathic pain rather than merely describing a visible action (Zhang et al., 11 Jul 2026).

The paper formalizes the benchmark as structured generation from visual input $V$ and a task instruction $Z_{\text{task}}$ into a ground-truth sequence

$S_{\text{gt}} = (S_1^*, \ldots, S_t^*)$

where each segment is

$$S_i^* = (t_{i,\text{start}}^*, t_{i,\text{end}}^*, e_i, r_i).$$

Each $S_i^*$ contains a start time, end time, emotion label, and rationale. DAR-R1 is trained to emit this segmented emotional timeline directly (Zhang et al., 11 Jul 2026).

2. Dataset, annotation schema, and benchmark tasks

The DAR dataset contains 15,087 videos, 36,908 event-aligned affective segments, and 27 emotion categories, with a split of 13,646 training videos and 1,441 test videos (Zhang et al., 11 Jul 2026). The average video duration is 14.3s, the average segment duration is 5.8s, and the average rationale length is 118.8 words. The annotations are segment-level rather than video-level, and the paper states that the construction process enforces dense, temporally grounded, causally explicit reasoning chains.

The 27-category viewer-centric taxonomy is inherited from VCE and includes: Admiration, Amusement, Anger, Anxiety, Awe, Boredom, Calmness, Confusion, Craving, Disgust, Empathic Pain, Entrancement, Excitement, Fear, Horror, Interest, Joy, Nostalgia, Relief, Romance, Sadness, Satisfaction, Sexual Desire, Surprise, Awkwardness, Adoration, and Aesthetic Appreciation (Zhang et al., 11 Jul 2026).

DAR defines three coupled tasks. Affective Segmentation predicts the start and end times of emotional phases. Fine-grained Emotion Classification assigns one of the 27 emotions to each predicted segment. Affective Reasoning generates a causal explanation for each segment, grounded in visual evidence and consistent with preceding affective context. The benchmark evaluates these levels with Segment Count Accuracy, mIoU (%), Emo-Acc, and GPT-4o-based reasoning scores over visual grounding (VG), causal logic (CL), viewer centricity (VC), temporal consistency (TC), and answer consistency (AC) (Zhang et al., 11 Jul 2026).

The annotation schema is conceptually

$$(\text{start\_time}, \text{end\_time}, \text{emotion}, \text{reason}),$$

and the paper further describes a bottom-up reasoning logic:

$$\text{Visual Evidence} \rightarrow \text{Contextual Appraisal} \rightarrow \text{Emotion Trigger}.$$

This means that segment rationales are expected to identify visible evidence, appraise it in context, and infer the induced viewer emotion, often explaining the transition from $e_{i-1}$ to $e_i$ rather than describing only a local shot (Zhang et al., 11 Jul 2026).

3. Architecture and two-stage training pipeline

DAR-R1 uses Qwen2.5-VL-3B as its backbone in the main implementation. The vision encoder is frozen, while the LLM and aligner are fine-tuned (Zhang et al., 11 Jul 2026). The method does not introduce a separate temporal transformer, external event encoder, or runtime module such as Grounding DINO or Gemini. Instead, it relies on the backbone video MLLM to process the video and emit temporally structured textual outputs.

The pipeline is described as follows. At inference time, DAR-R1 receives a video $V$ and a prompt asking it to segment the video into emotion phases, assign a viewer emotion, and explain why. It outputs a structured list of segments containing start time, end time, emotion label, and reasoning text. During training, the generated output is parsed into segment tuples and optimized in two stages (Zhang et al., 11 Jul 2026).

The first stage is cold-start supervised fine-tuning (SFT). Each example contains visual input $Z_{\text{task}}$0, task instruction $Z_{\text{task}}$1, and ground-truth target sequence $Z_{\text{task}}$2. The model is optimized by standard autoregressive negative log-likelihood. The paper states: “We optimize the parameters $Z_{\text{task}}$3 by minimizing the standard negative log-likelihood over the ground-truth tokens.” This stage is intended to teach format compliance, task decomposition, and viewer-centric affective reasoning prior alignment, but the paper states that SFT alone often still lacks precise event boundary placement, deeply grounded reasoning, and strong logical coherence (Zhang et al., 11 Jul 2026).

The second stage is Group Relative Policy Optimization (GRPO). For each input query $Z_{\text{task}}$4, the model samples a group of outputs

$Z_{\text{task}}$5

The objective is given as

$Z_{\text{task}}$6

with

$Z_{\text{task}}$7

where

$Z_{\text{task}}$8

The advantage is therefore a z-scored relative reward within the sampled group. The paper explicitly positions this stage as improving event count accuracy, boundary precision, localization-classification consistency, and rationale quality (Zhang et al., 11 Jul 2026).

Training details reported for the main implementation are: 4 × H100 GPUs; SFT stage: 0.5 epochs, AdamW, learning rate $Z_{\text{task}}$9; GRPO stage: 1 epoch, learning rate $S_{\text{gt}} = (S_1^*, \ldots, S_t^*)$0 (Zhang et al., 11 Jul 2026).

4. Reward design and reasoning constraints

DAR-R1’s main training novelty lies in its task-specific reward decomposition rather than in a new RL algorithm. The total reward is

$S_{\text{gt}} = (S_1^*, \ldots, S_t^*)$1

with implementation weights

$S_{\text{gt}} = (S_1^*, \ldots, S_t^*)$2

The reward is designed to regularize structure, temporal segmentation, emotion correctness, and rationale quality jointly (Zhang et al., 11 Jul 2026).

The Structural Constraints Reward is

$S_{\text{gt}} = (S_1^*, \ldots, S_t^*)$3

with $S_{\text{gt}} = (S_1^*, \ldots, S_t^*)$4, $S_{\text{gt}} = (S_1^*, \ldots, S_t^*)$5, $S_{\text{gt}} = (S_1^*, \ldots, S_t^*)$6, and $S_{\text{gt}} = (S_1^*, \ldots, S_t^*)$7. The soft length penalty uses $S_{\text{gt}} = (S_1^*, \ldots, S_t^*)$8, $S_{\text{gt}} = (S_1^*, \ldots, S_t^*)$9, and $$S_i^* = (t_{i,\text{start}}^*, t_{i,\text{end}}^*, e_i, r_i).$$0. This reward is intended to stabilize generation, prevent malformed outputs, and discourage excessive verbosity (Zhang et al., 11 Jul 2026).

The Segment Count Reward is

$$S_i^* = (t_{i,\text{start}}^*, t_{i,\text{end}}^*, e_i, r_i).$$1

with $$S_i^* = (t_{i,\text{start}}^*, t_{i,\text{end}}^*, e_i, r_i).$$2. This penalizes over-segmentation and under-segmentation. The Temporal Segmentation Reward matches each predicted segment to the best ground-truth segment by maximum temporal IoU and computes

$$S_i^* = (t_{i,\text{start}}^*, t_{i,\text{end}}^*, e_i, r_i).$$3

with $$S_i^* = (t_{i,\text{start}}^*, t_{i,\text{end}}^*, e_i, r_i).$$4 and $$S_i^* = (t_{i,\text{start}}^*, t_{i,\text{end}}^*, e_i, r_i).$$5. The final $$S_i^* = (t_{i,\text{start}}^*, t_{i,\text{end}}^*, e_i, r_i).$$6 is the average across predicted segments. This formulation explicitly rewards not only overlap but also boundary snapping precision (Zhang et al., 11 Jul 2026).

The Emotion Accuracy Reward is

$$S_i^* = (t_{i,\text{start}}^*, t_{i,\text{end}}^*, e_i, r_i).$$7

Thus a label is credited only when the segment is sufficiently aligned temporally and the emotion label matches. The Reasoning Quality Reward uses a target-length score

$$S_i^* = (t_{i,\text{start}}^*, t_{i,\text{end}}^*, e_i, r_i).$$8

and

$$S_i^* = (t_{i,\text{start}}^*, t_{i,\text{end}}^*, e_i, r_i).$$9

with $S_i^*$0, $S_i^*$1, and $S_i^*$2. According to the paper, this discourages repetitive rationales and helps prevent token-budget instability during GRPO (Zhang et al., 11 Jul 2026).

Reasoning quality is also judged externally along five benchmark dimensions—visual grounding, causal logic, viewer centricity, temporal consistency, and answer consistency—and the judge prompt requires that a rationale be grounded in visible evidence, specific to the target interval, viewer-centric rather than character-centric, and coherent with previous segments if not the first (Zhang et al., 11 Jul 2026).

5. Empirical results, ablations, and generalization

The paper evaluates 10+ MLLMs, including emotion-focused systems such as Videmo, EmotionLlama, and AffectGPT, and general MLLMs such as Video-ChatGPT, PandaGPT, Video-LLaVA, InternVL-3.5-2B, InternVL-3.5-8B, Qwen2.5-VL-3B, Qwen2.5-VL-7B, Qwen3-VL-4B, and Qwen3-VL-8B (Zhang et al., 11 Jul 2026). The reported headline result is that DAR-R1, despite using a 3B backbone, establishes the strongest overall performance on the DAR test set.

The core comparison in Table 2 is summarized below.

Model	Localization / classification	Reasoning
Qwen2.5-VL-3B	SC-Acc 25.4, mIoU 41.7, Emo-Acc 15.0	Avg GPT-Score 2.3
DAR-SFT	SC-Acc 37.7, mIoU 47.6, Emo-Acc 25.8	Avg GPT-Score 3.0
DAR-R1	SC-Acc 41.5, mIoU 52.3, Emo-Acc 28.6	Avg GPT-Score 3.3

DAR-R1’s reasoning subscores are reported as VG 3.1, CL 3.8, VC 3.2, TC 3.2, and AC 3.1 (Zhang et al., 11 Jul 2026). The paper emphasizes that from the base Qwen2.5-VL-3B to DAR-SFT, segment-count accuracy improves by 12.3% and emotion accuracy by 10.8%. The RL stage then further improves SC-Acc 37.7 → 41.5, mIoU 47.6 → 52.3, Emo-Acc 25.8 → 28.6, and CL 3.0 → 3.8 (Zhang et al., 11 Jul 2026). This is presented as evidence that SFT primarily teaches output structure and task format, whereas GRPO refines precision and reasoning quality.

Human evaluation over 100 randomly sampled videos reports average scores of 0.7 for AffectGPT, 2.5 for Qwen2.5-VL-3B, 3.0 for Qwen3-VL-4B, 3.6 for DAR-SFT, and 4.2 for DAR-R1. DAR-R1 receives particularly strong human scores in causal logic 4.5, viewer centricity 4.2, and answer consistency 4.4 (Zhang et al., 11 Jul 2026).

Reward ablations support the task-specific design. Without $S_i^*$3, performance drops to SC-Acc 38.3, mIoU 49.7, Emo-Acc 26.5. Without $S_i^*$4, results are 39.4, 48.0, 27.0. Without $S_i^*$5, DAR-R1 retains relatively strong localization but emotion accuracy falls to 24.2. The paper further states that $S_i^*$6 and $S_i^*$7 are retained as stability rewards, and that removing the reasoning reward causes rationales to become too long and destabilizes GRPO (Zhang et al., 11 Jul 2026).

Out-of-domain evaluation on TSL temporal localization is reported as [email protected] 13.2, [email protected] 11.4, [email protected] 9.4, [email protected] 7.5, and [email protected] 6.7, which the paper interprets as evidence that the segmentation-and-reasoning formulation is not purely overfit to DAR (Zhang et al., 11 Jul 2026).

6. Interpretation, limitations, and scope

DAR-R1’s reported significance lies in reframing video emotion analysis as structured temporal reasoning rather than static clip classification. The system jointly predicts segmentation, emotion, and rationale, and the paper argues that affect should be modeled as a response to event sequence, narrative change, prior emotional context, and causal interpretation by the viewer (Zhang et al., 11 Jul 2026). A plausible implication is that DAR-R1 belongs as much to temporally grounded multimodal reasoning as to emotion recognition in the narrow sense.

The paper also makes clear that DAR-R1 does not resolve all open problems in affective modeling. Viewer-centric emotion is inherently subjective; despite strong quality control, reported annotation errors remain: boundary error $S_i^*$8: 2.4%, emotion-label disagreement: 3.1%, visual-grounding error: 2.2%, and non-viewer-centric rationale: 1.3% (Zhang et al., 11 Jul 2026). Temporal grounding is difficult even for strong MLLMs, reasoning evaluation partly relies on LLM-as-a-judge, and two-stage RL tuning is more expensive than plain SFT. The paper states that generalization beyond the benchmark is promising but not fully resolved.

A common misconception is that DAR-R1 denotes a generic “DAR” system or an unrelated R1-family safety or reasoning model. In arXiv usage, however, DAR is polysemous: it can denote Deontic Agentic Reasoning in legal/policy reasoning (Dou et al., 3 Jun 2026), Data Agnostic Researcher in autonomous database exploration (Vykhopen et al., 16 Dec 2025), or Diversity–Area Relationship in ecology (Ma, 2017). The specific label DAR-R1 is defined in the viewer-centric video-emotion benchmark paper and refers to the RL-enhanced multimodal framework described above (Zhang et al., 11 Jul 2026).

Within that scope, DAR-R1 is best understood as a benchmark-specific, reinforcement-learning-enhanced MLLM for generating a structured emotional timeline over video: it predicts when emotions shift, what the viewer feels, and why the transition occurs, using SFT for structural adaptation and GRPO for temporal precision and causal refinement (Zhang et al., 11 Jul 2026).