PaperlayerView exact revision on arXiv

Source-linked AI summary

Taming Hallucinations: Boosting MLLMs' Video Understanding via Counterfactual Video Generation

Zhe Huang, Hao Wen, Aiming Hao, Bingze Song, Meiqi Wu, Jiahong Wu, Xiangxiang Chu, Sheng Lu, Haoqian Wang

arXiv:2512.24271v1cs.CVcs.AI

TL;DR

MLLMs can rely on language priors over visual evidence, particularly on counterfactual videos, while counterfactual data collection and annotation are difficult to scale. The paper introduces DualityForge, DualityVidQA, and DNA-Train to synthesize paired counterfactual data and train with it. It reports a 24.0% gain on DualityVidQA-Test and gains across hallucination and general-purpose benchmarks.

  • Problem

    MLLMs’ reliance on language priors can produce visually ungrounded hallucinations on counterfactual videos, while collecting and annotating counterfactual data is costly.

    Source: §1 · p. 2, §3.1 · p. 4

  • Method

    DualityForge edits real videos using structured counterfactual context to generate QA pairs and original–edited video pairs, while DNA-Train uses SFT followed by pair-wise contrastive RL.

    Source: §1 · p. 2, §4.2 · p. 6

  • Results

    24.0% gains on DualityVidQA-Test accompany significant gains across hallucination and general-purpose benchmarks.

    Source: §1 · pp. 2–3

  • Takeaways & Limitations

    Experiments report reduced hallucinations and improved performance across hallucination and general-purpose video benchmarks.

    Source: §1 · p. 2, §6 · p. 8

    The stated divergence-based optimization articulates the desired sample properties but is not applied literally in the pipeline.

    Source: §3.1 · p. 4

Abstract

from arXiv · show

Multimodal Large Language Models (MLLMs) have made remarkable progress in video understanding. However, they suffer from a critical vulnerability: an over-reliance on language priors, which can lead to visual ungrounded hallucinations, especially when processing counterfactual videos that defy common sense. This limitation, stemming from the intrinsic data imbalance between text and video, is challenging to address due to the substantial cost of collecting and annotating counterfactual data. To address this, we introduce DualityForge, a novel counterfactual data synthesis framework that employs controllable, diffusion-based video editing to transform real-world videos into counterfactual scenarios. By embedding structured contextual information into the video editing and QA generation processes, the framework automatically produces high-quality QA pairs together with original-edited video pairs for contrastive training. Based on this, we build DualityVidQA, a large-scale video dataset designed to reduce MLLM hallucinations. In addition, to fully exploit the contrastive nature of our paired data, we propose Duality-Normalized Advantage Training (DNA-Train), a two-stage SFT-RL training regime where the RL phase applies pair-wise $\ell_1$ advantage normalization, thereby enabling a more stable and efficient policy optimization. Experiments on DualityVidQA-Test demonstrate that our method substantially reduces model hallucinations on counterfactual videos, yielding a relative improvement of 24.0% over the Qwen2.5-VL-7B baseline. Moreover, our approach achieves significant gains across both hallucination and general-purpose benchmarks, indicating strong generalization capability. We will open-source our dataset and code.

1. Introduction

The paper targets visual hallucinations caused by MLLMs’ reliance on language priors, especially for counterfactual videos. It introduces counterfactual data synthesis and paired-video training, reporting gains on hallucination and general-purpose benchmarks.

1. Introduction

  • MLLMs can favor language priors over visual evidence, producing plausible narratives that conflict with counterfactual video events.The issue is particularly severe when videos defy common sense, such as showing an object vanish or violate physical laws.Source: §1 · p. 2
  • Text-focused mitigation does not address the underlying text–video data imbalance, while counterfactual-video collection and reliable QA annotation are costly.The authors describe an automation paradox: models’ limited comprehension makes them unreliable annotators for the phenomena being targeted.Source: §1 · p. 2
  • DualityForge uses controllable diffusion-based editing to embed structured context in real videos and generate QA pairs, yielding original–edited video pairs for contrastive training.The structured context can specify event type and temporal location.Source: §1 · p. 2
  • DNA-Train combines SFT on real and counterfactual videos with RL on paired contrastive QA, using pair-wise ℓ1 advantage normalization during RL.The normalization is intended to balance learning across real–counterfactual pairs and stabilize updates.Source: §1 · p. 2
  • Experiments report significant improvements on hallucination detection and leading general-purpose video-understanding benchmarks.The cited benchmarks include EventHallusion, TempCompass, MVBench, TOMATO, and TVBench.Source: §1 · p. 2
  • 24.0% gains on DualityVidQA-Test accompany significant gains across hallucination and general benchmarks.The reported examples include TempCompass and MVBench.Source: §1 · p. 3

2. Related Works

Prior work addresses language-prior effects, scales video annotations, and applies RL to multimodal models, but the cited approaches leave challenges around counterfactual video coverage and visual grounding.

2.1. Language Prior in MLLMs

  • Contrastive decoding adds negative views and inference costs, is hyperparameter-sensitive, and does not update the base model; training-based data construction is costly.The authors position their automated synthesis framework as a scalable alternative for video.Source: §2.1 · p. 3

2.2. Video Understanding Datasets

  • VLM-generated video annotations scale supervision, but existing pipelines have limited coverage of rare events, long-range dependencies, and edited counter-commonsense scenarios.The passage also notes category and domain imbalance in these datasets.Source: §2.2 · p. 3

2.3. Visual Reinforcement Learning

  • Recent multimodal RL methods use varied reward and reasoning strategies, yet the cited approaches often optimize textual traces more than visual evidence.The paper emphasizes that video understanding requires distinguishing visually plausible from counterfactual cues and grounding decisions in evidence.Source: §2.3 · p. 3

3. DualityVidQA

DualityVidQA frames hallucination mitigation as a counterfactual-video synthesis problem: structured context guides video editing and QA generation, while paired original and edited clips support contrastive training.

3.1. Problem Formulation

  • DualityVidQA targets the language-prior bias that can make MLLMs disregard salient counterfactual events in videos.The formulation pairs anomaly-probing questions with videos to encourage visual grounding.Source: §3.1 · pp. 3–4
  • The conceptual objective seeks contexts where video-conditioned MLLM responses diverge from human responses while text-only responses remain comparatively aligned; the optimization is not applied literally.The authors identify scarce naturally occurring counterfactual videos and unreliable automated annotation as practical bottlenecks.Source: §3.1 · p. 4

3.2. DualityForge

  • Counterfactual contexts are organized as visual pixel distortions, semantic object-level inconsistencies, and commonsense violations, with a distinct editing pipeline for each.The figure maps these types to OpenCV pixel editing, VACE-based semantic editing, and FLUX-Kontext editing followed by VACE interpolation for commonsense anomalies.Source: Figure 2 · p. 3, §3.2 · p. 5
  • DualityForge embeds a predefined context in a source video, then reuses that context as a blueprint for an oracle caption and grounded QA pairs.Shared-question pairs require different correct answers for original and edited videos, using their contrast to test visual grounding.Source: §3.2 · p. 5

3. DualityVidQA

  • The dataset includes SFT, RL, and human-annotated test splits, with about 144K training QA pairs and 600 paired test samples.The test set covers four counter-commonsense scenarios; the dataset contains 81,274 clips totaling 100 hours.Source: Figure 3 · p. 4, §3.2 · p. 5

4. DNA-Train

DNA-Train combines SFT and RL to mitigate hallucinations while preserving real-world performance, using dual advantage normalization to balance gradient updates.

4.1. Supervised Fine-Tuning

  • SFT trains on balanced batches of original and counterfactual videos to recognize edited-video context while maintaining performance on real-world videos.The stated objective uses cross-entropy loss.Source: §4.1 · p. 5

4.2. Reinforcement Learning

  • The RL component builds on DAPO, which the paper selects for stable optimization over long trajectories and applies to groups of outputs for each QA pair.Source: §4.2 · p. 6
  • RL uses a dual-component reward: a binary correctness score for selecting the right answer and a format reward for following the prescribed reasoning structure.Source: §4.2 · p. 6
  • Duality-Normalized Advantage scales the real and counterfactual groups’ advantages to give each an equal contribution to gradient updates.The normalization addresses the initial imbalance in their learning signals.Source: §4.2 · p. 6
  • Under binary rewards, the learning signal peaks at group accuracy R = 0.5 and diminishes when tasks are trivial or impossible.Source: §4.2 · p. 6

5. Experiment

Evaluation uses short-video benchmarks and strict pairwise accuracy on DualityVidQA-Test. DNA-Train-7B improves hallucination detection while retaining general-video gains, and ablations favor paired data and DNA normalization.

5.1. Experimental Setup

  • DualityVidQA-Test counts a sample only when both original- and edited-video answers are correct; evaluation benchmarks are limited to videos under 30 seconds.The evaluation uses each benchmark’s standard frame-sampling protocol.Source: §5.1 · p. 7

5.2. Experimental Results

  • DNA-Train-7B scores 76.8% on DualityVid-Test, outperforms open-source models on EventHallusion, and improves over Qwen2.5-VL-7B across general benchmarks, competing with GPT-4o on MVBench and TVBench.Table 2 compares models across benchmarks.Source: §5.2 · pp. 7–8, Table 2 · p. 7

5.3. Ablation Studies

  • Paired data reaches 70.6 on DualityVid-Test and +1.8 average general-benchmark improvement; real-only and counterfactual-only training score 29.0 and 13.1, respectively.The paired-data baseline is 52.8 before the paired-data ablation result.Source: §5.3 · p. 8, Table 3 · p. 8
  • Against GRPO and DAPO from the same SFT model, DNA improves hallucination detection by 10.8 on average and beats DAPO on every general benchmark.Source: §5.3 · p. 8, Table 4 · p. 8
  • Across scales, DNA-Train improves both task types; the smallest model gains 25.9 hallucination-average points, while 72B receives only 20 RL steps under resource constraints.SFT provides a foundation and RL yields the largest boosts, particularly on DualityVid-Test.Source: §5.3 · p. 8, Table 5 · p. 8
  • On LLaVA-Next-Video, DNA-Train gains 42.0 on hallucination and 3.7 on general benchmarks, while both tested architectures beat their baselines across metrics.Source: §5.3 · p. 8, Table 6 · p. 8

6. Conclusion

The paper presents DualityForge and the 144K-sample DualityVidQA dataset as a route to training on paired real and counterfactual videos. It reports reduced hallucinations alongside stronger general video understanding benchmarks.

  • DualityForge uses controllable editing to build DualityVidQA, a 144K-sample contrastive dataset of paired real and counterfactual videos.Source: §6 · p. 8
  • DNA-Train uses two training stages and ℓ1-normalizes advantages per real-counterfactual pair during RL to encourage balanced training and visual grounding.Source: §6 · p. 8
  • The approach reduces hallucinations and boosts performance on general video understanding benchmarks.Source: §6 · pp. 8–9

A. Datset Detail

Video anomalies are organized into visual, semantic, and commonsense levels, ranging from pixel-level distortions to implausible events.

  • The taxonomy distinguishes pixel-wise visual distortions, scene-semantic temporal inconsistencies, and broader spatiotemporal or physical implausibilities.Examples include abnormal brightness, object disappearance, unnatural deformation, and implausible movement.Source: A. Datset Detail · p. 13

A.1. DualityForge

DualityForge draws source clips from OpenVid and Pexels, then applies different anomaly-generation methods for visual, semantic, and commonsense cases.

  • The source pool contains 61,591 OpenVid clips and 36,333 Pexels clips, totaling 97,924 videos.OpenVid sampling draws around 3,000 videos from each of its 20 most populated categories.Source: §A.1 · p. 13
  • OpenCV adds pixel-level perturbations, VACE edits localized semantic anomalies, and FLUX-Kontext with VACE interpolation creates physical, causal, material, and human-motion anomalies.Source: §A.1 · p. 13
  • The pipeline produces 135,168 anomaly videos for quality screening and uses around 40k GPU hours on NVIDIA H20 GPUs.Source: §A.1 · p. 13

A.2. DUALITYVIDQA

DualityVidQA is built through separate SFT and RL data curation, alongside a human-reviewed paired-video test set. Its RL examples use shared questions whose correct answers differ between real and edited videos.

Training Data Construction

  • The initial curation generated 200K QA pairs from 80K videos, then filtered about 30% of samples associated with zero reward and low-quality video.The authors attribute most zero-reward samples to failed edits without meaningful visual changes.Source: §A.2 · pp. 13–14

SFT data construction

  • The appendix reports question-type frequency statistics for the SFT QA data and question-type descriptions.The statistics are identified in Tables A.3 and A.4.Source: Table A.3 · p. 14, Table A.4 · p. 14
  • SFT QA generation uses dense captions and 16 question types, producing 100K QA pairs from 25K real and 25K edited videos in an 8:2 open-ended-to-multiple-choice ratio.GPT-5 and Gemini 2.5 Pro generate QA; three same-category examples are sampled from a 5,000-example LLaVA-Video pool as in-context references.Source: §A.2 · p. 14

RL data construction

  • The RL set contains 20K counterfactual QA pairs, using the same question and answer candidates for each real-edited pair while changing the correct answer.Gemini 2.5 Pro generates pairs from captions by identifying visual differences.Source: §A.2 · p. 14

Test Set

  • The test set contains 600 real-counterfactual video pairs with shared questions and options but different answers, grouped into four major categories after human review.Three annotators and three expert reviewers filtered and refined generated QA pairs for validity, clarity, and answerability.Source: §A.2 · p. 14

B. Derivation

The derivation considers binary rewards and expresses group accuracy as the fraction of correct responses, then relates reward dispersion to the magnitude and summed ℓ1 norm of advantages.

B. Derivation

  • Under binary rewards, group accuracy is defined as the fraction of correct responses and is used to derive reward standard deviation.The passage gives the variance expression in terms of group accuracy and group size.Source: B. Derivation · pp. 14–15
  • The derivation proceeds from reward standard deviation to the magnitude of the token-level advantages averaged within a response.The text identifies the average token-level advantage and then introduces advantage magnitude.Source: B. Derivation · p. 15
  • It also derives the sum of the ℓ1 norms of the averaged advantages across the group.This quantity is introduced as the group-level sum of ℓ1 norm.Source: B. Derivation · p. 15
Loading 2512.24271v1…