Video-PLR: Hallucination-Resistant Video Reasoning

• Video-PLR is a video reasoning framework that alternates between localized perception and explicit reasoning to gather timestamped evidence.
• It employs a Perception Loop Reasoning paradigm and a Factual-Aware Evaluator with anti-hallucination rewards to ensure evidence accuracy.
• Empirical results show state-of-the-art performance on temporal benchmarks with enhanced data efficiency and reliable factual grounding.

Video-PLR is a video LLM framework for hallucination-resistant video understanding that formalizes video reasoning as an alternation between localized perception and explicit reasoning, rather than a single global description followed by answer generation. In its canonical form, introduced with Qwen2.5-VL backbones at 3B and 7B scales, it combines a loop-based Perception Loop Reasoning (PLR) paradigm, a learned Factual-Aware Evaluator (FAE) for segment-level factual judgment, a large hallucination preference dataset AnetHallu-117K, a cold-start reasoning dataset VideoPLR-14K, and a GRPO-based reinforcement-learning stage that rewards both task correctness and factual grounding (Pu et al., 23 Nov 2025).

1. Concept and problem setting

Video-PLR addresses a specific weakness of many VideoLLMs: the single-step perception pattern in which the model first produces a global description of a video and only then performs reasoning. That pattern creates what is described as a perception shortcut. The model can end up solving the task from its own generated description rather than from sufficiently grounded visual evidence, which leads to two characteristic failure modes: partial and incomplete perception, where key segments are omitted, and hallucinatory perception, where the description contains objects, actions, or temporal relations not supported by the video (Pu et al., 23 Nov 2025).

The framework is aimed at complex video understanding tasks requiring temporal causality, multi-hop reasoning, and temporal grounding. Benchmarks discussed alongside the framework include Video-Holmes, VCR-Bench, MMVU, VSI-Bench, Video-tt, TVG-Bench, VideoMME, and supplementary evaluations on NextGQA. This benchmark profile places Video-PLR in the regime where the model must not merely recognize salient frames, but also assemble evidence across multiple temporally separated events (Pu et al., 23 Nov 2025).

The central design claim is that sufficient visual perception is the foundation of video reasoning. Video-PLR operationalizes that claim by turning evidence acquisition into part of the reasoning trace itself. Instead of a monolithic caption, the model emits timestamped, segment-level evidence that can later be evaluated for factuality. This suggests a shift from answer-centric optimization toward evidence-centric optimization in video reasoning systems (Pu et al., 23 Nov 2025).

2. Perception Loop Reasoning paradigm

The Perception Loop Reasoning paradigm structures inference into three stages. The model first performs initial analysis of the question, then enters an iterative perception–reasoning loop, and finally produces summarization and conclusion. In each loop, the model selects a video segment with explicit timestamps, describes that segment, analyzes how the evidence bears on the question, and decides whether another segment should be inspected or whether it has enough information to answer (Pu et al., 23 Nov 2025).

Each evidence unit has a standardized form:

$$e_i = \langle \text{start} = t_i^{start},\, \text{end} = t_i^{end},\, \text{desc} = \text{description} \rangle.$$

During inference, the reasoning trace is placed in > ... and the final answer in <answer>...</answer>, while evidence is emitted in tags of the form: $$P_{Y/N} = E\left( \text{clip}(Video, t_{start}, t_{end}), desc \right),$$9 This makes temporal grounding explicit inside the chain of thought rather than attaching it as a separate post hoc output (Pu et al., 23 Nov 2025).

PLR is implemented through prompting and training format rather than a new video encoder. The base model remains Qwen2.5-VL, but the supervision format teaches it to plan, seek evidence, and reason over timestamped clips. The reasoning state is therefore not an abstract latent only; it is a textualized sequence of evidence-bearing actions. A practical consequence is that temporal grounding and reasoning become jointly learnable behaviors instead of separate subtasks (Pu et al., 23 Nov 2025).

Relative to one-pass caption-then-reason pipelines, PLR changes the granularity of supervision. Each local caption can be judged independently, and each reasoning step can be tied to a specific temporal slice. This reduces the chance that a single early hallucination contaminates the entire downstream answer.

3. Factual-Aware Evaluator and anti-hallucination reward

The Factual-Aware Evaluator is the framework’s learned factuality judge. It is implemented as a LoRA-tuned Qwen2.5-VL 7B model trained to determine whether a caption matches a given video segment. For a clip and candidate description, the evaluator produces correctness probabilities

$$P_{Y/N} = E\left( \text{clip}(Video, t_{start}, t_{end}), desc \right),$$

where $Y$ denotes that the description matches the video and $N$ denotes mismatch (Pu et al., 23 Nov 2025).

FAE is trained with ORPO (Odds Ratio Preference Optimization) on AnetHallu-117K. The stated objective is

$$\mathcal{L}_{ORPO} = \mathbb{E}_{(x, y_w, y_l) \sim \mathcal{D}}\left[\mathcal{L}_{SFT} + \lambda \cdot \mathcal{L}_{OR}\right],$$

with

$\mathcal{L}_{SFT} = -\log P_{\theta}(y_w | x),$

and

$$\mathcal{L}_{OR} = -\log \sigma \left( \log \frac{\text{odds}_{\theta}(y_w | x)}{\text{odds}_{\theta}(y_l | x)} \right).$$

The dataset mixes direct yes/no judgments with CoT-style judgments, and the reported conclusion is that this hybrid training regime gives the best overall balance of accuracy and bias (Pu et al., 23 Nov 2025).

For reinforcement learning, FAE supplies an anti-hallucination reward over the set of generated evidences. If the model emits $n$ evidences $e_i$, each evidence is weighted by an attenuation term

$$w_i = 1 - \max_{j \neq i} \left( \text{IoU}(e_i, e_j) \times \text{ROUGE}(e_i, e_j) \right),$$

which discourages repeated or near-duplicate evidence in both time span and description. The anti-hallucination reward is then

$$P_{Y/N} = E\left( \text{clip}(Video, t_{start}, t_{end}), desc \right),$$0

The denominator is deliberately non-linear so that the model is not rewarded for emitting only a minimal number of very safe evidences; instead, it is pushed toward multiple, diverse, factual evidence snippets (Pu et al., 23 Nov 2025).

On hallucination benchmarks, FAE is reported to perform comparably to GPT-4o. On VideoHallucer it reaches Acc 75.4 and Diff 7.3, compared with GPT-4o at Acc 74.6 and Diff 0.9, and on HEAVEN it reaches Acc 72.9 and Diff 2.1, substantially reducing the class bias seen in untuned Qwen2.5-VL. Supplementary analysis further reports an AUC of 76.3 on HEAVEN and a reward-margin gap $$P_{Y/N} = E\left( \text{clip}(Video, t_{start}, t_{end}), desc \right),$$1 of 29.62%, supporting its use as a reward model rather than merely as an offline evaluator (Pu et al., 23 Nov 2025).

4. Data construction and training pipeline

Video-PLR relies on two custom datasets with distinct functions. AnetHallu-117K is a hallucination judgment preference dataset derived from ActivityNet segments and used to train FAE. VideoPLR-14K is a cold-start reasoning dataset formatted in PLR style and used to initialize the policy model before reinforcement learning (Pu et al., 23 Nov 2025).

AnetHallu-117K is built through a Hallucinate-then-Correct process. ActivityNet segment descriptions are first recaptioned into high-fidelity dense captions using Qwen2.5-VL-32B. Negative captions are then synthesized by injecting exactly one hallucination from five types: Attribute Modification, Quantity Modification, Action Substitution, Detail Conflation, and Temporal Reordering. A corresponding positive caption is generated by correcting the hallucinated one while preserving style and length. The pipeline then applies iterative Text Bias Elimination by removing the most polarity-biased captions over 15 iterations, with $$P_{Y/N} = E\left( \text{clip}(Video, t_{start}, t_{end}), desc \right),$$2 biased words tracked at each round (Pu et al., 23 Nov 2025).

VideoPLR-14K provides PLR-style reasoning traces. It is built from NextQA and ActivityNet ReOrdering data. For NextQA-derived instances, the construction process includes entity extraction, frame-level captioning, raw PLR CoT construction, FAE-based filtering of hallucinated events, and logic refinement. The resulting data explicitly teaches the model to alternate timestamp selection, local description, and reasoning. The cold-start stage also includes 20K image CoT data from Reason-RFT-CoT-Dataset (Pu et al., 23 Nov 2025).

The policy model is trained in two stages. First, supervised fine-tuning uses a learning rate of $$P_{Y/N} = E\left( \text{clip}(Video, t_{start}, t_{end}), desc \right),$$3, 1 epoch, Deepspeed ZeRO3, batch size 6, and maximum output length 1600 tokens. Second, reinforcement learning applies GRPO with $$P_{Y/N} = E\left( \text{clip}(Video, t_{start}, t_{end}), desc \right),$$4 sampled generations for the 7B model and $$P_{Y/N} = E\left( \text{clip}(Video, t_{start}, t_{end}), desc \right),$$5 for the 3B model. The GRPO objective is

$$P_{Y/N} = E\left( \text{clip}(Video, t_{start}, t_{end}), desc \right),$$6

The total reward used in RL is

$$P_{Y/N} = E\left( \text{clip}(Video, t_{start}, t_{end}), desc \right),$$7

combining task accuracy, think-format validity, evidence-format validity, and anti-hallucination reward (Pu et al., 23 Nov 2025).

5. Empirical performance and data efficiency

Video-PLR is reported to achieve state-of-the-art performance at both 3B and 7B parameter scales among the compared open-source systems, while also showing the best data efficiency in the study’s comparison set. At 3B scale, it reaches an average score of 40.8 across the main benchmark suite, compared with 35.3 for Time-R1 and 34.7 for Video-RFT. At 7B scale, it reaches 44.9, exceeding VideoChat-R1.5 at 43.2 and Video-RFT at 41.6 (Pu et al., 23 Nov 2025).

Its strongest gains appear on tasks that require evidence integration across multiple temporal segments. On Video-Holmes, the reported scores are 39.8 at 3B and 42.6 at 7B. On TVG-Bench, it reaches 24.9 at 3B and 31.7 at 7B, outperforming Time-R1 despite using only 0.5K VTG data versus Time-R1’s 2.5K VTG-focused training. On VideoMME, it reaches 57.3 at 3B and 61.8 at 7B (Pu et al., 23 Nov 2025).

The paper also emphasizes data efficiency. For the 7B model, Video-PLR uses 34K cold-start samples and 6K RL problems, whereas Video-R1 uses 165K + 8K, Video-RFT uses 102K + 8K, VideoChat-R1 uses 0 + 18K, and VideoChat-R1.5 uses 0 + 80K. The claim is not merely that Video-PLR is smaller-data, but that its segment-level reasoning format and reward design allow more effective use of the available supervision (Pu et al., 23 Nov 2025).

Ablations indicate that the anti-hallucination reward is most useful when paired with PLR-style cold-start supervision and with the non-linear normalization in $$P_{Y/N} = E\left( \text{clip}(Video, t_{start}, t_{end}), desc \right),$$8. Using only accuracy reward improves over no RL, but adding anti-hallucination reward without the normalization gives smaller or unstable gains. With the normalization, benchmark scores improve alongside increases in the number of perceptions, suggesting that the relevant behavioral change is expanded grounded evidence collection rather than longer free-form reasoning alone (Pu et al., 23 Nov 2025).

6. Terminological scope and related usages

The label “Video-PLR” is not uniform across the broader literature. In the hallucination-resistant video reasoning sense, it specifically denotes the Perception Loop Reasoning framework described above (Pu et al., 23 Nov 2025). Outside that usage, closely related or acronymically similar interpretations appear in several distinct research areas.

In networking, PLR usually means packet loss ratio. A study of LTE downlink scheduling for video traffic evaluates PLR as a QoS metric and concludes that Frame Level Scheduler achieves the lowest video PLR among the compared algorithms in a high-speed vehicular setting (Sahoo, 2013). A metro-haul smart-city video-surveillance demonstration discusses low-latency, high-bandwidth, QoS-sensitive video transport, but explicitly does not define or measure a “Video-PLR” metric (Dochhan et al., 2020).

In video-language modeling, the term can be interpreted differently. PPLLaVA does not use “Video-PLR” explicitly, but it combines prompt-guided pooling over video token sequences with video-specific Direct Preference Optimization, which the source material characterizes as a near-instantiation of “video-level pooling and preference(-based) learning for representation” (Liu et al., 2024). Perceptual Learned Video Compression likewise does not name “Video-PLR” directly, but it learns an entropy-constrained spatio-temporal latent representation shaped by adversarial perceptual objectives, making “perceptual learned representation” a plausible alternate expansion in the compression literature (Yang et al., 2021).

Open video perception-language work adds another nearby usage. PerceptionLM defines an explicitly open Perception LLM for detailed video understanding and introduces PLM-FGQA, PLM-STC, and PLM-VideoBench to evaluate the “what”, “where”, “when”, and “how” of video (Cho et al., 17 Apr 2025). This is conceptually adjacent to Video-PLR’s emphasis on evidence-rich video reasoning, but it is a different framework. The same holds for memory-efficient streaming systems such as ProVideLLM, which focuses on multimodal caches and long-horizon procedural understanding rather than hallucination-resistant segment-level reasoning (Chatterjee et al., 10 Apr 2025).

The principal misconception, therefore, is to treat “Video-PLR” as a settled universal acronym. In current arXiv usage, it is best understood as a term whose meaning depends on subfield context, with the most direct and explicit usage referring to the Perception Loop Reasoning framework for video understanding (Pu et al., 23 Nov 2025).

7. Limitations and prospective directions

The reported limitations of Video-PLR arise mainly from dependence on temporal grounding and the cost of iterative reasoning. The framework assumes that the base VideoLLM can generate accurate timestamps, yet precise time perception remains difficult for many open-source models. The experiments are also limited to 3B and 7B scales, and the authors note that they do not explore larger 10B+ or 72B-scale VideoLLMs (Pu et al., 23 Nov 2025).

Inference is slower than that of some comparator systems because the model generates longer reasoning traces with multiple evidence segments. Supplementary runtime analysis reports roughly 16.9s per sample for Video-PLR versus about 10.6–11.5s for Video-R1 and Video-RFT on MMVU. The framework also does not fully solve reasoning honesty: factual perception is improved, but slight mismatches between the thinking trace and final answer can remain. An attempted external “honesty” reward using a 32B judge reportedly reduced benchmark performance, which suggests that judge quality and reward bias remain open problems (Pu et al., 23 Nov 2025).

The future work directions stated for the framework include improving temporal grounding ability in the underlying video backbone, scaling PLR to larger models and longer videos, developing less biased judges for reasoning honesty, and extending the loop-based paradigm toward broader multimodal or tool-augmented video agents. A plausible implication is that the framework’s lasting contribution may be methodological: it turns timestamped evidence collection into a trainable intermediate action space for video reasoning, providing a template for future systems that combine perception control, factuality evaluation, and reinforcement learning in a single video-language stack (Pu et al., 23 Nov 2025).