Revisiting Logit Distributions for Reliable Out-of-Distribution Detection

W1: The biggest issue may be that the method itself is very trivial and lacks novelty. Research around logit magnitudes can be traced back to the last century, and the concept of logit gap proposed in this paper has similar concepts that were introduced as early as 2001 (Margin Sampling).

We agree that the general idea of utilizing logits has a long history, our method makes a novel contribution in the specific context of OOD detection:

• Revisit Logit Distribution in OOD Detection. Traditional methods (e.g. MaxLogit and MSP) use only the maximum logit or probability as OOD score, overlooking the rich information in non-maximum logits. Recent Energy-based methods instead link high energy to low likelihood, but they struggle in zero-shot (CLIP-based) settings and fail to explicitly model the logit distribution. In contrast, our method revisits the full logit distribution and proposes a novel logit-gap-based scoring function that better distinguishes ID from OOD samples.
• Different Context and Objective. While Margin Sampling also involves the concept of logit gap, it was designed for active learning to identify uncertain samples for labeling. In contrast, our LogitGap leverages the logit gap to score and detect OOD samples, particularly from unseen classes or styles. To our knowledge, this is the first work to adapt the logit gap specifically for OOD detection in zero- and few-shot settings.
• Simple but effective. Although the proposed method is conceptually simple, it consistently demonstrates strong performance across various OOD detection tasks, including zero-shot, few-shot, hard and traditional settings.

---

W2: The comparisons in the method section are not comprehensive enough.

W2.1: Zero-shot comparison baselines in Tables 1 and 2 are from three years ago.

In response, we have added two recent zero-shot OOD detection methods [1,2]. See response to reviewer L4Ge's Weakness 2.1(W2.1) for details.

• Training-free, post-hoc simplicity: Many recent zero-shot OOD methods require additional training data [3,4] or external models [5]. Differently, LogitGap is a simple, training-free, post-hoc method directly applicable to pre-trained CLIP models.
• Specialized for vision-language models: Most post-hoc OOD methods target unimodal visual backbones (e.g., ResNet), and often struggle with vision-language models like CLIP, underscoring the need for our specialized approach.

W2.2: Effectiveness when combined with other methods in Table 4 are negative.

Table 4 evaluates the compatibility of LogitGap with existing fine-tuning methods. By simply replacing the OOD score, LogitGap improves performance on 9 out of 12 average metrics. The largest performance drop is under 0.2%, indicating minimal downside. These results suggest that LogitGap can serve as a drop-in replacement in existing OOD frameworks, offering consistent gains with negligible risk.

W2.3: Comparison in Table 5 did not include several top-ranking works from the OpenOOD.

• Training-Free & Data-Free Advantage: Unlike top OpenOOD methods (e.g., OE-based approaches) that require large-scale data and training, LogitGap operates in a fully training-free and data-free manner.
• Fair Comparison with Post-Hoc Baselines: We benchmark against training-free post-hoc methods (e.g., SHE, NCI) in Table 5, where LogitGap shows superior performance.
• Compatibility with Training-Based Methods: LogitGap can also enhance training-based approaches—when combined with OE, performance further improves, demonstrating its versatility in OOD detection pipelines.

---

W3.1: The ablation study on hyperparameter N with only one figure (Fig 3) is insufficient. It would be better to provide more comprehensive data.

In below table, we provide a more ablation study on hyparameter $N$. Specifically, we conduct experiments under the zero-shot OOD detection setting, using either ImageNet-100 or ImageNet-1K as ID dataset. For simplicity, we report the FPR95 of our LogitGap method based on CLIP ViT-B/16 model. We can observe that: (1) Optimal $N$ varies by dataset (e.g., 19 for ImageNet-100, 195 for ImageNet-1K); (2) Setting $N$ to 20% of total classes consistently provides strong performance. Therefore, we adopt this value as default in LogitGap.

W3.2: Whether synthetic OOD data aligns with real OOD distributions, and the impact of the alignment degree on final performance, need further discussion.

• As stated in the paper, we propose a synthetic OOD sampling scheme to adaptively select $N$. Importantly, LogitGap also performs well without synthetic OOD, by simply setting $N$ to a fixed proportion (e.g., 20% of the number of classes).
• We observe that while synthetic data may not fully capture real OOD characteristics, it is effective for selecting a suitable $N$ for LogitGap*. Please see our response to Reviewer BQ1B's Weakness 3 (W3) for more details.

---

Minor W1. The title of Table 2 in the supplementary material is incorrect.

Thank you for pointing this. We will fix it.

[1] GL-MCM: Global and Local Maximum Concept Matching for Zero-Shot Out-of-Distribution Detection, IJCV 2025.

[2] TAG: Text Prompt Augmentation for Zero-Shot Out-of-Distribution Detection, ECCV 2024.

[3] CLIPN for Zero-Shot OOD Detection: Teaching CLIP to Say No, ICCV 2023.

[4] Zero-Shot Out-of-Distribution Detection Based on the Pre-trained Model CLIP, AAAI 2022.

[5] Envisioning Outlier Exposure by Large Language Models for Out-of-Distribution Detection, ICML 2024.

W1: The introduction section is poorly written.

W1.1&1.2: Energy-based method is the most representative OOD score, which also utilizes information from non-maximum logits. Could the authors clarify their reasoning for overlooking this method?

We would like to clarify that the energy-based method differs significantly from the other methods mentioned in the introduction, such as MaxLogit, MSP, and LogitGap. The energy-based method distinguishes between ID and OOD samples based on likelihood density, with the idea that high-energy points correspond to low likelihood density, making them more likely to be OOD samples. In contrast, the other methods focus on the logit pattern, with the idea that ID samples typically produce more peaked logits, while OOD samples corresponds to more uniform logit patterns. For this reason, we chose not to discuss energy-based methods in the introduction. In the experments, we have compared energy-based methods with our LogitGap, and results show that LogitGap outperforms the energy score across multiple benchmarks (see Table 1, 2 and 3). In addition, while energy-based method utilizes the non-maximum logits, we observe notable limitations of it from experimental results:

• Energy-based method is not compatible with zero-shot OOD scenarios (see Tables 1, 2, and 3) when using a pre-trained CLIP model. We attribute this to different training losses: CLIP model is trained using an image-text contrastive loss, which differs significantly from the NLL loss commonly assumed in energy-based methods. As a result, energy-based methods perform poorly in this setting, since high energy values do not necessarily correspond to low likelihood densities during inference.
• The log-sum-exp formulation is numerically dominated by the largest logit, which makes it behave similarly to MaxLogit in many cases.

W1.3: The caption of Figure 1 lacks information about datasets and models used. Additionally, the comparsion with energy-based method should be included in Figure 1.

Thanks to point out this problem. The results in Figure 1 are reported on the zero-shot OOD scenario with CLIP ViT-B/16 model. In this experiment, categories from ImageNet-20 are used as ID, while those from ImageNet-10 are used as OOD. For the comparison with energy-based method in Figure 1, we will revise Figure 1 in our updated version.

W2: The experiments are not sufficient.

W2.1: Missing baselines like GL-MCM and TAG in zero-shot OOD detection.

We report the results of GL-MCM and TAG in the table below. The experiments are conducted using CLIP ViT-B/16, with ImageNet categories as ID classes and categories from other benchmarks as OOD classes. As shown in Table 2.1, LogitGap achieves comparable or better performance than the two methods on average across multiple benchmarks.

W2.2: The performed OOD benchmark is not standard. Evaluation for far-OOD detection is missing.

Our benchmark design is inspired by recent observations [1], which highlight that existing OOD datasets often mix covariate and semantic shifts, making it difficult to isolate and accurately measure a model’s ability to detect true semantic shift. To address this, we construct a more challenging benchmark using NINCO, ImageNet-O, and ImageNet-OOD to emphasize semantic shift under a high-coverage, diverse, and realistic setting.

Moreover, we have followed the OpenOOD protocol and added results on the far OOD scenario in Table 2.2 and included a comparison with the SSB_Hard dataset in Table 2.1 with ImageNet as ID set. These additions confirm LogitGap's robustness under a broader distributional shift.

W2.3: Missing baselines like ViM and GEN in traditional OOD detection.

We report the results of ViM and GEN in the table below. The experiments are conducted in traditional OOD detection setting, using CIFAR-10 as ID dataset. Specifically, since VIM requires training set statistics, we randomly selected 100 samples from the training set for estimation. In Table 2.3, LogitGap remains competitive, further validating its general applicability.

W2.4: Is there any reason that the authors switch from the large OOD benchmark (ImageNet-1k as ID dataset) to small OOD benchmark (CIFAR-10 as ID dataset)?

Thank you for this question. We report results on the small-scale OOD benchmark to demonstrate the generality of our method. We also conduct experiments on a larger OOD benchmark (CIFAR-100 as ID dataset) in the Appendix C.5. Furthermore, we additionally report results on ImageNet-1k with ResNet50 under the traditional OOD setting in following table, using SSB-Hard and NINCO as Near-OOD datasets, and iNaturalist, Textures, and OpenImage-O as Far-OOD datasets. Due to space limitations, we report the average performance of the method across all datasets. As shown in the table, our LogitGap achieves comparable or better performance than other methods across different benchmarks, and obtains the best average performance overall.

W2.5: For traditional OOD detection, the results across different models such as ViT and BiT are also expected to present.

Here, we report additional results in traditional OOD setting with different models, including Swin-T (Table 2.5.1) and ViT-B/16 (Table 2.5.2), to highlight the method's generalization. In experiments, we utilize ImageNet-1k as ID dataset and other benchmarks as OOD datasets. As shown in Table 2.5.1 and 2.5.2, our LogitGap achieves better performance than other comparative methods across different models.

W2.6: For zero-shot OOD detection, it is expected to evaluate on the other CLIP variants, including ViT models and ResNet models.

We extend our experiments to other CLIP variants including ViT-L/14 and ResNet-101. Results in Table 2.6 demonstrate that LogitGap generalizes well across architectures. As shown in the table below, our method consistently outperforms baselines across all backbones, indicating strong generalization. These results have been included in the revised version of the paper.

W2.7: To support the claim in Line 159 -161, a more comprehensive evaluation across different ID/OOD datasets with different models should be done.

To further support our claim, we report statistical information on the logits distribution across multiple datasets (see our response to Reviewer 4oYE’s Comment W1). These results show that the observed logit pattern is consistently present across diverse datasets.

MW1: What datasets and models are utilized to produce Figure 2?

Figure 2 was produced using the CLIP ViT-B/16 model, with ImageNet100 as the ID and iNaturalist as the OOD dataset.

MW2: Compare LogitGap with the energy score using top 10% logit information.

We have compared LogitGap with top-10% energy scoring and found LogitGap achieves consistently better results (see Table 2.3, 2.4), validating its stronger separability for OOD detection.

MW3: Experimental setup on ImageNet-10/ImageNet-20 is not particularly challenging.

We follow the experimental setup from previous works [2], alternately using ImageNet-10 and ImageNet-20 as the ID and OOD datasets to create a more challenging scenario. While existing methods achieve high AUROC scores in this setting, obtaining a low FPR95 remains difficult. For example, LogitGap reduces FPR95 by 8.40% compared to MCM, highlighting its advantage in handling subtle distribution shifts.

[1] IMAGENET-OOD: Deciphering Modern Out-of- Distribution Detection Algorithms, ICLR 24

[2] Delving into Out-of-Distribution Detection with Vision-Language Representations, NeurIPS 22

W1.1 The theoretical motivation ... are not fully verified or supported.

• Our theoretical motivation stems from the observed distinction in logit patterns between ID and OOD samples: (i) Higher maximum logit values for ID samples; (ii) Higher non-predicted class logits in OOD samples.

• To verify the generality of this phenomenon, we conduct experiments across various ID/OOD datasets and model architectures. Specifically, we report the average maximum logit, and the average logit of non-predicted classes, using ResNet-50, ViT-B/16, and ViT-L/14 as backbone models. As shown in the table below, the results support our hypothesis.

---

W1.2: Figure 2 lacks critical context — it is unclear under what setting it was generated.

Figure 2 was generated using the CLIP ViT-B/16 model, with ImageNet100 as ID dataset and iNaturalist as OOD dataset. The logit curves represent the average logits computed over the entire test set, offering a generalizable summary of the model's behavior rather than illustrating a specific example.

---

W2 & L3. Limited generalization beyond vision classification tasks.

While our method is primarily designed for visual classification (a core OOD detection scenario), we also demonstrate its broader applicability. We evaluate it on audio classification using ESC-50 dataset, where 50 classes are randomly split into 25 ID and 25 OOD classes. The method is applied with the CLAP [3] model, following the original hyperparameter strategy with $N$ set to 20% of the class count ($N = 5$).

As shown in the table below, our method is also effective beyond vision. Crucially, our method does not rely on modality-specific features or architectures, making it easily portable to other domains.

---

W3 & L2. Excessive reliance on supplementary material for theoretical explanation.

We agree that adding clearer intuition in the main text would improve accessibility and we will revise the main paper accordingly. Specifically:

• We will provide a clearer explanation for FPR: False Positive Rate (FPR) measures the likelihood of an ID sample being misclassified as OOD. A lower FPR indicates better OOD detection performance.
• we will expand the explanations for Theorem 4.1: Our analysis (Theorem 4.1) shows that LogitGap achieves a lower FPR than MCM when the temperature $\tau$ exceeds a threshold $z'_1\cdot \frac{K-1}{\ln K}$. This implies that LogitGap better preserves discriminative logit information, avoiding the over-smoothing effects introduced by softmax in MCM.

Thus, the lower FPR of LogitGap confirms its effectiveness in distinguishing ID from OOD samples compared to MCM.

---

L1: Motivation Example in Section 4.1 may be unintuitive.

We agree that the current example might be unintuitive from a classification perspective. In the revision, we will clarify this point by rephrasing Line 139-141 as follows:

"Since OOD samples tend not to show confident alignment with any known class, sample z2 is more likely to be in-distribution than sample z1. However, the maximum softmax probability for both is identical: ..."

---

Q1. The choice subtraction-based formulation in LogitGap needs further justification, with experimental comparisons to alternative designs like weighting the MCM score or other aggregation strategies to strengthen the claim of its effectiveness.

We subtract the average of remaining logits from the maximum logit because ID samples typically show larger maximum logits and smaller remaining logits than OOD samples. This pattern enhances the ID/OOD distinction, improving detection performance.

We evaluate alternatives on CLIP ViT-B/16 with ImageNet as ID data:

1. MCM_MaxLogit, a weighted combination of MCM and MaxLogit, which improves upon both but is still outperformed by LogitGap (see table below).
2. Two LogitGap variants—LogitGap_exp and LogitGap_sqrt—applying exponential and square transformations to amplify score differences. Results show these variants perform comparably to the original LogitGap, indicating different transformations can yield similarly effective discrimination. These comparisons will be included in the revised manuscript.

---

Q2. A fair evaluation of MCM should consider N values derived similarly to LogitGap’s.

To address this concern, we introduce MCM_topN, a comparative baseline that applies the top-N strategy to MCM. For fair comparison, we use the same $N$ for MCM_topN and our LogitGap. Experiments on CIFAR-10 (ID dataset) show that applying the top-N strategy to MCM does not consistently improve performance. Moreover, LogitGap outperforms both MCM and MCM_topN across all 6 benchmarks (see table below).

[1] A Baseline for Detecting Out-of-Distribution Examples in Image Captioning. ICLR 2017.

[2] Delving into Out-of-Distribution Detection with Vision-Language Representations. NeurIPS 2022.

[3] CLAP: Learning Audio Concepts From Natural Language Supervision. ICASSP 2023.

We sincerely thank the reviewer for the insightful comments. Below, we address each concern.

W1. The LogitGap requires the computation for all DNN layers to obtain the logits to determine whether a sample is ID or OOD.

We would like to clarify that OOD detection is typically performed alongside the primary classification task, whose goal is to identify OOD samples without degrading ID classification accuracy [1]. In this context, the logits required for OOD detection are already produced during the standard forward pass of classification. Therefore, our method leverages these existing computations and adds no extra computational cost beyond standard inference.

[1] Generalized Out-of-Distribution Detection: A Survey, IJCV 2024.

---

W2. Determining the correct N is not entirely free, as LogitGap requires access to a small ID validation set, which may be unrealistic in zero-sample deployment scenarios.

We would like to clarify that our proposed method can be applied in the zero-sample scenario. Specifically, our method includes two variants (Lines 230-232): LogitGap (fixed N = 20% of classes) and LogitGap* (adaptive N via ID validation set). As shown in Tables 1 and 2, LogitGap performs comparably to LogitGap*, demonstrating that our method remains effective even without access to a validation set.

---

W3. Relies on synthetic OOD samples generated by noise/interpolation, whose representativeness for real-world OOD samples is unclear and not justified.

Thanks for this valuable comment. We agree that synthetic OOD samples may not fully reflect the characteristics of real-world OOD samples. However, our empirical results show that these synthetic samples are sufficiently informative for selecting a robust hyperparameter $N$.

As shown in the table below, on ImageNet, the optimal $N$ selected using synthetic samples (i.e., $N_{syn}$) is close to the optimal $N$ derived from multiple real OOD datasets (i.e., $N_{real}$), and achieves comparable performance. This demonstrates the practicality and generalizability of our strategy for selecting $N$ without requiring access to real OOD data.

---

Q1: Can LogitGap be applied to features at an earlier layer in the DNN in order to detect an OOD sample without performing the full inference computation?

Detecting OOD samples using only shallow features is challenging. In zero/few-shot OOD detection settings, OOD samples often involves semantic shifts, which are difficult to capture without the high-level abstractions provided by deeper layers [2].

To address this, existing works [3, 4] explicitly leverage deep features to capture high-level semantic cues for OOD detection. However, these methods may incur additional computational overhead as they require calculating feature correlations for each layer. In contrast, our approach remains computationally efficient while still effectively identifying OOD samples.

[2] Full-Spectrum Out-of-Distribution Detection. IJCV 2023.

[3] A simple unified framework for detecting out-of-distribution samples and adversarial attacks. NeurIPS 2018.

[4] Detecting out-of-distribution examples with gram matrices. ICML 2020.