acceptodds
Under review as a conference paper at ICLR 2027

SafeLens: Deliberate and Efficient Video Guardrails with Fast-and-Slow Screening

Abstract

The rapid growth of online video platforms and AI-generated content has made reliable video guardrails a key challenge for safety and real-world deployment. While most videos can be screened through fast pattern recognition, a small subset requires deeper reasoning over temporally complex content and nuanced policy constraints. Existing approaches typically rely on large vision-language models applied uniformly across all inputs, resulting in high inference costs and inefficient allocation of computation. We propose SafeLens, a video guardrail framework that introduces a fast-and-slow inference architecture for efficient and accurate content moderation with variable computational cost across inputs. Additionally, we construct a high-quality dataset by applying influence-guided filtering to the SafeWatch Dataset, retaining only 2.4% of the original data. To further address limitations of training-time scaling, we enable test-time reasoning by augmenting the filtered data with structured Chain-of-Thought traces. Across real-world and AI-generated video benchmarks, SafeLens achieves state-of-the-art performance, outperforming strong open-source video guardrails (e.g., SafeWatch-8B, OmniGuard-7B) and closed-source models (e.g., GPT-5.4, Gemini-3.1-pro) while significantly reducing inference cost, demonstrating that efficient design serves to be more effective than scaling data or model size alone.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Video Guardrails, Fast-and-Slow Reasoning, AI Safety

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_v8tKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

SafeLens proposes a "fast-and-slow" cascading framework for video content moderation: S1 uses a lightweight probe + frame-level captioning for initial screening, and S2 uses a Chain-of-Thought VLM for in-depth reasoning on low-confidence samples. On the data side, the authors apply TracIn-based influence filtering to the SafeWatch public subset (roughly 6.5% of the original 2M), retaining only 48K samples (2.4% of the full set) and augmenting them with CoT traces. The backbone is Qwen3-VL-2B; CoT generation uses Qwen3.5-27B; captioning uses Florence-2-Large. On the SafeWatch-GenAI test set, the system reports 76.7% Avg ACC and 75.3% Macro F1, with inference latency of approximately 1.76s.

Critical strengths
  1. Clearly motivated problem: The tension between "uniform large-model inference is too expensive" and "simple classifiers are insufficiently accurate" is a genuine pain point in real deployment. A fast-and-slow cascade has direct engineering value.
  2. Data filtering angle is meaningful: Using influence functions to curate a high-quality 2.4% subset from noisy data, with an ablation in Appendix K showing filtered vs. unfiltered (73.9→76.7%), supports the "quality over quantity" argument.
  3. Efficiency-accuracy trade-off analysis is fairly complete: Threshold τ sweep (Figure 4), nine backbone combinations (Figure 5, Tables 9/10), runtime breakdown (Table 11), and FLOPs comparison (Figure 6) map out a deployable parameter space.
  4. Code and data release: An anonymous repository is available, and the filtered dataset is also released, which is favorable for reproducibility.
Critical weaknesses

Methodological credibility of influence filtering

  • Circularity risk: Influence analysis requires a backbone model to compute gradients (Eq. 1). The authors first fine-tune Qwen3-VL-2B on the noisy full training set to obtain Mθ_sw, then use this noise-contaminated model to determine which samples are "influential." If the backbone itself is biased by mislabeled examples, negatively influential samples may precisely be the correct examples that disagree with the erroneous labels. The paper does not validate this.

  • Validation set noise: The reference point for influence is SafeWatch-Real (644 samples). If this validation set itself has label noise (the authors themselves show noise examples from the SafeWatch training set in Appendix L, but do not characterize validation set quality), then the selection criterion of "positive average influence within class" rests on a noisy reference.

  • Threshold sensitivity: The two conditions in Algorithm 1 Line 4 (positive within-class influence, non-negative aggregate influence) are hard thresholds. The paper does not show how many samples were removed at each step, how the class distribution of the retained set shifts, or how final performance varies with threshold choices. From 130K public subset to 48K, the retention rate is roughly 37%, but is this number reproducible?

  • TracIn approximation limitations: Using gradient inner products of the last layer as TracIn is a very rough approximation. The authors acknowledge in Limitations that Hessian-based methods might be better, but provide no comparison quantifying the error introduced by this approximation.

Questions for authors

see as the above

4 comments
AuthorsOpenReview ↗

Reviewer v8tK - 1/2

Dear Reviewer v8tK,

Thank you for your time and effort in reviewing our paper. We appreciate your observations and provide clarifications for each point below.



W-1: Circularity risk: Influence analysis requires a backbone model to compute gradients (Eq. 1). The authors first fine-tune Qwen3-VL-2B on the noisy full training set to obtain Mθ_sw, then use this noise-contaminated model to determine which samples are "influential." If the backbone itself is biased by mislabeled examples, negatively influential samples may precisely be the correct examples that disagree with the erroneous labels. The paper does not validate this.


WR-1: Thank you for this observation. We provide empirical validation of this concern below.

  • The selection criterion is anchored to the human-verified validation set, not to the backbone's predictions. A sample is kept only if its gradient aligns with the validation (human-verified) gradients (Eq. 1, Alg. 1 Line 4). A correct sample that contradicts the backbone's noisy behavior should still align with the clean human-verified validation gradient.
  • Fine-tuning a backbone on the training data before influence analysis follows prior influence-based curation work [1, 2, 3].
  • Filtered data outperforms the unfiltered data. If the filter removed mostly correct samples, the filtered 48K would not beat the full unfiltered set. We compare the final SafeLens pipelines (CoT + probe-score conditioning) trained on both:
ModelAvg. Acc.Macro F1
SafeLens (Unfiltered)73.9 ± 0.673.6 ± 0.7
SafeLens (Filtered)76.7 ± 0.275.3 ± 0.2
  • Filtered data outperforms random selection of the same size and class distribution. We randomly sampled the same number of samples with the same class distribution from the unfiltered data:
DatasetAvg. Acc.Macro F1
Random (same size, same class distribution)73.0 ± 0.970.9 ± 0.9
Influence-filtered76.7 ± 0.275.3 ± 0.2
  • Filtering gives +3.7 Avg. Acc. and +4.4 Macro F1 over random selection. The gain is therefore not a data-reduction or class-balance effect. It comes from which samples are selected.
  • All evaluation is on the test set, which is not used in influence analysis, so there is no leakage.
  • Our goal is a net reduction in noise, and the results above show this is achieved.
  • We have added these results to Appendix L.


W-2: Validation set noise: The reference point for influence is SafeWatch-Real (644 samples). If this validation set itself has label noise (the authors themselves show noise examples from the SafeWatch training set in Appendix L, but do not characterize validation set quality), then the selection criterion of "positive average influence within class" rests on a noisy reference.


WR-2: Thank you for raising this point. We characterize the validation set quality below.

  • The SafeWatch-Real validation set is human-verified, according to the SafeWatch paper [4]. The noise examples in Appendix M are from the training set, which is not fully human-verified.
  • Class-level average influence (Alg. 1 Line 4) is computed over 30-160 validation samples per class, so the few isolated label errors have limited effect on the class-level average.
  • The downstream results (comparison provided in WR-1) show that filtering with this validation set gives consistent gains on the independent test set.


AuthorsOpenReview ↗

Reviewer v8tK - 2/2



W-3: Threshold sensitivity: The two conditions in Algorithm 1 Line 4 (positive within-class influence, non-negative aggregate influence) are hard thresholds. The paper does not show how many samples were removed at each step, how the class distribution of the retained set shifts, or how final performance varies with threshold choices. From 130K public subset to 48K, the retention rate is roughly 37%, but is this number reproducible?


WR-3: Thank you for the observation. In influence analysis, using a zero threshold to distinguish beneficial from detrimental samples is a standard practice in prior work [5, 6]. Therefore, we believe that varying the threshold is not necessary for our analysis. We clarify below.

  • There is no tunable threshold. Both criteria use the sign of the influence score. In prior work, positive influence marks a beneficial sample and negative influence marks a detrimental sample [5,6]. Zero is therefore the natural decision boundary, and varying it would require an arbitrary hyperparameter with no principled meaning.
  • Class-wise sample reduction. We provide the percentage of each class removed and each class's share of the unfiltered data:
Class% of Class RemovedShare of Unfiltered Data
Sexual67.8%0.35
Abuse44.5%0.03
Violence79.7%0.13
Misinfo34.4%0.10
Illegal26.8%0.02
Extreme57.7%0.05
Safe57.2%0.30
  • The filtered set is a deterministic function of the training data, the validation set, and the backbone. Code, filtered subset, datasets, and models are all publicly available, so the retention rate is reproducible.

  • We have added this table to the revised Appendix B.



W-4: TracIn approximation limitations: Using gradient inner products of the last layer as TracIn is a very rough approximation. The authors acknowledge in Limitations that Hessian-based methods might be better, but provide no comparison quantifying the error introduced by this approximation.


WR-4: Thank you for the observation. For large datasets, TracIn is a scalable approach to influence analysis that has been adopted in prior work and these studies demonstrate its effectiveness in their respective analyses [5, 6]. Our results provide further empirical support: the filtered data consistently outperforms the unfiltered data. We clarify below.

  • Last-layer, Hessian-free TracIn is a scalable choice made in existing work [5, 6]. Hessian-based estimation for a 2B VLM over about 100K+ ×\times 644 training-validation pairs is impractical at this scale. This is the reason we use a Hessian-free method.
  • The relevant measure is the quality of the selected data. Filtered data achieves +2.8 Avg. Acc. over unfiltered data and +3.7 Avg. Acc. over random selection of equal size and class distribution.
  • These results show the approximation is sufficient to select high-quality data, and SafeLens outperforms all baselines with this setting.
  • The influence method is a design choice that practitioners can adapt to their resources, and a Hessian-based method can be plugged into our pipeline without changes to the rest of SafeLens.


References:

  • [1] Humane, P., et. al., Influence Functions for Efficient Data Selection in Reasoning. NeurIPS 2025 Workshop on Efficient Reasoning.
  • [2] Kwon, Y., et. al., DataInf: Efficiently Estimating Data Influence in LoRA-tuned LLMs and Diffusion Models. ICLR 2024.
  • [3] Yang, et. al., Dataset Pruning: Reducing Training Data by Examining Generalization Influence. ICLR 2023
  • [4] Chen et al., SafeWatch: An Efficient Safety-Policy Following Video Guardrail Model with Transparent Explanations, ICLR 2025; OpenReview.
  • [5] Pruthi et al., Estimating Training Data Influence by Tracing Gradient Descent, NeurIPS 2020.
  • [6] Vitel et. al., First is Not Really Better Than Last: Evaluating Layer Choice and Aggregation Strategies in Language Model Data Influence Estimation, ICLR 2026.

Dear Reviewer v8tK,

Thank you once again for your insightful review. With the deadline approaching in just one day, we would be happy to answer any remaining questions.

If our rebuttal has addressed your concerns, we would also kindly request for a reconsideration of the score.

Kind Regards,

Authors.

Dear Reviewer v8tK,

We would like to thank you once again for taking the time to review our work and share your valuable feedback. We are happy to answer any additional questions you may have.

As the rebuttal period is ending in a few hours, if our responses have addressed your main concerns, we would be grateful if you would reconsider your assessment.

Kind Regards,

Authors.

Reviewer_i4f5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Some videos can be easily classified as safe, while others require detailed analysis. Aiming for deliberate and efficient video guardrails, the authors propose a fast-and-slow video guardrail framework, combining a lightweight probe-based screening module (SAFELENS-S1) with a test-time scaling-enabled reasoning module (SAFELENS-S2). They also construct a video moderation dataset from SafeWatch and augment it with CoT traces to train their framework. Results show improvements in moderation performance and inference efficiency.

Critical strengths
  • S1: The proposed dataset could be useful. (Though not released yet)
  • S2: The method shows a clear advance in the accuracy–runtime trade-off.
Critical weaknesses
  • W1: It seems this work mainly narrows general fast-and-slow techniques down to the video-guardrail domain. The novelty needs to be stated much more clearly.
  • W2: As shown in Table 3, combining S1 and S2 does not show clear improvement, maybe lightweight S1 is enough
Questions for authors
  • A small-scale human verification would make the "high-quality dataset" statement more convincing
3 comments

Dear Reviewer i4f5,

Thank you for your time and effort in reviewing our paper. We appreciate your observations and suggestions, and provide clarifications and additional experiments below to address the queries raised.



W-1: It seems this work mainly narrows general fast-and-slow techniques down to the video-guardrail domain. The novelty needs to be stated much more clearly.


WR-1: Thank you for the observation. We agree the novelty should be stated more clearly, and clarify it below.

  • Our contribution is not a generic application of fast-and-slow thinking. Existing fast-and-slow methods [1, 2] switch between short and long reasoning inside the same model. Video guardrails need a different design because the fast path must be a cheap classifier with a calibrated confidence for routing, and the slow path must reason over the fast path's outputs.
  • Specific design contributions:
    • S1 is a probe on hidden states of the fine-tuned VLM. It needs one forward pass and no autoregressive decoding (0.04 s vs. 4.45 s for S2 decoding), and its confidence is the routing signal.
    • S2 reasons over S1's output. S2 receives frame captions and probe confidences, and is trained to reason over them. Removing these S1 signals drops S2 from 76.7 to 69.7 Avg. Acc. (Table 8), so the interface between S1 and S2 is a key part of the method.
    • Confidence-based cascade with a controllable trade-off. The threshold τ\tau moves SafeLens between speed and accuracy, and S1 confidence is well calibrated where it operates (ECE = 0.054; 94.9% accuracy on the 67.4% of samples S1 handles at τ\tau = 0.9).
  • Data contribution. To our knowledge, this is the first use of influence-based curation for video-level data. We train on 2.4% of SafeWatch, and the filtered data outperforms both the full unfiltered data (76.7 vs. 73.9 Avg. Acc.) and a random subset of the same size and class distribution (76.7 vs. 73.0 Avg. Acc. in Table 15).
  • CoT data tailored to the cascade. Existing datasets have no reasoning traces grounded in a fast system's outputs. We construct such traces (Alg. 1, Stage 2), which S2 requires.
  • Empirical outcome. SafeLens outperforms larger open-source and closed-source models and existing guardrails, at 1.76 s average runtime, with a 2B backbone.


W-2: As shown in Table 3, combining S1 and S2 does not show clear improvement, maybe lightweight S1 is enough.


WR-2: Thank you for this observation. However, we believe there may be a misunderstanding, as SafeLens (S1+S2) outperforms S1 on both the validation and test sets, achieving the highest Avg. Acc and Macro F1 scores. We provide the clarification below.

  • Validation set (Table 3): SafeLens achieves 82.9 Avg. Acc. and 81.7 Macro F1, vs. 82.0 and 80.8 for S1.
  • Test set (Table 1): SafeLens achieves 76.7 Avg. Acc. and 75.3 Macro F1, vs. 73.1 and 72.6 for S1. This is +3.6 Avg. Acc. and +2.7 Macro F1.
  • S1 alone fails on harder classes. On the test set, S1 gets 52.0% on Misinfo and 52.2% on Abuse, vs. 76.4% and 62.3% for SafeLens. These are the cases that need deeper reasoning, and S2 recovers them.
  • S1 alone is a valid operating point, not a replacement. SafeLens provides flexibility to tune τ\tau according to deployment requirements.


Q-1: A small-scale human verification would make the "high-quality dataset" statement more convincing.


QR-1: Thank you for the suggestion. We randomly sampled 100 videos from the filtered training set and asked two graduate students to independently annotate them using the guidelines in Appendix K. We computed multi-class Cohen's κ\kappa between the original labels and the two annotators.

ComparisonCohen's κ\kappa
Annotator 1 vs. Annotator 20.890
Original label vs. Annotator 10.868
Original label vs. Annotator 20.856
  • All three scores fall in the almost-perfect agreement range (κ\kappa = 0.8-1.0) [3].
  • The agreement between the original labels and each annotator (0.868, 0.856) is close to the agreement between the two annotators (0.890). The filtered labels are therefore about as consistent with humans as humans are with each other.

We have included this analysis in the revised manuscript in Appendix B.



References:

  • [1] Su et al., DualFormer: Controllable Fast and Slow Thinking by Learning with Randomized Reasoning Traces, ICLR 2025.
  • [2] Pan et al., DynaThink: Fast or Slow? A Dynamic Decision-Making Framework for Large Language Models, EMNLP 2024.
  • [3] Landis & Koch, The Measurement of Observer Agreement for Categorical Data, Biometrics, 1977.

Dear Reviewer i4f5,

Thank you once again for your insightful review. With the deadline approaching in just one day, we would be happy to answer any remaining questions.

If our rebuttal has addressed your concerns, we would also kindly request for a reconsideration of the score.

Kind Regards,

Authors.

Dear Reviewer i4f5,

We would like to thank you once again for taking the time to review our work and share your valuable feedback. We are happy to answer any additional questions you may have.

As the rebuttal period is ending in a few hours, if our responses have addressed your main concerns, we would be grateful if you would reconsider your assessment.

Kind Regards,

Authors.

Reviewer_VBShOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies efficient video content moderation. It filters SafeWatch data using influence scores and adds reasoning text. A fast probe handles confident cases, a slow model combines video frames, captions and probe scores to handle difficult cases. The method reports 75.3% Macro F1 on SafeWatch-GenAI and lower latency than only using the slow model.

Critical strengths
  1. The improvement on the GenAI set is meaningful. Macro F1 acheieves 75.3%.

  2. The cascade keeps the reported average performance of S2, and the latency reduces from 5.02s to 1.76s.

Critical weaknesses
  1. Evidence of generalization to unseen real videos is limited. The experimental setup uses SafeWatch-Real for influence analysis and validation, while the main test results are on GenAI.
Questions for authors
  1. Does influence-based filtering outperform random selection with the same sample size and class distribution?

  2. The improvements are on GenAI. Do the gains hold on real videos?

4 comments

Dear Reviewer VBSh,

Thank you for your time and effort in reviewing our paper. We provide clarifications on the points raised below.



W-1: Evidence of generalization to unseen real videos is limited. The experimental setup uses SafeWatch-Real for influence analysis and validation, while the main test results are on GenAI.

Q-2: The improvements are on GenAI. Do the gains hold on real videos?


WR-1 & QR-2: Thank you for the observation and the clarification question. To evaluate generalization to real videos, we evaluated SafeLens on an additional held-out set of 1004 clean real videos provided by SafeWatch. These videos are not used in influence analysis, training, or threshold selection.

ModelSexualAbuseViol.Misinfo.IllegalExtremeSafeAvg. ACCMacro F1
GPT-5.489.5±1.014.5±1.686.8±2.69.7±1.078.9±0.010.8±2.790.8±0.254.5±0.849.8±1.2
Gemma-494.2±0.24.8±1.659.2±6.613.1±1.589.5±0.010.8±2.790.7±0.651.8±0.647.3±1.6
SafeWatch-8B88.7±2.311.3±4.873.7±7.942.7±3.931.6±15.829.7±8.195.8±0.053.4±4.856.9±4.7
SafeLens88.2±0.889.2±1.593.8±1.270.2±1.293.0±2.561.3±2.590.5±0.683.7±0.676.0±0.5
  • SafeLens outperforms all strong baselines (Closed-source, open-source, and guardrails) on held-out real videos: +29.2 Avg. Acc. and +19.1 Macro F1 over the best baseline in each metric.
  • The gains on real videos are consistent with the GenAI results, so we would like to emphasize that the improvement is not specific to synthetic content.

We have added these results to Appendix D.



Q-1: Does influence-based filtering outperform random selection with the same sample size and class distribution?


QR-1: Thank you for the clarification question. We randomly sampled the same number of samples, with the same class distribution, from the unfiltered data and compared the performance of SafeLens on the test set.

DatasetSexualAbuseViol.Misinfo.IllegalExtremeSafeAvg. ACCMacro F1
Random97.5±1.247.9±2.156.7±3.092.7±2.067.3±1.986.7±0.062.1±0.073.0±0.970.9±0.9
Filtered92.3±0.062.3±2.765.7±1.276.4±3.570.5±1.889.8±1.479.5±0.976.7±0.275.3±0.2
  • Influence-based filtering outperforms random selection by +3.7 Avg. Acc. and +4.4 Macro F1.
  • The gain is therefore not a data-reduction or class-balance effect. It comes from which samples are selected.
  • We have added these results to Appendix L.


Dear Reviewer VBSh,

Thank you once again for your insightful review. With the deadline approaching in just one day, we would be happy to answer any remaining questions.

If our rebuttal has addressed your concerns, we would also kindly request for a reconsideration of the score.

Kind Regards,

Authors.

Thank you for the additional experiments. The held-out real-video evaluation substantially addresses my main concern, since these videos were not used for influence analysis, training or threshold selection. The results support gains beyond the GenAI test set.

The random-selection control with the same sample size and class distribution also addresses my question about filtering. It provides useful evidence that the selected samples matter, beyond data reduction or class balance.

Although broader cross-dataset generalization remains to be tested, the reported additions address my main concerns. I am raising my rating from 2 to 3 (Weak acceptance).

Dear Reviewer VBSh,

Thank you for increasing your assessment. We really appreciate it, and we are glad that we could address your concerns. Thank you again for putting your time and effort into reviewing our work and helping us improve it.

Kind Regards,

Authors.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Manuscript Updates

Dear Chairs and Reviewers,

Thank you for your time and effort in reviewing our paper. We have conducted additional experiments to address the concerns raised and updated the manuscript accordingly. To distinguish the newly added section from the previous version, we have highlighted all updates in blue font throughout the revised PDF.

The major additions and improvements are summarized below:

  • Generalization to real-world videos: We evaluated SafeLens on an additional held-out set of 1,004 real videos, achieving 83.7% Avg. Accuracy and 76.0 Macro F1, outperforming strong open-source, closed-source, and guardrail baselines.

  • Validation of influence-based data filtering: We added a controlled comparison against random selection using the same sample size and class distribution. Influence-based filtering improves Avg. Accuracy by 3.7 percentage points and Macro F1 by 4.4 percentage points.

  • Human verification of data quality: We conducted an independent annotation study involving two graduate students on 100 randomly sampled videos. The Cohen's Kappa scores demonstrate almost-perfect agreement between human annotators and the original labels.

  • Clarification of methodological contributions: We clarified the novelty of SafeLens, highlighting the probe-based fast screening mechanism, confidence-guided cascading, reasoning conditioned on fast-path outputs, and influence-based video data curation.

All corresponding experiments, results, and analyses have been incorporated into the revised manuscript and its appendices.

We sincerely appreciate the reviewers' constructive feedback, which has helped us strengthen our work.

Kind Regards,

Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.