Beyond Sparse Rewards: A New Benchmark and Structure-Aware Graph Alignment for Micro-Drama Understanding¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/qyx1121/MDrama_SAGA
Area: Multimodal VLM
Keywords: micro-drama understanding, reinforcement learning, graph matching reward, structure-aware graph alignment, temporal causal reasoning
TL;DR¶
Addressing character identity confusion and temporal hallucinations caused by sparse rewards in hyper-dense micro-dramas, this paper introduces M-Drama—the first large-scale bilingual benchmark spanning 35K+ instances—and proposes SAGA, a structure-aware graph alignment reward that decouples semantic triplet matching and transitive closure temporal order over narrative graphs, substantially outperforming vanilla DAPO and SFT baselines on Qwen3-VL-8B.
Background & Motivation¶
Driven by the exponential growth of mobile video consumption, vertical micro-dramas have rapidly surged into a dominant global entertainment format. Unlike conventional movies or television series that follow gradual cinematic exposition, micro-dramas feature hyper-compressed narratives within 1-to-3 minute vertical episodes (9:16 aspect ratio). Each episode is engineered with rapid scene transitions, sudden character arrivals, abrupt emotional reversals, and suspenseful plot hooks designed to capture user engagement. Compressing intricate character networks, underlying motivations, and multi-step causal chains into such short durations imposes unprecedented cognitive demands on vision-language models (VLMs), requiring them to aggregate fine-grained visual details and perform coherent, long-horizon narrative deductions. However, existing video benchmarks—ranging from early clip-QA suites like MovieNet and DramaQA to recent long-form benchmarks like MovieChat and InfiniBench—remain exclusively centered on horizontally framed, slowly developing cinematic content, leaving the hyper-dense dynamics of micro-dramas largely unaddressed.
To unlock deep narrative reasoning in VLMs, recent research has embraced reinforcement learning (RL) post-training paradigms (such as DeepSeek-R1-style pipelines). Nonetheless, designing effective reward signals for open-ended video comprehension and long-form narrative generation presents formidable challenges. Standard lexical metrics (such as ROUGE-L and CIDEr) merely assess surface n-gram overlap, inadvertently incentivizing verbose yet ungrounded responses. Meanwhile, scalar feedback from LLM judges remains sparse and coarse-grained, lacking structural awareness of fine-grained multi-entity interactions. Consequently, when confronted with the rapid identity swaps and abrupt plot shifts inherent to micro-dramas, RL-tuned VLMs frequently suffer from severe hallucinations—conflating distinct character identities (e.g., swapping roles or misattributing familial ties) and distorting chronological event order.
To resolve these compounding issues, the authors argue that narrative reasoning must be supervised through explicit topological structures rather than flat token sequences. Intermediate descriptive outputs should be grounded into structured entity-action-time graphs where relational integrity and causal precedence can be objectively verified. Core idea: formalize micro-drama narrative comprehension as a heterogeneous graph alignment task, introducing SAGA—a graph-matching reward function that decouples semantic triplet matching from transitive closure temporal alignment to provide dense structural supervision for video RL—accompanied by M-Drama, the first large-scale bilingual micro-drama benchmark.
Method¶
Overall Architecture¶
The framework encompasses two interconnected contributions: the construction of the large-scale bilingual M-Drama benchmark and the SAGA reinforcement learning framework tailored for multimodal narrative models. To compel models to ground reasoning in concrete visual observations before drawing conclusions, the inference output is formulated into a tripartite structure: <caption>[Video Description]</caption> <think>[Thinking Process]</think> <answer>[Response]</answer>.
The SAGA architecture consists of an LLM-based Graph Builder (Qwen3-14B), a dual-branch graph alignment reward module, and a group-relative policy optimization engine (DAPO). Given the predicted intermediate caption and ground truth description, the Graph Builder first converts both into heterogeneous narrative graphs \(G = (V, E)\), consisting of four node types (Character, Event, Prop, Scene) and five relation types. SAGA then computes dense rewards along two decoupled dimensions: Semantic Triplet Matching aligns entity roles and event contexts using character name parsing adjustments and relation-weighted neighborhood fingerprints, producing a semantic F1 score \(F_1^{sem}\) via maximum bipartite matching. Concurrently, Structural Temporal Matching computes the transitive closure over narrative event chains to evaluate chronological fidelity, yielding a temporal F1 score \(F_1^{struct}\). These scores are blended into a dense intermediate caption reward \(R_{cap}\) and combined with the final answer reward \(R_{ans}\) under segment-specific masking within the DAPO optimization framework.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Input Drama Video + Task Prompt"] --> B["Tripartite Structured Generation<br/>caption + think + answer"]
B --> C["Heterogeneous Narrative Graph Construction<br/>Scene/Character/Prop/Event Nodes & Edges"]
C --> D["Semantic Triplet Matching<br/>Entity Adjustments + Fingerprints + Bipartite Matching"]
C --> E["Structural Temporal Matching<br/>NEXT_EVENT Transitive Closure + Partial Order Alignment"]
D --> F["Dense Structural Reward Integration<br/>Weighted Combination of Semantic & Temporal F1"]
E --> F
F --> G["Segment-Specific Masked Policy Optimization<br/>Decoupled Advantage Estimation & Gradient Updates"]
Key Designs¶
1. Heterogeneous Narrative Graph Construction: Abstracting continuous video descriptions into atomic story-beat topologies
To eliminate unstructured linguistic redundancy and make graph-matching computationally tractable, SAGA defines a strict ontology tailored to micro-dramas. Nodes \(V\) are categorized into Scene (spatiotemporal locations), Character (strictly formatted as Role(Name) to disambiguate actors and roles), Prop (narrative-critical artifacts), and Event. Each Event node features an atomic verb name alongside an unambiguous, full-sentence semantic_context. Directed edges \(E\) comprise four semantic relations (LOCATED_IN, AGENT_OF, TARGET_OF, HAS_PROP) and one structural relation (NEXT_EVENT). To ensure high-quality graph construction, the LLM Graph Builder enforces story-beat event aggregation: rapid, continuous micro-actions with identical intent (such as multiple verbal jabs during an argument) are merged into a single macro-event (e.g., "Confronts"). Furthermore, topological integrity rules mandate that all nodes are connected without orphans, and all events form an unbroken, acyclic chronological chain.
2. Semantic Triplet Matching: Cascaded identity resolution and neighborhood fingerprinting for global bipartite alignment
To overcome the fragility of raw text embeddings—which frequently trigger false positives on possessive variations (e.g., "John" vs. "John's mother") and false negatives on lexical expansions (e.g., "Anderson" vs. "Peter Anderson")—SAGA deploys a cascaded rule-based adjustment for character matching. Core names are parsed from the Role(Name) schema; possessive relations incur a heavy penalty \(\lambda_{poss} = 0.30\), word-level subsets receive an explicit perfect score of 1.0, and completely disjoint core names are forced to 0. Partial substring matches on Props and Scenes receive a containment penalty \(\lambda_{cont} = 0.50\). For Event nodes, recurring verbs in differing narrative contexts are disambiguated via neighborhood fingerprints: non-event neighbors are grouped by relation type and aligned via the Hungarian algorithm, with higher weights assigned to core agency relations (\(w_{agent} = 1.5\), \(w_{target} = 1.2\)) versus contextual relations (\(w_{loc} = 0.8\), \(w_{prop} = 0.5\)). The final event similarity linearly interpolates textual and structural signals:
with \(\alpha = 0.65\). Triplet-level alignment then performs maximum bipartite matching across ground-truth triplets \(T_{gt}\) and predicted triplets \(T_{pred}\) using the Hungarian algorithm, enforcing strict relation and node-type compatibility to compute a soft semantic F1 score:
3. Structural Temporal Matching: Transitive closure expansion for robust chronological and causal fidelity
To evaluate chronological sequencing without overly punishing the omission of minor transitional events, SAGA eschews rigid step-by-step matching of NEXT_EVENT edges in favor of transitive closure \(O(G)\). Given a linear narrative chain \(A \to B \to C\), the transitive closure extracts all valid reachability pairs: \(O(G) = \{(A,B), (A,C), (B,C)\}\). Leveraging the event mapping \(\phi\) established during semantic matching, the set of correctly predicted temporal pairs is defined as:
Structural temporal precision, recall, and F1 score are computed as:
This mathematical formulation gracefully grants partial credit when the global narrative trajectory is preserved despite minor event omissions, while strictly penalizing catastrophic temporal reversals and hallucinated event loops.
Loss & Training¶
The intermediate caption reward combines structural and semantic F1 scores: \(R_{cap} = \beta \cdot F_1^{struct} + (1 - \beta) \cdot F_1^{sem}\) with \(\beta = 0.30\). Within the DAPO optimization framework, segment-specific token masking is enforced. The caption loss indicator mask \(I_{cap}(y_{i,t})\) restricts gradient backpropagation exclusively to tokens inside the <caption>...</caption> tags, whereas \(I_{ans}\) spans the entire sequence to steer downstream reasoning. The overall objective minimized during policy optimization is:
where \(\gamma = 0.75\), and each component objective is formulated as:
Training starts with a 1-epoch cold-start SFT phase on the full M-Drama training set using Qwen3-VL-8B-Instruct (global batch size 64, learning rate \(1\times 10^{-5}\)). The first-epoch checkpoint is selected to retain optimal output entropy and response diversity. Subsequently, 6 candidate trajectories per instance are sampled, stratified into easy, middle, and hard tiers, and filtered via prioritized sampling (retaining 44% high-quality trajectories). The RL policy is trained for 1 epoch using AdamW in BF16 with learning rate \(1\times 10^{-6}\) and PPO clipping ratios \(\varepsilon_L = 0.20, \varepsilon_H = 0.28\).
Key Experimental Results¶
Main Results¶
Evaluation on the M-Drama benchmark test set (1,562 Multiple-Choice, 513 Open-Ended, and 1,036 Summary tasks) confirms that SAGA consistently outperforms all 8B-scale baselines and specialized caption rewards.
| Model | Thinking Mode | MC Pass@1 (%) | OE Pass@1 (%) | OE Pass@3 (%) | Summary (DeepSeek-V3.2) | Summary (GPT-5-mini) |
|---|---|---|---|---|---|---|
| Zero-shot (Small Size) | ||||||
| Qwen2.5-VL-7B | ✗ | 38.6 | 24.0 | 36.8 | 3.67 | 4.54 |
| Qwen3-VL-8B | ✓ | 56.5 | 33.3 | 47.2 | 4.18 | 4.81 |
| InternVL3.5-8B | ✓ | 52.9 | 25.6 | 36.3 | 3.96 | 4.77 |
| Keye-VL 1.5-8B | ✓ | 50.3 | 23.2 | 37.4 | 3.54 | 4.30 |
| MiMo-VL-7B | ✓ | 60.5 | 25.7 | 41.3 | 4.16 | 5.65 |
| Zero-shot (Middle & Flagship) | ||||||
| Qwen3-VL-32B | ✓ | 66.5 | 38.6 | 55.4 | 4.93 | 5.50 |
| Gemini-3-Flash | ✓ | 85.3 | 56.0 | – | 6.21 | 6.53 |
| GPT-5 | ✓ | 83.7 | 60.0 | – | 7.54 | 7.09 |
| Post-Training Baselines (8B Scale) | ||||||
| Qwen3-VL-8B-SFT (best epoch) | ✓ | 69.3 | 35.9 | 54.4 | 3.40 | 3.66 |
| Qwen3-VL-8B-DAPO (vanilla) | ✓ | 68.8 | 34.9 | 53.2 | 5.22 | 5.97 |
| Qwen3-VL-8B-DAPO w/ VideoCap-R1 | ✓ | 69.2 | 36.3 | 55.0 | 5.40 | 6.23 |
| Qwen3-VL-8B-DAPO w/ AutoDQ | ✓ | 69.0 | 34.5 | 51.5 | 5.15 | 5.83 |
| SAGA (Ours) | ✓ | 69.3 | 40.0 | 56.3 | 5.74 | 6.41 |
Furthermore, evaluating out-of-domain generalization across mainstream video benchmarks confirms that SAGA enhances specialized narrative reasoning without degrading broader multimodal capabilities:
| Model | Frames | ShortVid-Bench | Video-MME (w/o sub.) | Video-Holmes |
|---|---|---|---|---|
| Qwen2.5-VL-7B-Instruct | 192 | 69.7 | 65.7 | 44.4 |
| ARC-Hunyuan-Video-7B | 150 | 73.0 | 58.2 | 40.9 |
| Qwen3-VL-8B-Instruct | 192 | 70.6 | 68.4 | 44.8 |
| Qwen3-VL-8B-DAPO | 192 | 70.5 | 63.5 | 45.5 |
| SAGA (Ours) | 192 | 72.8 | 64.9 | 46.7 |
Ablation Study¶
The table below isolates the individual contributions of SAGA's semantic and structural reward components on M-Drama:
| Configuration | MC Pass@1 (%) | OE Pass@1 (%) | Summary (DeepSeek-V3.2) | Summary (GPT-5-mini) |
|---|---|---|---|---|
| Full SAGA Model | 69.3 | 40.0 | 5.74 | 6.41 |
| w/o structural reward (\(F_1^{struct}\)) | 69.7 | 35.9 | 5.62 | 6.34 |
| w/o semantic reward (\(F_1^{sem}\)) | 69.7 | 36.7 | 5.69 | 6.50 |
| w/o both (vanilla DAPO) | 68.8 | 34.9 | 5.22 | 5.97 |
Key Findings¶
- Complex generative reasoning is uniquely responsive to graph rewards: Removing either the semantic or structural reward causes a steep drop in OE Pass@1 (\(40.0\% \to 35.9\%\) or \(36.7\%\)), collapsing to \(34.9\%\) when both are omitted (a 5.1% absolute drop). In contrast, multiple-choice accuracy fluctuates minimally (\(68.8\% \sim 69.7\%\)), confirming that structural rewards primarily penalize nuanced narrative hallucinations that surface in generative tasks.
- Granular task gains mirror reward component formulations: Fine-grained task breakdowns reveal that the \(+14.3\%\) OE surge in Character Motivation (\(28.6\% \to 42.9\%\)) directly stems from semantic triplet matching resolving agent-target interactions. Meanwhile, gains on Narrative Analysis (\(+8.4\%\) MC, \(+8.1\%\) OE) and Logical Reasoning (\(+8.0\%\) OE) are propelled by transitive temporal chains enforcing strict causal sequences.
- Summarization improvements stem from factual accuracy and completeness: Decomposing 5-point summary dimensions reveals that SAGA's advantage over vanilla DAPO is driven by Accuracy (\(+0.22\) on GPT-5-mini / \(+0.27\) on DeepSeek-V3.2) and Completeness (\(+0.12 / +0.16\)), while Fluency remains universally saturated (\(\ge 4.6\)). GPT-5-mini score distributions show that SAGA increases top-tier summaries (\(\ge 7\)) from \(36.7\%\) to \(44.4\%\), while cutting poor summaries (\(\le 3\)) from \(11.7\%\) down to \(7.5\%\).
- Perceptual trade-offs in fine-grained scene tasks: SAGA incurs mild regressions on certain fine-grained visual MC categories (e.g., Key Object Interaction \(-4.3\%\), Spatiotemporal \(-3.6\%\)). Qualitative error inspection indicates that optimizing for macro event-level coherence occasionally reduces sensitivity to hyper-specific background nuances (e.g., predicting "Inside the pharmacy" instead of "dusty abandoned warehouse").
Highlights & Insights¶
- Discrete graph topology as an objective reward engine: Instead of gambling on fragile text embeddings or uncalibrated scalar judge scores, SAGA demonstrates that parsing intermediate reasoning into an ontological scene-character-prop-event graph enables transparent, mathematically bounded Hungarian bipartite matching for RL supervision.
- Transitive closure resolves graph matching rigidity: In long-form video reasoning, requiring exact consecutive edge alignment fails when non-essential intermediate actions are omitted; expanding sequential paths into reachability closures enables flexible yet mathematically strict causal verification.
- Computationally efficient graph builder scaling: Rigorous auditing reveals that a 14B model achieves near-flawless graph extraction (\(0.37\%\) node error rate, \(3.70\%\) edge error rate), and downstream policy gradient optimization using a 14B builder matches the performance of a 32B builder with no statistical difference (\(p = 0.35\)), drastically lowering deployment overhead.
- Micro-dramas as a stress test for multimodal intelligence: With \(6.32\) events/min and \(1.94\) characters/min (1.4–3.7× denser than traditional films), micro-dramas serve as an aggressive benchmark for diagnosing long-context visual forgetting, role swaps, and causal hallucinations in VLMs.
Limitations & Future Work¶
- Omission of acoustic and speech modalities: Micro-dramas rely extensively on fast-paced verbal sparring, voice-over exposition, and dramatic musical cues. The current pipeline operates strictly on visual frames, establishing a ceiling for audio-dependent comprehension.
- Granularity limits in spatial and object attribute modeling: The
LOCATED_INandTARGET_OFrelations currently operate at the macro-entity level, lacking micro-level attribute nodes (e.g., object color, exact room corner coordinates), which occasionally blunts fine-grained multiple-choice discrimination. - Dependence on closed-source models for pipeline initialization: Initial benchmark generation and golden test verification rely heavily on Gemini-2.5-Pro and GPT-5, introducing potential external reproducibility dependencies despite rigorous human verification audits.
- Evaluation scale confined to 8B backbone: Due to computational constraints, RL post-training was demonstrated primarily on Qwen3-VL-8B; scaling SAGA across larger backbones (32B/72B) remains an open exploration area.
Related Work & Insights¶
- Comparison with cinematic benchmarks (MovieNet / TVQA / MovieChat / InfiniBench): Conventional movie and drama benchmarks focus on horizontal cinematic layouts and slow-paced exposition. M-Drama fills a crucial industry void by formalizing vertical mobile dramas characterized by hyper-dense narrative turns and emotional cliffhangers.
- Comparison with video RL reward functions (VideoCap-R1 / AutoDQ): VideoCap-R1 computes cosine similarity across flat subject-attribute-action embeddings, failing on entity disambiguation and temporal inversions. AutoDQ relies on multi-pass atomic NLI checks, which incur massive computational latency without global topological awareness. SAGA bridges this gap through bipartite graph matching combined with transitive reachability closures.
- Transferability to embodied AI and sequential planning: The concept of parsing unstructured CoT or intermediate descriptions into dynamic causal graphs and applying transitive closure alignment can be seamlessly adapted to robot task planning, procedural instruction following, and surgical video analysis to prevent step inversions and tool misattributions.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Pioneering benchmark for the vertical micro-drama domain paired with an innovative structure-aware narrative graph reward function.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Rigorous evaluation across 8 open-source and 3 proprietary models, 3 external out-of-domain benchmarks, human alignment audits, and noise perturbation analyses.
- Writing Quality: ⭐⭐⭐⭐⭐ Exceptionally clear narrative, elegant formal mathematical definitions, and comprehensive qualitative case studies.
- Value: ⭐⭐⭐⭐⭐ Delivers both a foundational benchmark for short-video understanding and a robust blueprint for structured reinforcement learning in long-form multimodal reasoning.