acceptodds
Under review as a conference paper at ICLR 2027

C2I-CasITI-VTG: Confidence-to-IoU Cascaded Inference-Time Intervention for Video Temporal Grounding

Abstract

Video temporal grounding (VTG) aims to localize the temporal segment in an untrimmed video that corresponds to a natural-language query. Recent inference-time intervention (ITI) methods improve model behavior by steering internal activations. A natural VTG strategy is to subtract the mean activation of low-IoU samples from that of high-IoU samples, encouraging better-localized predictions. However, high IoU does not necessarily indicate that the underlying activation reliably represents successful localization. IoU and model confidence are sometimes misaligned: predictions can have high IoU with low confidence, or high confidence with poor overlap with the ground truth. Thus, IoU-only contrasts may mix inconsistent internal behaviors and produce noisy steering directions. We propose a confidence-to-IoU cascaded ITI framework for VTG (C2I-CasITI-VTG). First, confidence-based ITI calibrates the model's confidence behavior; IoU-based ITI then builds on these confidence-aligned representations to improve localization. Technically, the interventions cannot be naively combined as independently estimated steering vectors may point in incompatible directions and their superposition may cancel useful components. We therefore apply confidence-based ITI during prefill stage to reshape the subsequent decoding activations toward a confidence-conditioned representation space, from which we extract the IoU steering direction. In summary, our key contribution is to demonstrate that ITI can be cascaded to steer model behavior sequentially along complementary directions. Experiments on Charades-STA, ActivityNet-Captions, and QVHighlights show consistent gains across frozen MLLMs. On Charades-STA, our method improves Qwen3.5-9B from 53.93% to 57.47% mIoU and from 32.72% to 41.16% [email protected].

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video Temporal Grounding, Multimodal Large Language Model, Inference-Time

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_aZ9XOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes C2I-CasITI-VTG, a confidence-to-IoU cascaded inference-time intervention framework for video temporal grounding. The method is motivated by the mismatch between prediction confidence and localization accuracy, which may introduce noise when constructing IoU-based steering directions. To address this issue, the authors first apply confidence-based intervention during the prefill stage to obtain confidence-conditioned representations. The resulting samples are then filtered according to their confidence scores, and an IoU-based steering direction is constructed from the high-confidence subset. During inference, only the resulting IoU direction is applied to the decoding activations. Experiments on three VTG benchmarks with four MLLMs demonstrate consistent improvements over the corresponding frozen backbones.

Critical strengths
  1. The paper investigates the relationship between prediction confidence and localization accuracy, which is an interesting perspective for improving video temporal grounding.

  2. The experiments cover multiple MLLMs and three VTG benchmarks. The reported improvements, particularly under strict IoU thresholds, are encouraging. The cross-dataset evaluations also provide some evidence of transferability.

Critical weaknesses
  1. Although the proposed method only injects a single steering direction during inference, it still relies on a relatively large MLLM for temporal localization. The paper does not report inference latency, computational cost, or memory consumption. Considering that efficiency is an important factor for video temporal grounding, especially for long videos and large-scale applications, the authors should provide a more comprehensive efficiency analysis. Moreover, Table 1 mainly compares MLLM-based approaches, without including representative conventional VTG methods.

  2. In Section 3.2, confidence is estimated using the weighted entropy of timestamp-generation tokens. However, low token entropy does not necessarily indicate accurate temporal localization, especially when a model confidently generates incorrect timestamps. Furthermore, Table 2 shows that pure confidence intervention substantially decreases mIoU from 53.93% to 44.66%. This raises an important question about whether the confidence intervention actually improves the reliability of the internal representations. Although Fig. 3 provides some qualitative observations, quantitative analysis of the correlation between confidence and IoU is missing. The authors should further investigate this relationship and compare the proposed weighted entropy with other confidence estimation strategies.

Questions for authors
  1. The paper argues that independently constructed confidence and IoU steering directions may interfere with each other. Is there any more direct evidence of such interference? Moreover, how does the proposed method compare with a simple parallel combination of confidence and IoU steering?

  2. According to Section 3.4, only the final IoU direction is applied during test-time inference, while confidence intervention is used during vector construction. Why is the confidence intervention not retained during inference?

  3. I also suggest that the authors provide visualization examples comparing the predicted temporal boundaries of the baseline, pure IoU intervention, and the proposed method against the ground truth. Both successful and failure cases would help illustrate how the proposed intervention improves temporal localization. In addition, visualization of the hidden representations before and after confidence intervention could provide further insights into the effectiveness of the proposed cascaded design.

5 comments

W1: Efficiency analysis and comparison with conventional VTG methods. Thank you for highlighting these omissions. We have expanded Table 1 to include representative conventional VTG methods: TR-DETR, UVCOM, Sim-DETR, and SE-DETR+. Their published results are included below alongside the existing comparisons.

Updated Table 1. Comparison on Charades-STA (C), ActivityNet-Captions (A), and QVHighlights (Q). All values are percentages. A dash indicates an unreported result.

MethodC: [email protected]C: [email protected]C: [email protected]C: mIoUA: [email protected]A: [email protected]A: [email protected]A: mIoUQ: [email protected]Q: [email protected]Q: [email protected]Q: mIoU
Conventional VTG methods
TR-DETR (2024)—57.6133.52——————64.6648.96—
UVCOM (2024)—59.2536.64——————63.5547.47—
Sim-DETR (2025)73.0961.3439.6252.56—————67.6450.91—
SE-DETR+ (2026)71.1060.2039.7051.20————————
Frozen backbone baselines and our method
InternVideo2.5-Chat-8B54.8930.4811.2434.2522.4611.465.0517.1547.5528.9713.6133.09
+ C2I-CasITI-VTG58.2633.2711.9736.0626.4214.146.5419.5154.9736.9018.5238.14
InternVL3-8B33.7112.905.4025.3421.5911.316.4517.6951.2932.1919.6138.06
+ C2I-CasITI-VTG64.4934.5712.1838.9142.3224.4112.3029.5966.8442.5825.1045.91
Qwen3-VL-8B69.6045.1319.0143.9249.0832.9520.0335.7767.7454.9742.3253.09
+ C2I-CasITI-VTG75.0352.2623.3948.0352.1635.2321.0337.6968.4556.1342.9053.72
Qwen3.5-9B79.7062.6332.7253.9353.0236.3922.5638.8379.5567.0355.3564.14
+ C2I-CasITI-VTG82.1267.1541.1657.4755.5637.8623.6540.1981.8769.2957.1666.10

Conventional methods use their respective task-specific training settings. The reported TR-DETR, UVCOM, and Sim-DETR results use SlowFast+CLIP video features and CLIP text features, while SE-DETR+ uses multi-layer CLIP video features (K=4) and CLIP text features.

For efficiency, we measured Qwen3.5-9B with and without our intervention on the same Charades-STA samples using a single NVIDIA RTX 3090 GPU:

MetricWithout interventionOurs
Mean latency (s/sample)3.07773.0790
Mean generated tokens16.2616.26
Throughput (tokens/s)5.28365.2815
Mean per-sample peak memory (GiB)18.778418.7784
Maximum peak memory (GiB)20.170720.1707

At inference time, our method simply adds a precomputed vector to the model’s hidden states, without additional model passes. The measured latency increase is only 0.040%, while generation length and peak memory remain unchanged at the reported precision. Offline vector construction requires two activation-extraction rounds using forward passes, without training or backpropagation.

The MLLM backbone remains the main source of computational and memory cost. These results show that our method improves localization while maintaining essentially the same inference latency and GPU memory usage as the backbone. We will include these efficiency measurements in the revision.

W2: Confidence estimation and its role in the cascade. Thank you for raising this important distinction.Our method estimates confidence through uncertainty in timestamp generation, using weighted entropy computed from the model’s internal token probability distributions. The decrease from 53.93 to 44.66 mIoU under confidence-only intervention illustrates why confidence is used as an auxiliary signal in our method, rather than the final localization objective.

To quantitatively examine the relationship between confidence and localization quality, we computed the Spearman correlation between negative weighted entropy and IoU on 7,000 paired samples. After confidence intervention, the correlation increases from 0.197 to 0.258, indicating a modest improvement in their alignment. We also visualize 500 randomly sampled pairs using a jointly fitted PCA, which shows a shift in the activation distribution after intervention. Together, these observations support the role of confidence intervention in conditioning representations for subsequent IoU-vector construction.

Our aim is to link confidence to the model’s own timestamp-generation behavior, so that confidence and the behavior it describes remain consistent. We therefore need to obtain confidence from signals provided by the model itself. Two common approaches are to derive confidence scores from token prediction probabilities or log-probabilities, or to prompt the model to report its confidence directly, known as verbalized confidence. Our weighted entropy follows the first approach, using uncertainty in timestamp generation to measure confidence.

To evaluate verbalized confidence, we first prompted the model to generate a temporal interval together with its confidence. Across the tested 100-point, 5-point, and binary scales, the model consistently reported maximum confidence, providing little distinction between predictions. We therefore adopted a two-pass self-evaluation procedure: the model first predicts the interval and then assesses its confidence in that prediction in a separate pass.

We used these self-reported scores to construct the confidence direction and select high-confidence samples for subsequent IoU-vector construction. The results on Charades-STA using Qwen3.5-9B are reported below; all metrics are percentages.

Confidence estimatorInterventionmIoU[email protected][email protected][email protected]
—Backbone53.9379.7062.6332.72
Two-pass self-evaluationConfidence → IoU56.7981.2967.0239.09
Weighted entropyConfidence → IoU (Ours)57.4782.1267.1541.16

With either estimator, the confidence-to-IoU cascade outperforms the corresponding confidence-only intervention. The cascade using weighted entropy achieves 57.47 mIoU, compared with 56.79 using two-pass self-evaluation. Weighted entropy therefore yields better localization performance in this comparison and eliminates the second inference pass needed for self-evaluation.

These results support weighted entropy as a useful confidence measure for our framework: it provides a model-derived signal for vector construction, achieves better localization performance than the tested self-evaluation alternative, and avoids an additional inference pass.

Q1: Parallel combination of confidence and IoU steering. Thank you for suggesting this comparison. We evaluated a weighted-sum baseline on Charades-STA using Qwen3.5-9B, combining the independently constructed IoU and confidence vectors into a single intervention vector. We swept the IoU:confidence ratio from 1:0 to 0:1 in steps of 0.1 under otherwise identical settings. All metrics below are percentages.

IoU:ConfmIoU[email protected][email protected][email protected]
1.0:0.056.1380.9165.6738.12
0.9:0.156.0980.9465.1137.72
0.8:0.255.7480.8664.8937.07
0.7:0.355.0780.4664.1135.75
0.6:0.454.2679.9762.7433.87
0.5:0.552.9978.7161.3731.59
0.4:0.651.8277.9859.1129.22
0.3:0.750.1976.5355.9726.64
0.2:0.848.8775.5153.2524.84
0.1:0.947.4274.1951.0822.61
0.0:1.044.6671.9144.9218.12
Ours57.4782.1267.1541.16

The best mIoU in the sweep occurs at the pure-IoU endpoint. Increasing the confidence weight progressively reduces mIoU, [email protected], and [email protected], showing that adding the independently constructed confidence direction can weaken localization performance under the tested settings.

These results provide empirical support for our concern about combining the two steering objectives directly. Our cascade instead uses confidence to condition the representations from which the IoU direction is constructed, achieving 57.47 mIoU. We will include these results in the revision to show why we use confidence during vector construction instead of combining it directly with IoU steering.

Q2: Why is confidence intervention not retained at inference? Confidence intervention is used to condition the representations from which the final IoU direction is constructed. The deployed direction is then explicitly guided by localization quality, so confidence does not need to be applied again as a separate steering signal.

We directly tested retaining confidence intervention during prefill while applying the constructed IoU direction during decoding. Results on Charades-STA using Qwen3.5-9B are shown below; all metrics are percentages.

Test-time interventionmIoU[email protected][email protected][email protected]
Backbone53.9379.7062.6332.72
Confidence during prefill + final IoU direction during decoding29.8347.7228.7911.21
Final IoU direction only (Ours)57.4782.1267.1541.16

Retaining confidence intervention substantially reduces localization performance in this experiment. One possible explanation is that the two interventions shift the model’s representations in different directions: confidence steering changes the prefill representations, while the subsequent IoU steering shifts the decoding representations toward localization quality. These successive shifts may interfere with each other and weaken the final prediction. Confidence first conditions the representations used for vector construction, while IoU guides the final intervention toward the downstream localization objective. We therefore apply confidence intervention only during vector construction and retain a single IoU direction at inference.

Q3: Qualitative examples and representation visualization. We sincerely thank the reviewer for these helpful suggestions. We have added qualitative case studies to the appendix and will expand them to compare the baseline, pure IoU intervention, and our method against the ground truth, including both successful and failure cases. We will show the predicted temporal boundaries and corresponding IoU values to make the improvements and remaining errors easier to inspect.

We also appreciate the suggestion to visualize hidden representations before and after confidence intervention. We will include this analysis to help illustrate how confidence intervention changes the representations used for subsequent IoU-vector construction. We will present these visualizations as supporting observations, alongside the quantitative ablations, to provide a clearer understanding of the cascaded design.

Thank you again for the constructive suggestions, which help us better communicate both the effectiveness and limitations of our method.

Reviewer_4pfNOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies the mismatch between model confidence and temporal IoU, arguing that IoU-only activation contrasts can produce noisy steering directions and that naively combining independently estimated confidence and IoU vectors may cause interference. It proposes C2I-CasITI-VTG, a training-free method that applies confidence intervention during prefill when constructing the steering vectors, then estimates an IoU direction from the resulting high-confidence decoding activations. At test time, only the final IoU direction is applied during decoding. Experiments show improvements across multiple benchmarks and frozen MLLMs, with ablations supporting the proposed construction.

Critical strengths

1. Simple and effective design.

The method improves temporal grounding without updating backbone parameters. Consistent gains across three benchmarks and multiple models, including cross-dataset transfer of the same steering vector, demonstrate its practical usefulness.

2. Supporting ablation studies.

The ablations examine confidence steering, IoU steering, and high-confidence filtering, providing evidence that the full cascaded construction outperforms the evaluated alternatives.

Critical weaknesses

1. The benefit of the confidence-first cascade needs stronger validation.

In Qwen3.5-9B, simpler IoU-based interventions already account for much of the improvement: pure IoU and single-stage high-confidence IoU achieve 56.13 and 56.44 mIoU, respectively, compared with 57.47 for the full method. Although the gains are larger on other backbones, further controlled comparisons are needed to establish why confidence-first construction is preferable to alternative filtering and steering strategies.

2. Limited evaluation of robustness beyond the standard benchmarks.

The three benchmark evaluation supports effectiveness, but robustness to annotation quality and different video settings remains insufficiently explored. Evaluation on TimeLens-Bench’s refined annotations and Ego4D-NLQ’s egocentric videos would strengthen the generalization claims.

Questions for authors
  1. Compare the proposed confidence-to-IoU cascade with two alternatives: (i) IoU-to-confidence, applying IoU-based steering during prefill and confidence-based steering during decoding; and (ii) IoU-to-IoU, applying IoU-based steering during both prefill and decoding. These comparisons would clarify whether the confidence-first ordering is important and whether confidence steering provides benefits beyond an IoU-only cascade.

  2. Evaluate on TimeLens-Bench and Ego4D-NLQ, including the pure IoU and single-stage high-confidence IoU baselines, to test whether the cascade’s additional benefit generalizes.

  3. Clarify whether the method can extend to one-to-many grounding, such as OMTG Bench, where a query corresponds to multiple disjoint temporal segments. Explain how confidence aggregation and IoU-based vector construction would be adapted.

6 comments

W1&Q1: Cascade ordering and objective choice. Thank you for suggesting these comparisons. We evaluated IoU-to-confidence and IoU-to-IoU on Charades-STA using Qwen3.5-9B. For IoU-to-confidence, the second-stage vector points from low-confidence to high-confidence activations. We tested its construction both with and without high-IoU sample filtering. All metrics below are percentages.

MethodmIoU[email protected][email protected][email protected]
Backbone53.9379.7062.6332.72
Pure IoU56.1380.9165.6738.12
Single-stage high-confidence IoU56.4481.3665.9038.76
IoU → Confidence (with high-IoU filtering)47.3774.8449.9722.34
IoU → Confidence (without high-IoU filtering)46.0972.9847.3720.46
IoU → IoU54.3180.0362.7734.46
Confidence → IoU (Ours)57.4782.1267.1541.16

Both IoU-to-confidence variants reduce localization performance. This is consistent with the distinction between confidence and localization accuracy: the confidence direction targets more confident predictions, but does not directly target accurate temporal boundaries. Even within high-IoU samples, contrasting low- and high-confidence activations does not explicitly distinguish better from worse localization. Confidence therefore serves as an auxiliary signal for conditioning representations in our method, rather than the final steering objective.

IoU-to-IoU also falls below pure IoU intervention, despite its second-stage vector being re-estimated from the representations obtained after the first-stage intervention. Repeating the IoU objective may reinforce information already captured by the first stage, limiting the additional benefit.

Together, these results support the complementary roles of the two stages under the evaluated settings: confidence first conditions the representations used for vector construction, while IoU guides the final intervention toward the downstream localization objective.

W2&Q2&Q3: Evaluation on refined annotations and egocentric videos. Thank you for this helpful suggestion. We additionally evaluated Qwen3.5-9B on the 3,363 samples of Charades-TimeLens, including both requested IoU-based baselines. All metrics below are percentages.

MethodmIoU[email protected][email protected][email protected]
Backbone50.8573.5153.7630.48
Pure IoU50.8973.7154.5031.52
Single-stage high-confidence IoU50.9273.6855.8731.82
Ours51.0273.7156.1130.87

Our method achieves the highest mIoU and [email protected], but the additional mIoU gain is small, and [email protected] is lower than that of single-stage high-confidence IoU. These results indicate that the cascade’s additional benefit is limited on this refined benchmark.

To better understand this limitation, we examined the duration distributions of ground-truth and predicted intervals:

SourceTotal intervalsIntervals ≤2 sProportion ≤2 s
Charades-STA ground truth3,720180.48%
Charades-TimeLens ground truth3,3631,29938.63%
Qwen3.5-9B backbone predictions3,3631364.04%
Predictions after our intervention3,3632547.55%

Charades-TimeLens contains substantially more short ground-truth intervals, requiring more precise temporal boundaries. For example, if a prediction fully contains a 2-second ground-truth interval but extends it by 1 second in total, its IoU is only 2/3, already below the 0.7 threshold.

The backbone produces far fewer short intervals than appear in the refined annotations. Our intervention increases their proportion from 4.04% to 7.55%, but a substantial gap remains relative to the ground truth’s 38.63%. This suggests that the intervention only partially addresses the model’s tendency to predict longer intervals. This tendency may reflect a duration preference learned during training, although the present analysis does not establish its cause. Addressing it more fully may require stronger adaptation, such as fine-tuning, rather than activation steering alone.

We sincerely apologize that, due to the limited time and current experimental progress, we have not yet completed evaluations on Ego4D-NLQ or OMTG Bench. We will prioritize these experiments and aim to provide the results within the discussion period.

Thank you again for the constructive suggestions, which help us better communicate both the effectiveness and limitations of our method.

Dear Reviewer,

Thank you for your review. The authors have submitted their rebuttal. Could you please take a moment to review their responses?

Thank you for your time and contribution!

Best regards, AC

Thank you for the additional experiments. However, the benefit of the proposed cascade over simpler IoU-based steering remains limited, particularly on Charades-TimeLens. The shorter GT intervals do not fully explain this gap; please provide a performance breakdown by GT interval duration and further clarify applicability to one-to-many grounding. I am inclined to lower my rating from 3 to 2, but remain open to reconsideration if additional results address these concerns.

Reviewer_pMtXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

For inference-time intervention (ITI) in video temporal grounding, direct IoU-based approaches will face the problem: confidence and IoU are often mismatched, and a weighted sum requires an additional balancing coefficient. To address these questions, the authors first use confidence-based ITI during prefill to modify the model’s confidence behavior; based on these confidence-aligned representations, they then extract an IoU-based steering direction to improve localization. Experiments on three benchmarks show consistent improvements.

Critical strengths
  • S1: The setting is practical and low-cost.
  • S2: The motivation is reasonable, and the corresponding method shows improvement compared to the backbone.
Critical weaknesses
  • W1: It seems the paper does not compare with the weighted-sum intervention baseline.
  • W2: Compared with the "Only high-confidence IoU" and the weighted-sum baseline, what is the extended time cost for different models?
Questions for authors
  • In Table 3, the gain from "Only high-confidence IoU" to the proposed method is much smaller for Qwen3.5-9B than for the other two backbones. Maybe for stronger models, this method is not that necessary.
  • Better show some case studies.
3 comments

W1: Weighted-sum baseline. We added this comparison on Charades-STA using Qwen3.5-9B, sweeping the IoU:confidence ratio from 1:0 to 0:1 in steps of 0.1 under otherwise identical settings. All metrics below are percentages.

IoU:ConfmIoU[email protected][email protected][email protected]
1.0:0.056.1380.9165.6738.12
0.9:0.156.0980.9465.1137.72
0.8:0.255.7480.8664.8937.07
0.7:0.355.0780.4664.1135.75
0.6:0.454.2679.9762.7433.87
0.5:0.552.9978.7161.3731.59
0.4:0.651.8277.9859.1129.22
0.3:0.750.1976.5355.9726.64
0.2:0.848.8775.5153.2524.84
0.1:0.947.4274.1951.0822.61
0.0:1.044.6671.9144.9218.12
Ours57.4782.1267.1541.16

Weighted-sum intervention on Charades-STA using Qwen3.5-9B.

The best mIoU in the weight sweep occurs at the pure-IoU endpoint, and increasing the confidence weight progressively reduces mIoU. Our method achieves 57.47 mIoU%, supporting the use of confidence to condition representations for IoU-vector construction rather than mixing the two directions.

W2: Computational cost. At inference time, our method adds a precomputed vector to the model's hidden states. As shown below, this has almost no effect on inference latency or GPU memory usage. We measured both settings on the same Charades-STA samples using Qwen3.5-9B.

MetricWithout interventionOurs
Mean latency (s/sample)3.07773.0790
Mean generated tokens16.2616.26
Mean per-sample peak memory (GiB)18.778418.7784
Maximum peak memory (GiB)20.170720.1707

Inference cost on Charades-STA using Qwen3.5-9B.

Latency increases by only 0.040%, while generation length and peak memory remain unchanged at the reported precision. Weighted-sum, “Only high-confidence IoU,” and our method all apply a single precomputed vector at inference time and therefore have the same deployment cost in terms of model operations.

The additional cost is incurred during offline vector construction: weighted-sum requires one activation-extraction round, whereas our method requires two sequential rounds, approximately doubling the activation-extraction time. This does not add any extra model passes during inference.

Q1: Smaller gains on stronger backbones. Thank you for highlighting the difference in gains across backbones. Stronger baselines often leave less room for further improvement, which may partly explain the smaller mIoU gain on Qwen3.5-9B. Nevertheless, compared with “Only high-confidence IoU,” our method improves Qwen3.5-9B from 56.44 to 57.47 mIoU and, more notably, from 38.76 to 41.16 at [email protected]. This gain of 2.40 percentage points shows that the method remains beneficial when precise temporal localization is required.

We also believe the method is valuable for the study and practical use of smaller models. Vector construction uses standard forward passes without additional training or backpropagation, while deployment requires only a single vector addition with almost no extra inference cost, as reported in our response to W2. These properties may also make the method useful for local model deployment and resource-constrained settings, such as edge devices.

Q2: Case studies. We sincerely thank the reviewer for this helpful suggestion. We have added case studies to the appendix of the revised manuscript to illustrate how our method affects temporal predictions. We hope these examples help readers better understand its effects on localization.

We sincerely thank the reviewer for the thoughtful questions and constructive suggestions, which have helped us improve the paper.

Dear Reviewer,

Thank you for your review. The authors have submitted their rebuttal. Could you please take a moment to review their responses?

Thank you for your time and contribution!

Best regards, Area Chair

We sincerely thank the reviewers for their time, careful evaluations, and constructive suggestions. The feedback has helped us strengthen the experimental analysis and clarify the contribution of our work.

In response, we have provided additional controlled comparisons, benchmark evaluations, and efficiency measurements. We hope these results help demonstrate the value of our method, as well as its practical benefit of improving frozen backbones without parameter updates and with almost no additional inference cost.

We appreciate that time for further discussion may be limited. When finalizing your assessments, we would be grateful if you could consider the additional evidence and clarifications provided in our responses, and reconsider your scores where appropriate. We fully respect your independent judgment and sincerely appreciate your consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.