ELV-Halluc: Benchmarking Semantic Aggregation Hallucinations in Long Video Understanding

TL;DR

This paper introduces ELV-Halluc, a benchmark for studying semantic aggregation hallucinations in long videos, revealing their causes and proposing mitigation strategies to improve video understanding models.

Contribution

It is the first benchmark dedicated to long-video hallucinations, specifically addressing Semantic Aggregation Hallucination (SAH) and proposing methods to reduce it.

Findings

SAH increases with semantic complexity

Models are more prone to SAH during rapid semantic changes

Positional encoding and DPO strategies reduce SAH

Abstract

Video multimodal large language models (Video-MLLMs) have achieved remarkable progress in video understanding. However, they remain vulnerable to hallucination-producing content inconsistent with or unrelated to video inputs. Previous video hallucination benchmarks primarily focus on short-videos. They attribute hallucinations to factors such as strong language priors, missing frames, or vision-language biases introduced by the visual encoder. While these causes indeed account for most hallucinations in short videos, they still oversimplify the cause of hallucinations. Sometimes, models generate incorrect outputs but with correct frame-level semantics. We refer to this type of hallucination as Semantic Aggregation Hallucination (SAH), which arises during the process of aggregating frame-level semantics into event-level semantic groups. Given that SAH becomes particularly critical in long videos due to increased semantic complexity across multiple events, it is essential to separate and thoroughly investigate the causes of this type of hallucination. To address the above issues, we introduce ELV-Halluc, the first benchmark dedicated to long-video hallucination, enabling a systematic investigation of SAH. Our experiments confirm the existence of SAH and show that it increases with semantic complexity. Additionally, we find that models are more prone to SAH on rapidly changing semantics. Moreover, we discuss potential approaches to mitigate SAH. We demonstrate that positional encoding strategy contributes to alleviating SAH, and further adopt DPO strategy to enhance the model's ability to distinguish semantics within and across events. To support this, we curate a dataset of 8K adversarial data pairs and achieve improvements on both ELV-Halluc and Video-MME, including a substantial 27.7% reduction in SAH ratio.

Tables

Table 1. Table 1: Statistical comparison with previous video hallucination benchmarks

	Videos	QAs per Video	Avg Video Length (s)
Videohallucer	948	1.89	85.6
EventHallusion	397	1.77	11.2
VidHalluc	5002	1.86	24.7
ARGUS	500	19	26.3
Ours	200	24	672.4

Table 2. Table 2: Main results on ELV-Halluc. Diff. denotes the gap between in-video and out-video accuracy. Note that the semi-automatic annotation pipeline uses Gemini for initial captioning, which may introduce bias; therefore, metrics for Gemini-2.5 Flash should not be directly compared with other models. All accuracies are shown as percentages.

Models	LLM size	Visual Details			Object			Action			Declarative Content			Avg Acc↑	Avg Diff.↓	SAH Ratio↓
		In.	Out.	Diff.	In.	Out.	Diff.	In.	Out.	Diff.	In.	Out.	Diff.
Open Source Models
InternVL3-1B	0.5B	8	11	3	8.7	11	2.3	8.7	12.5	3.8	11.3	8.3	-3	9.9	1.5	1.6
InternVL3-2B	1.5B	7	15.5	8.5	8.7	17.2	8.5	7.2	10.5	3.3	10	13	3	11.1	5.8	6.3
SmolVLM-2.2B	1.7B	0	0	0	3	5	2	0	0	0	0	0	0	1	0.5	0.5
Qwen2.5VL-3B	3B	2.2	10.5	8.3	7.7	13.8	6.1	5	8	3	6	6	0	7.4	4.3	4.5
LLaVA-Video-7B	7B	3.7	3.7	0	4.5	2.5	-2	3.7	3.2	-0.5	4	4	0	3.6	-0.6	-0.6
Video-chatgpt-7B	7B	2	2.5	0.5	2.5	1.7	-0.7	1.2	1.2	0	2.2	3.2	1	2.0	0.2	0.1
LLaVA-OV-7b	7B	8	13.2	5.2	9.5	13.7	4.2	8.7	10.7	2	7.7	7.5	-0.2	9.9	2.8	3.0
Qwen2.5VL-7b	7B	10.2	26	15.8	17.5	30.7	13.2	13	20.7	7.7	16.8	10.5	-6.3	18.1	7.6	8.8
InternVL3-8B	7B	12.5	19.5	7.0	14.5	19.5	5.0	13.5	20.5	7.0	12.8	17.7	4.9	16.3	5.9	6.8
InternVL3-14B	14B	17.5	24.5	7.0	22.8	24.5	1.7	16.3	17.7	1.4	15.2	15.5	0.3	19.2	2.6	3.1
Qwen2.5VL-32B	32B	16.5	24.5	8.0	21.7	24.5	2.8	17.2	15.0	-2.2	15.2	7.2	-8	17.7	0.1	0.2
InternVL3-38B	32B	25.3	29	3.7	24.2	28	3.8	24	30	6	24.5	24.2	-0.3	26.1	3.3	4.3
Qwen2.5VL-72B	72B	24	35.5	11.5	35.7	41.5	5.8	27.8	32.3	4.5	32.3	27	-5.3	32.0	4.1	5.8
InternVL3-78B	72B	25	31.2	6.2	32	36.5	4.5	28.5	31.2	2.7	24.2	26.5	2.3	29.3	3.9	5.4
Closed Source Models
GPT-4o	/	7.7	8.3	0.6	8	8.7	0.7	8.7	10.2	1.5	8.5	9.5	1	8.7	0.9	1.0
Gemini2.5-Flash	/	47	58	11	56.5	58.8	2.3	50.5	53.2	2.7	48.7	52	3.3	53.1	4.8	9.8

Table 3. Table 3: Different RoPE Strategies on ELV-Halluc. Bold values indicate the best performance. In. and Out. denote the average in-video and out-video accuracy, respectively.

Method	In. ↑	Out. ↑	SAH Ratio↓
vanillarope	0.94	2.75	1.82
tad_rope	0.44	2.62	2.18
mrope	1.12	2.06	0.95
videorope	1.19	2.06	0.88

Table 4. Table 4: Performance comparison of base model and DPO variants on ELV-Halluc and VideoMME benchmarks.

Model	ELV-Halluc		VideoMME
	Avg Acc↑	SAH Ratio↓	Short	Medium	Long	Avg↑
Qwen2.5vl-7B	15.9	8.3	72.7	61.7	51.3	61.9
+ invideo-4k	16.2	6.0 (-27.7%)	72.4	62.6	51.9	62.3
+ outvideo-4k	16.0	8.6 (+3.6%)	73.8	62.0	52.6	62.8
+ together-8k	16.4	8.4 (+1.2%)	73.4	62.1	51.8	62.4

Equations

$$\text{SAH Ratio}=\frac{\text{OutAcc}-\text{InAcc}}{1-\text{InAcc}}$$

Full text

Introduction

Video multimodal large models have demonstrated strong capabilities in visual understanding(Bai et al. 2025; Comanici et al. 2025; Zhu et al. 2025). However, a serious challenge still remains—the hallucination, where models generate content that is inconsistent with or even fabricated beyond the video content, thereby impacting the reliability of the models in practical applications. Many works(Rawal et al. 2025; Zhang et al. 2024a; Li, Im, and Fazli 2025; Kong et al. 2025; Wang et al. 2024c) have attempted to measure hallucination in video MLLMs, but they primarily focus on short videos ranging from a few seconds to tens of seconds, leaving hallucination issues in long-video contexts largely unexplored. They attribute hallucinations in Video-MLLMs to factors such as vision–language misalignment, poor frame quality, or suboptimal frame sampling strategies, which cause the model to rely on incomplete or inaccurate visual evidence. Alternatively, model may correctly perceive visual semantics but over-rely on strong language priors, disregarding visual input and producing incorrect content.

While above causes indeed covers large proportion of hallucinations, another cause has been overlooked in prior short video hallucination benchmarks: cases where the model correctly perceives and outputs accurate frame-level semantics but still produces incorrect content by misattributing semantics across event. For example, in Figure 1, the model attributes “Starbucks” to the first event, where the host is holding “some paper” while explaining the news. However, the mention of “Starbucks” actually corresponds to a later event in the video. In this case, while the perception of frame-level visual semantics is correct, the error arises from misaggregating information across temporal segments—incorrectly linking visual cues from one event to concepts from another. We refer to this phenomenon as Semantic Aggregation Hallucination (SAH).

In short-video scenarios, the impact of SAH is limited because frame-level semantics usually map directly to a single, self-contained event. As a result, logically consistent event semantics—especially those aligned with language priors—relatively rare (e.g.,“person”, “horse”, “riding”; it is unlikely for the model to hallucinate “horse riding a person,”). In contrast, long videos often contain multiple temporally extended, yet semantically coherent events, increasing the risk of misattributing concepts across events. As illustrated in Figure 1, this richer temporal structure amplifies the likelihood of SAH, where the model confuses when an event occurs, even if all the visual elements are correctly perceived.

To address the aforementioned limitations, we introduce ELV-Halluc, the first long video hallucination benchmark. As SAH is particularly prominent and challenging in long videos yet remains underexplored, ELV-Halluc is designed for studying SAH by quantifying semantic complexity through event-based videos and categorizing hallucination aspects based on semantic granularity, including visual details, action, object, and declarative content. To facilitate focused investigation, it adopts an adversarial triplet question pair design: (1) Ground Truth Question paired with In-Video Hallucinated Question, and (2) Ground Truth Question paired with Out-of-Video Hallucinated Question. We use the accuracy gap between in-video and out-of-video hallucinated question pairs to quantify a model’s sensitivity to semantic misalignment across events—a key aspect of SAH. Furthermore, we define the SAH Ratio as the proportion of such cases among all hallucinations, enabling systematic and interpretable analysis.

We conducted extensive experiments on ELV-Halluc, covering 14 open-source MLLMs and 2 closed-source models. Our findings confirm the existence of SAH and reveal that it does not necessarily correlate with overall hallucination rates. Notably, SAH becomes more severe as semantic complexity increases—such as with more events or denser frame sampling—and is more likely to occur in fine-grained, rapidly changing aspects (e.g., visual details rather than declarative content). Since SAH arises from incorrect aggregation of frame-level semantics across events, we show that strengthening the mapping between frames and events—such as through improved positional encodings—can help reduce its occurrence. We further adopt DPO(Rafailov et al. 2023) strategy that explicitly discourages the model’s preference for hallucinated semantics. Our contributions are summarized as follows:

• •

We introduce ELV-Halluc, the first long-video benchmark designed specifically to evaluate SAH.

• •

We conduct extensive experiments demonstrating that SAH positively correlates with semantic complexity and semantic variation rate(e.g., more SAH on visual details than declarative content). This relationship causes SAH to sometimes exhibit trends opposite to overall hallucination levels (e.g., when more frames are sampled).

• •

We validate the effectiveness of multimodal positional encoding in mitigating SAH and further adopt DPO strategy to reduce SAH. By curating 8K QA pairs with and without hallucinations, we achieve a maximum 27.7% reduction in SAH ratio while also improving overall performance (+0.9% on VideoMME).

Related Works

Video Understanding Benchmarks

Video Understanding benchmarks such as Video-MME(Fu et al. 2025) and MVBench(Li et al. 2024c) aim to provide a comprehensive evaluation of video understanding capabilities, covering multiple video lengths and diverse aspects of comprehension. Some benchmarks, however, focus on specific abilities of video models; for example, ETBench(Liu et al. 2024) emphasizes temporal localization and time-awareness, while Video-Holmes(Cheng et al. 2025) evaluates complex reasoning capabilities through QA pairs requiring strong reasoning skills.

In long video contexts, LVBench(Wang et al. 2024b) evaluates model comprehension for ultra-long videos exceeding one hour, while MLVU(Zhou et al. 2024) designs tasks with different requirements—such as holistic understanding, single-detail comprehension, and multi-detail reasoning—to assess long-video capabilities. Similarly, EgoSchema(Mangalam, Akshulakov, and Malik 2023) emphasizes evaluating model performance in egocentric video scenarios. However, hallucination—an important and relatively independent aspect of model reliability—remains largely underexplored in these general-purpose video understanding benchmarks.

Hallucination Evaluation in Video-MLLMs

Several prior efforts have aimed to construct hallucination-specific benchmarks. VideoHallucer(Wang et al. 2024c) categorizes hallucinations into two types: intrinsic, where the model outputs content inconsistent with the original video, and extrinsic, where correctness cannot be determined solely from the video. EventHallusion(Zhang et al. 2024a) further identifies two main sources of hallucination in Video-MLLMs: language priors and vision-language bias, and investigates these through QA designs involving rare events and misleading contexts. VidHalluc(Li, Im, and Fazli 2025) focuses on evaluating hallucinations in dynamic video segments and argues that the inductive bias inherent in visual encoders makes hallucinations more likely when processing semantically similar videos. ARGUS(Rawal et al. 2025), on the other hand, emphasizes hallucination evaluation in open-ended video captioning tasks.

Nevertheless, these existing benchmarks share two major limitations: (1) They primarily target short videos with relatively simple semantics. (2) They lack explicit discussion of Semantic Aggregation Hallucination (SAH), a critical challenge in long-video understanding.

ELV-Halluc

To address above issues, we propose ELV-Halluc, a Event based Long Video Hallucination benchmark and conduct in-depth analysis on SAH. Table 1 shows statistical comparison with other video hallucination benchmarks.

Event by event Video collection

Our benchmark is composed of Event-by-Event Videos. We define event-by-event Videos as videos that consist of multiple clearly separated events sharing the same overall topic. (e.g., a news broadcast with multiple news items).

Event-by-event Videos offer several advantages for establishing a long-video hallucination benchmark:

• •

Videos with clearly separated events can reduce captioning difficulty by isolating semantic units.

• •

Increase the likelihood of inducing SAH, as the semantics of event by event videos can be reorganized into multiple plausible yet incorrect descriptions.

• •

In event by event videos, the number of events can serve as an intuitive indicator of semantic complexity.

Finally, we manually collected 500 videos from YouTube. We removed overlapping samples with datasets such as YouCook2 to prevent potential data leakage.

Semi-automated Caption Pipeline

As shown in Figure 2, we adopt a three stage semi-automated caption pipeline to reduce human effort while ensuring annotation quality.

Video Quality Recheck

To reduce annotator disagreement on the Event-by-Event concept, we conducted a quality recheck. Annotators retained videos with 2–10 clearly distinguishable events and provided a keyword summarizing each video’s core event (e.g., scoring moments in basketball, news coverage in broadcasts). A total of 348 videos were retained, each reviewed by at least two annotators.

Automated Caption Generation with Gemini

We used Gemini-2.5 Flash to generate initial captions. Gemini was prompted to segment videos based on annotated keywords, exclude transitional or non-essential parts, and produce detailed captions for each identified event.

Human Verification and Refinement

Annotators are asked to revise the gemini generated captions through the following steps:

1. Correcting inaccurate time ranges;
2. Fixing factual errors in the captions;
3. Removing redundant segments (e.g., introductions, summaries, transition parts);
4. Adding missing events and manually annotating captions.

Following this semi-automated process, we obtained 348 high-quality Event-by-Event videos with human-refined ground truth captions, ensuring accuracy while substantially reducing manual annotation costs.

Hallucinate QAs curation

We design adversarial question pairs for better hallucination evaluating. A model should be able to both chose correct caption and reject hallucinated captions. Based on this principle, we use GPT-4o to modify ground truth captions by introducing hallucination elements. Each modified caption is paired with its ground truth to form a Question pair, and the pair is considered correct only if model answers both questions correctly.

Our modifications specifically target the semantics of four aspects: Visual details: attributes such as color, shape, size, patterns, spatial relationships, or on-screen text (OCR). Action: denotes the key activity or motion being performed. Object: refers to humans or physical objects mentioned in the caption. Declarative content: descriptive or propositional statements summarizing a situation, asserting an outcome, or conveying a belief or result, rather than concrete actions or events. (e.g., “Team A is leading,” “The match seems intense,”)

GPT-4o is instructed to modify captions by altering one of these aspects. To further investigate SAH, we design two modification strategies:

In-video modification: GPT replaces an object in the ground truth caption with an object drawn from another event within the same video.

Out-video modification: GPT replaces an object in the ground truth caption with a fabricated object that does not appear in any captions from the video.

Captions after modification must remain plausible and reasonable, such that correctness cannot be judged without watching the video. If a model is misled by an in-video hallucinated caption, all hallucination types could be responsible causes. In contrast, if the model is misled by an out-video hallucinated caption, SAH won’t serve as possible cause, since the hallucinated content does not exist in the video. Therefore, **subtracting the out-video mislead rate from the in-video mislead rate approximates the contribution of Semantic Aggregation Hallucination.**As shown in Figure 4, We use an object as an example to demonstrate the In-video and Out-video modifications.

After applying these modifications, we obtain 20072 hallucinated captions across 348 videos.

Hallucinated Caption Quality Check

We used GPT-4o to automatically recheck all modified captions, ensuring that in-video captions introduce the desired aspect change present in other events’ ground truths, while out-video captions introduce changes absent from all ground truths. Captions meeting above criteria were retained, yielding 348 Event-by-Event videos and 8,630 hallucinated caption pairs.

Final Benchmark and Evaluation Metrics

We select 200 videos from the original set of 348, leaving the remaining 148 videos as the training set for DPO. For each selected video, we choose 24 captions to construct binary QA pairs by appending the question prefix: “Is the following caption totally correct? Reply with ‘Yes’ or ‘No’ only.”

These QA pairs cover the aforementioned four aspects: visual details, objects, actions, and declarative content. Each aspect includes 6 questions, formed from 2 triplets drawn from different events within the same video. Each triplet consists of three captions: ground truth, in-video hallucinated, and out-of-video hallucinated. Examples of the final QA pairs are shown in Figure 5.

We form adversarial QA pairs by combining a ground-truth question with a hallucinated question, resulting in two pairs per triplet: (GT, In-Video Hallucination) (GT, Out-of-Video Hallucination) A pair is considered correct only if the model predicts “Yes” for the ground-truth question and “No” for the hallucinated question. Overall, the benchmark contains 4,800 binary QA pairs, which can be further grouped into 3,200 adversarial QA pairs. Figure 3 presents detailed statistics of ELV-Halluc, illustrating its diversity in video length, topics, and number of events.

Accuracy

We use Accuracy to evaluate the overall hallucination level of models in long-video scenarios. Specifically, we report the following metrics: In-Video Accuracy: Accuracy on QA pairs containing in-video hallucinations. Out-Video Accuracy: Accuracy on QA pairs containing out-of-video hallucinations.

SAH Ratio

We further propose the SAH Ratio to quantify the proportion of Semantic Aggregation Hallucination (SAH) among all hallucination errors. If a model achieves high accuracy on out-of-video hallucinations but significantly lower accuracy on in-video ones, it indicates difficulty in resolving semantic misalignment across events—the hallmark of SAH. Therefore, the accuracy gap reflects how prone the model is to confusing correct frame-level content with incorrect event-level attribution, making it a suitable proxy for SAH severity. Therefore, we use the SAH Ratio instead of the absolute difference between Out-Video and In-Video accuracy. This approach enables a more precise measurement of the relative severity of SAH while minimizing the influence of the model’s absolute performance level. Consequently, it facilitates targeted solutions specifically addressing SAH. The metric is computed as follows:

[TABLE]

where OutAcc and InAcc denote the accuracy on out-of-video and in-video hallucination pairs, respectively.

Experiments and discussions

We evaluate 14 open-source models ranging from 1B to 78B parameters, along with two closed-source MLLMs on ELV-Halluc. As shown in Table 2, the results demonstrate that current SOTA LLMs continue to face significant challenges with hallucination in long-video understanding(Note: The results of Gemini-2.5 Flash should not be directly compared with other models, as it was used to generate the initial captions, which may introduce bias). The results also validate the existance of SAH, as most MLLMs exhibit substantially lower accuracy on in-video hallucinated captions than on out-of-video hallucinated captions. Among the open-source models, Qwen2.5-VL-32B achieves the lowest SAH Ratio, at only 0.2%.

SAH increase as the semantic complexity rises.

In Figure 6, we observe that Qwen2.5-VL-3B, Qwen2.5-VL-72B, InternVL3-8B, and Gemini2.5-Flash (representing different model sizes and families) exhibit a positive correlation between the SAH ratio and the number of video events, as a larger number of events typically introduces more complex semantics. Meanwhile, we find that the SAH ratio does not show any consistent relationship with video length, since in our event-by-event dataset, video length does not necessarily correspond to a higher number of events.

SAH Occurs More Frequently at Rapidly Changing Semantic

We found that model tend to exhibit SAH more frequently on semantics that change rapidly over time. During modification, we define four aspects—visual details, action, object, and declarative content—each representing a different level of semantic granularity. Among them, visual details vary the fastest, followed by actions, as objects often engage in multiple actions over time. Objects themselves change less frequently, while declarative content represents higher-level semantics that evolve the slowest.

Based on the results of 14 open-source models, we plot the SAH ratio for these four aspects (see Figure 7). Results reveal that Video-MLLMs exhibit the highest SAH ratio on visual details, followed by actions and objects, and the lowest on declarative content, suggesting that models are more prone to SAH on semantics with higher temporal variability.

Hallucination under Varying Frame Numbers and Model Sizes.

As shown in Figure 8, We use the Qwen2.5-VL and InternVL3 series models to investigate the relationship between the number of video frames sampled and the occurrence of hallucinations. For Qwen2.5-VL, we disable the dynamic resolution mechanism to ensure that the number of frames serves as the sole varying factor. For accuracy, we observe that for most models, increasing the number of frames generally leads to higher overall hallucination accuracy. For the SAH ratio, most models tend to exhibit higher values as the number of frames increases. We attribute above findings to the fact that an increased number of frames provides the model with richer and more complex semantic information. On one hand, this additional information reduces uncertainty about the overall video content, thereby improving the model’s robustness against overall hallucinations. On the other hand, the added complexity increases the likelihood of semantic mismatches, leading to more SAH. Interestingly, the Qwen2.5-VL-32B model exhibits a different trend. We hypothesize that this may be due to the reinforcement learning-based post-training of Qwen2.5-VL-32B-Instruct, which likely enhances its ability to aggregate visual semantics effectively.

Furthermore, we observe that larger language models generally achieve higher overall hallucination accuracy, suggesting improved robustness against global hallucinations. However, we do not observe a consistent correlation between model size and the SAH ratio, indicating that increasing model capacity does not necessarily mitigate semantic aggregation hallucinations.

Strategies for Mitigating SAH

Position Encoding Decease SAH

As SAH occur due to errors introduced during the semantic grouping process. We argue that a stronger RoPE mechanism could enhance the model’s ability to bind semantic relationships, thereby reducing errors that occur during grouping. Our experiments validate this hypothesis, as shown in Table 3. Specifically, we evaluate checkpoints based on Qwen2-VL, applying different positional encoding strategies: vanilla RoPE(Su et al. 2024), TAD-RoPE(Gao et al. 2024), m-RoPE(Wang et al. 2024a), and VideoRoPE(Wei et al. 2025). The results demonstrate that VideoRoPE achieves the lowest SAH ratio. However, the findings also indicate that stronger RoPE variants do not necessarily mitigate oveall hallucinations.

Mitigate SAH with DPO

Considering that SAH mainly arise from incorrect grouping of correct video semantics, applying a method to suppress the model’s attention to hallucinatory semantics should reduce SAH possibilities. Therefore, we adopt the Direct Preference Optimization(Rafailov et al. 2023) approach to further optimize the model.

We leverage the remaining 148 videos’ ground-truth and hallucinated caption of all events to construct the positive and negative response pairs required for DPO. Specifically, we use the following template to generate data pairs: “Please provide a detailed caption for this segment during mm:ss - mm:ss. Chosen: Ground truth caption Rejected:Hallucinated caption”

Finally, we create two separate datasets of 4k positive-negative pairs each: one using in-video hallucinated captions and the other using out-of-video hallucinated captions.

We use Qwen2.5-VL-7B as the base model and conduct three training settings:

1. In-video pairs,
2. Out-video pairs,
3. Combining above 2 types of pairs.

As shown in Table 4, applying DPO with in-video pairs yields the most notable improvement, reducing the SAH ratio from 8.3 to 6.0. This indicates that aligning the model’s preference toward correct event semantics within the same video is highly effective for mitigating semantic aggregation hallucinations. Additionally, the overall average accuracy on ELV-Halluc improves by 0.3 points, while general video understanding performance on VideoMME also improves slightly (+0.4), demonstrating that mitigating hallucination does not compromise general capability.

In contrast, DPO with out-of-video pairs reduces overall hallucination slightly but unexpectedly increases the SAH ratio. This suggests that optimizing the model to reject content entirely irrelevant to the video does not effectively improve SAH and may even bias the model toward over-reliance on language priors.

When combining in-video and out-of-video pairs (8k samples together), the model achieves a balanced trade-off: it retains most of the benefits of in-video optimization, while maintaining robustness against out-of-video hallucinations. However, the combined setting does not surpass in-video DPO in reducing SAH.

As illustrated in Figure 9, attention visualization reveals that after DPO with in-video pairs, the model significantly reduces its focus on incorrect yet semantically plausible regions, shown in blue areas indicating decreased attention weights. This suggests that DPO effectively reshapes the model’s internal preference distribution, promoting stronger grounding of responses in relevant visual semantics. Such alignment at the attention level provides a mechanistic explanation for the observed improvement in SAH mitigation.

Conclusion and Limitations

In this work, we addressed the underexplored challenge of hallucination in long-video understanding by introducing ELV-Halluc, the first benchmark tailored to evaluate Semantic Aggregation Hallucination (SAH). We identified SAH as a distinct and increasingly prominent error type in semantically complex, multi-event videos. To enable comprehensive evaluation, we benchmarked 14 open-source Video-MLLMs (1B–78B) and two closed-source models (GPT-4o and Gemini 2.5 Flash). Our experiments revealed key SAH patterns and overall hallucination trends. To mitigate SAH, we proposed a DPO-based approach and curated an 8K-pair adversarial dataset, achieving a 27.7% reduction in SAH and a 0.9% gain on VideoMME.

Despite its contributions, our work has several limitations. First, although our semi-automated pipeline reduces manual effort, Gemini-generated captions may introduce bias, potentially inflating Gemini 2.5 Flash’s performance. Second, while event-based video construction improves control and diversity, it still differs from real-world long videos. Third, the dataset scale is limited due to the high annotation cost. Nonetheless, we believe our benchmark, analysis, and mitigation strategies lay a solid foundation for advancing reliable long-video understanding.