CounterVid: Counterfactual Video Generation for Mitigating Action and Temporal Hallucinations in Video-Language Models

AI summary

Overview

• Research area: Computer Vision / multimodal machine learning — specifically hallucination mitigation in video-language models (VLMs).
• Technical level: Intermediate. Readers should be comfortable with preference optimization (DPO), vision-language model fine-tuning, and diffusion-based image/video generation.
• Scope: The paper introduces a pipeline that synthesizes counterfactual action videos, packages them into a ~26,000-pair preference dataset (CounterVid), and uses a unified textual + visual DPO objective (MixDPO) to reduce action-recognition and temporal-ordering hallucinations in VLMs.

What This Paper Is About

Video-language models often hallucinate: they describe actions that did not occur, or they assume a typical event order based on language priors rather than what the video actually shows. Existing fixes rely on textual filtering or random video perturbations such as frame shuffling, which disrupt temporal coherence but rarely produce semantically meaningful alternatives. This paper builds a scalable pipeline that generates videos differing only in their action or action order while preserving the same scene, then uses those paired videos to fine-tune VLMs away from language-prior guessing.

Key Contributions

1. A modular counterfactual video generation pipeline that combines language-guided action proposal (multimodal LLM) with diffusion-based image editing and image-to-video generation to produce semantic hard negatives at scale.
2. CounterVid, a synthetic preference dataset of 26,167 pairs (plus 2,910 held-out samples) designed to expose action-recognition and temporal-ordering hallucinations under controlled visual context, in free-form, binary-choice, multiple-choice, and order-list formats.
3. MixDPO, a unified preference-based alignment framework that jointly optimizes textual preferences (grounded vs. hallucinated answers under fixed video) and visual preferences (correct vs. counterfactual video under fixed text).
4. Empirical validation across Qwen2.5-VL-3B, Qwen2.5-VL-7B, and InternVL3-9B, showing gains on the CounterVid benchmark, mostly positive transfer to established hallucination benchmarks, and largely preserved general video understanding.

Main Findings

• Large gains on the CounterVid benchmark. MixDPO raises average accuracy from 36.9% to 45.2% on Qwen2.5-VL-3B, from 57.8% to 66.2% on Qwen2.5-VL-7B, and from 53.8% to 74.0% on InternVL3-9B.
• Temporal ordering is where the method helps most. Order-list accuracy rises from 1.6% to 16.7% for Qwen2.5-VL-3B and from 16.5% to 43.8% for Qwen2.5-VL-7B. On Qwen2.5-VL-7B, MixDPO beats text-only DPO (T-pref-DPO) by +6.7pp on order-list and +5.3pp on binary temporal questions.
• Visual preferences add something textual supervision cannot. On InternVL3-9B, MixDPO improves over T-pref-DPO by 2.0pp on average, with the largest gains on temporal ordering.
• SFT on the source domain does not help. The SFT baseline yields minimal or negative changes (e.g., -0.7pp for Qwen2.5-VL-7B), and PaMi-VDPO (frame-shuffle/crop augmentation negatives) struggles on CounterVid, especially on temporal tasks.
• Positive but non-uniform transfer to external hallucination benchmarks. MixDPO improves EventHallusion over the base model on all three backbones, reaching 64.4%, 72.5%, and 73.7% for the 3B, 7B, and 9B models. The clearest extra gains over T-pref-DPO are on VidHalluc TSH (+2.7pp on Qwen2.5-VL-7B, +5.3pp on InternVL3-9B; tied on the 3B). Across 12 matched external comparisons, MixDPO improves over T-pref-DPO in eight cases and ties in one; the three regressions are -0.2pp on EventHallusion and -0.5pp on VideoHallu for Qwen2.5-VL-7B, and -0.6pp on VideoHallucer for InternVL3-9B.
• General video understanding is largely preserved. VideoMME stays stable across Qwen2.5-VL and InternVL3; NExT-QA improves for Qwen2.5-VL-3B and stays close to baseline for 7B and 9B; TempCompass improves for Qwen2.5-VL, including a caption-matching gain from 76.2% to 80.2% for 7B, while the caption-matching subset decreases for InternVL3.
• Human evaluation supports dataset quality. On 300 randomly sampled held-out examples (10.3% of the 2,910), reviewed by ten evaluators each, 87.8% of samples were judged correct and only 5.4% wrong. On visual quality, 66.6% were rated good or excellent and 88.0% at least acceptable, with 12.0% poor or very poor and 2.3% rated very poor. Per-format correctness was highest for multiple choice (95.0%) and lowest for order-list (81.1%).
• The method is insensitive to narrow hyperparameter tuning. Changing λ fourfold alters average accuracy by at most 0.4pp on Qwen2.5-VL-7B and 0.2pp on InternVL3-9B. Moving from a 50/50 to a 90/10 visual/textual mixture changes the average by 0.5pp and 1.0pp respectively; 90/10 exceeds the default 70/30 by only 0.4pp on each backbone, so 70/30 is retained.
• Data generation is a substantial compute effort. Producing the ~26k preference pairs took approximately 3,000 GPU-hours across up to three NVIDIA A100 GPUs (40GB), plus about $60 in LLM API cost. Roughly 0.6% of samples were removed by automated safety classifiers (NudeNet, Q16).

Methodology in Plain English

The researchers start from real video-caption pairs in the PE Video Dataset (PVD), keeping only clips shorter than 10 seconds that likely contain a single action. From each clip they extract a representative keyframe using SigLIP-SO400M in a coarse-to-fine retrieval scheme (2 fps coarse, then 12 fps refined, with similarity averaged over center/left/right crops).

That keyframe and its caption are passed to Claude-4-Sonnet, which proposes several alternative actions that are plausible for the scene, distinct from one another, and expressible in a short clip; the same LLM filters them for realism, feasibility, and clarity. Each accepted action is turned into structured editing instructions, and Qwen-Image-Edit generates an end frame showing the action completed from the starting keyframe. Because image editing can drift semantically, an iterative refinement loop (up to five attempts) has the LLM check whether the result matches the intended action and preserves scene identity, revising the prompt if not. Wan2.2-I2V-14B then synthesizes a roughly three-second clip at 680×384 from the start frame, end frame, and action caption.

From these generated clips, which share an anchor frame but differ in action dynamics, the authors build preference pairs. For action recognition, the chosen answer is the clip's own caption and the rejected answer is a caption sampled from another action from the same anchor (textual preference), or the same question and answer are paired with the correct versus a different clip (visual preference). For temporal ordering, several clips are concatenated; the chosen order is the concatenated order and the rejected order is a non-identity random permutation, with the analogous textual and visual preference constructions. The dataset keeps a consistent 70% visual / 30% textual preference ratio, and the train-test split is done at the anchor level so no anchor appears in both sets.

Fine-tuning uses MixDPO, a DPO-style objective that sums a textual preference loss and a λ-weighted visual preference loss. The textual term pushes the model to prefer grounded answers over fluent but visually unsupported ones, while the visual term forces the same answer to receive high likelihood only under the correct video — penalizing the situation where the model assigns similar probabilities to a correct and a counterfactual clip. Training uses parameter-efficient fine-tuning with the vision encoder frozen. Baselines include the un-tuned model, SFT on original PVD captions, text-only DPO (T-pref-DPO), and PaMi-VDPO, all under a matched three-epoch budget.

Why This Matters

The work targets the root cause of a specific class of VLM failure — over-reliance on language priors about what "usually" happens — rather than patching symptoms with architectural changes or extra inference passes. Because all videos and answers are generated, the approach scales without manual annotation of counterfactual content, which is expensive and hard to collect naturally. The pipeline is modular, so components (LLM proposer, image editor, video generator) can be swapped as generative models improve.

Real-world applications:

• Video search and recommendation: more reliable action and event-order recognition improves indexing and retrieval of short clips, which matters for a catalog-heavy streaming context.
• Content moderation and safety review: distinguishing what actually happens in a clip from what a model assumes could reduce both missed violations and false flags.
• Accessibility: audio description and captioning systems that describe action sequences correctly benefit users who cannot see the video.
• Surveillance and security analysis: correct temporal ordering of events is essential when reconstructing what happened and in what sequence.

Industry relevance: the paper is a collaboration with Amazon Prime Video, and the problem it addresses — models confidently describing actions and event orders that are not in the footage — directly affects large-scale video platforms where automated understanding is applied at catalog scale. The relatively modest reported API cost ($60) alongside the 3,000 GPU-hours also gives a concrete sense of the practical trade-off between synthetic data generation and annotation.

Future Directions

• Stronger automated quality control. The authors suggest refining the evaluation mechanisms for action-proposal filtering and video generation, for example through more comprehensive sanity checks, to raise the proportion of high-quality samples.
• Longer and more complex temporal reasoning. Current counterfactuals cover short actions (~3 seconds) and controlled concatenations. Long-range causal dependencies, overlapping events, and multi-agent interactions are explicitly left unaddressed.
• Breaking dependence on specific generators. The pipeline is modular at the interface level and alternative VLM judges reproduce most filtering decisions, but invariant quality under replacement of the image editor or video generator is not established; cross-generator validation remains future work.
• Full end-to-end training. The current setup freezes the vision encoder; unfreezing it might improve sensitivity to fine-grained visual cues but brings substantially higher compute and optimization complexity.
• Addressing residual synthesis artifacts. The authors note that residual artifacts could act as shortcut signals; their quality controls and external transfer results make a purely artifact-driven explanation less likely, but do not eliminate the confound.

Target Audience

Researchers and engineers working on multimodal large language models, video understanding, and preference-based alignment will get the most from this paper, particularly those dealing with hallucination in video QA or temporal reasoning. It is also relevant to practitioners building video search, moderation, or captioning systems who need to know whether synthetic counterfactual data is a practical alternative to manual preference annotation, and to anyone studying how DPO-style objectives can be extended from answer-level to input-level supervision.

Authors’ abstract

Video-language models (VLMs) achieve strong multimodal understanding but remain prone to hallucinations, especially when reasoning about actions and temporal order. Existing mitigation strategies, such as textual filtering or random video perturbations, often fail to address the root cause: over-reliance on language priors rather than fine-grained visual dynamics. We propose a scalable framework for counterfactual video generation that synthesizes videos differing only in actions or temporal structure while preserving scene context. Our pipeline combines multimodal LLMs for action proposal and editing guidance with diffusion-based image and video models to generate semantic hard negatives at scale. Using this framework, we build CounterVid, a synthetic dataset of ~26k preference pairs constructed from short counterfactual action clips and targeting both action recognition and controlled action-sequence ordering. We further introduce MixDPO, a unified Direct Preference Optimization approach that jointly leverages textual and visual preferences. Across Qwen2.5-VL and InternVL3 backbones, MixDPO substantially improves action recognition and temporal ordering, yields gains on most standard video hallucination benchmarks, and largely preserves general video understanding. Our source code, trained models, and dataset are available at https://aimagelab.github.io/CounterVid.