acceptodds
Withdrawn from ICLR 2027

Open-World Zero-Shot Image Classification using an LLM-guided Discovery Loop

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Contrastive Vision-Language models such as CLIP have enabled zero-shot image classification using only the pretrained model and a complete list of category names, vastly lowering labeling costs compared to traditional supervised approaches. However, in practice the full class list may not be available, creating an obstacle to applying CLIP-style models. Vocabulary-free image classification (VIC) approaches this problem by trying to infer the class list from just the images, however VIC is fundamentally ambiguous as images may be classified according to different attributes and granularities, each of them equally valid. We therefore propose Open-World Zero-Shot Image Classification (OWIC) as a well-defined alternative setting in which some category names are given, and the task is to discover the remaining ones based on the pattern defined by the existing classes. To tackle this task, we propose Multimodal Agentic Novelty Detection (MAND), an LLM-driven discovery loop in which a multimodal LLM identifies novel-class images and proposes category names for them that are then used by a contrastive image-text model such as CLIP. A similarity-based stopping criterion helps minimize LLM usage to lower compute cost. Once all category names are discovered the LLM can be discarded, maintaining deployment efficiency. Experimental results show MAND outperforming strong baselines based on vocabulary-free methods adapted to this new task. In addition to leveraging pretrained models, we also show that finetuning can further improve the LLM's performance within the discovery loop, generalizing even to domains beyond the training data. Code will be published.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
category discovery, zero-shot classification, multimodal large language models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_DHwnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Open-World Zero-Shot Image Classification (OWIC), a setting in which only a subset of category names is provided, while the remaining categories must be discovered without sample-level annotations. To address this task, the authors propose Multimodal Agentic Novelty Detection (MAND), which combines a multimodal large language model with a contrastive vision-language model. MAND iteratively identifies images from potentially novel categories, generates category names, and filters duplicate categories using CLIP-based similarity. The framework further incorporates prior correction and a similarity-gap-based stopping criterion to improve category discovery and reduce LLM inference costs. After discovery, the category prototypes are refined, enabling classification without the LLM at deployment. Experiments on ten image classification datasets compare MAND with adapted vocabulary-free classification methods, including CaSED and FiNDR. The authors also provide component ablations, comparisons of stopping criteria, sensitivity analyses, and an optional LLM finetuning pipeline.

Critical strengths
  1. A practically motivated problem formulation. The proposed OWIC setting addresses a meaningful limitation of conventional zero-shot classification, where a complete category vocabulary is typically assumed to be available. Providing partial category information without sample-level annotations is a reasonable setting that may facilitate category discovery in practical applications. The distinction from fully vocabulary-free classification is clearly motivated.
  2. Relatively comprehensive experimental evaluation. The paper evaluates the proposed framework on ten datasets spanning different visual domains and includes comparisons with adapted CaSED and FiNDR baselines. The component ablations and comparisons of stopping criteria provide useful insights into the framework. In particular, the similarity-based stopping criterion achieves competitive classification accuracy while requiring substantially fewer discovery steps than the RichnEst-based alternative.
Critical weaknesses
  1. Limited methodological novelty and insufficient justification of the proposed task formulation. Although OWIC is presented as a new classification setting, its relationship to existing vocabulary-free classification and generalized category discovery frameworks raises questions about the significance of the conceptual contribution. Furthermore, the claim that partial category names make the classification problem well-defined is not fully established, since a limited set of known names may still be consistent with multiple classification rules or levels of granularity. Methodologically, MAND mainly integrates existing techniques, including LLM-based category generation, CLIP-based similarity filtering, prior correction, and prototype refinement. While their integration is reasonable, the paper does not clearly establish a sufficiently distinctive algorithmic contribution beyond this combination of established components.
  2. Insufficient experimental evidence for the necessity and practical advantages of the proposed framework. Although the experiments cover multiple datasets, they do not sufficiently establish whether the proposed iterative discovery pipeline is necessary compared with simpler alternatives. In particular, a direct MLLM-based category generation baseline, followed by CLIP-based filtering and classification, would help clarify the benefits of the proposed discovery loop. Moreover, the evaluation primarily uses a fixed 50% known-category split, without systematically investigating the effects of different known-category proportions or class distributions. The additional benefit of finetuning is also limited: Table 2 reports 75.1% cAcc for the pretrained model and 75.0% for the finetuned model before reference images are introduced, suggesting that the improvement of the final finetuned configuration cannot be attributed to finetuning alone. Finally, the efficiency analysis mainly reports discovery steps rather than end-to-end runtime or computational cost, leaving the practical efficiency advantage insufficiently demonstrated.
Questions for authors
  1. Task formulation: How do the authors justify the claim that partial category names uniquely determine the intended classification rule? Could the authors clarify under what conditions the ambiguity is sufficiently reduced, particularly when only a small or semantically narrow subset of categories is known?
  2. Simpler baselines: How does MAND compare with a simpler approach that directly prompts an MLLM to infer the classification rule and generate candidate categories from a subset of images and the known category names, followed by CLIP-based filtering and classification? Such a comparison would help establish the necessity of the iterative discovery loop.
  3. Finetuning effectiveness: Table 2 shows that the finetuned model without reference images achieves 75.0% cAcc, compared with 75.1% for the pretrained model under the corresponding configuration. Could the authors clarify the independent contribution of finetuning and reference images, and explain why finetuning provides limited benefits without the additional visual context?
  4. Robustness and computational cost: How sensitive is MAND to the proportion of known categories, especially when substantially fewer than 50% are provided? Additionally, could the authors report the end-to-end computational cost, including total LLM calls and runtime, to better substantiate the claimed efficiency advantages?
2 comments

Thank you for your review comments, we will give a general response in the main rebuttal comment, while addressing specific points here.

W1, Q1:

You are correct that our claim regarding task identifiability was too broad and not well-supported, we have revised the paper and included a detailed discussion on task identifiability in Section 3.1. The discussion shows that under some assumptions the known categories can substantially reduce ambiguity, and thus enable discovery. The degree to which ambiguity is reduced depends on the attribute variation among the known categories.

Regarding the conceptual significance of OWIC as a task, we believe that there are many practical tasks that fit OWIC, but not any existing zero-shot setting. Examples include surveying the species that occur in an area from a wildlife camera database, tracking failure types in a factory, or categorizing media libraries. In all these cases, we will usually have an initial category set that we are looking to extend with novel categories that are present in the data. Standard zero-shot classification methods will be unable to discover the novel categories, while VIC methods may not follow the classification scheme of the known classes. OWIC covers the requirements of these problems better than either setting, showing its real-world relevance. Compared to GCD, OWIC differs in that it requires to labeled data, and permits no per-dataset training.

Regarding the methodological novelty, there are several parts of MAND that are novel, including the CLIP-based prior estimation and recall gate, the granularity-adaptive duplicate filter, the similarity-based stopping criterion, and the hierarchical task sampling for the finetuning. As shown in Table 2, these components are critical to the method's accuracy, and especially important for preventing overprediction of novel categories, increasing cluster accuracy from 67.4% to 75.1% and decreasing the mean category error from 157.2% to 19.1%.

W2, Q2, Q3:

The finetuned model outperforms the pretrained one considerably in the baseline configuration (see Table 2). However, adding additional improvements such as the duplicate filter erodes this advantage, showing that these improvements improve the methods along similar lines as the finetuning. Concretely, the improvements mainly address the model's tendency to overpredict categories, whereas this problem is much less severe in the finetuned model, so it already performs well without them. In addition, the finetuned model can make use of the reference images, which the pretrained model cannot do. This enables the finetuned model to leverage more information to solve the task, achieving a novel-class accuracy gain of 3.9% in our main results. Also, we show in our revision that the gap between the pretrained and the finetuned model grows as the novel percentage increases, further strengthening the case for finetuning.

As for the additional simple baseline, we will provide the requested baseline in the second revision.

Q4: We have provided a study in which the known-class percentage is varied between 10% and 90% in the revised paper. The results in Figure 4 show that accuracy declines gracefully as the proportion of known classes decreases. The requested performance benchmarks will be provided in the second revision.

Here is the second part of our response:

Q2: We add a new method designated as MAND SP (single pass) based on your suggestions. MAND SP combines elements of FiNDR and MAND to create a method that processes the dataset in a single pass without a discovery loop, while keeping the filtering and refitting logic of MAND (see appendix C.2 for details). The results in appendix D.2 show that while MAND SP underperforms standard MAND on cluster accuracy, especially on novel classes (mean 65.3% for SP vs 70.2% for MAND pre), it achieves stronger results on semantic accuracy, about half of that driven by a 10-point improvement on EuroSAT. This seems to be a case of MAND SP not proposing many novel classes (1.6 novel classes detected on average for MAND SP vs 6.4 for MAND pre) for a dataset where other methods propose low-similarity names, resulting in MAND SP achieving better sAcc. Compared to FiNDR, MAND SP clearly outperforms on both cAcc and sAcc, showing the impact of our filtering pipeline. Note that we set the discovery set size based on MAND pre step counts to ensure equal budget, however in practice the budget would have to be set blindly, so MAND SP in this form is not practical.

Q4: We provide various performance metrics for MAND and FiNDR in appendix D.10 (CaSED uses no LLM and as a result has negligible cost). The results show that while MAND requires more wall-clock time than FiNDR, the time spent running LLM calls is very similar. A detailed inspection of LLM usage patterns reveals that MAND makes more LLM calls, but generates fewer tokens with them owing to the decision stage that discards known-class samples with a single-token response. Our efficiency claim concerns the stopping criterion relative to alternative criteria and to processing all images. MAND is not cheaper than FiNDR in wall-clock time, but uses comparable LLM time and about 30× fewer generated tokens. As the discovery stage is a one-time cost, the observed wall clock times are not problematic in practice as inference remains cheap regardless of dataset size.

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

The paper introduces Open-World Zero-Shot Image Classification (OWIC), in which a subset of category names is provided and the remaining categories must be discovered from unlabeled images. The proposed Multimodal Agentic Novelty Detection (MAND) loop uses a multimodal LLM to decide whether sampled images are outside the current class list and to name novel categories. A CLIP-like model is used for prior correction, duplicate filtering, a similarity-gap stopping rule, and final prototype-based classification; the LLM can then be removed at deployment. On ten image datasets, MAND outperforms adaptations of CaSED and FiNDR, and LLM finetuning gives additional gains on cluster accuracy.

Critical strengths
  1. The paper identifies a meaningful middle ground between standard zero-shot classification and vocabulary-free classification. Supplying a partial class list is a realistic source of task structure, and the motivation is clearly illustrated.
  2. MAND is a well-motivated modular pipeline: novelty detection and naming are separated from final CLIP-only inference, and the duplicate filter, prototype refitting, prior correction, and stopping rule are tested in ablations.
  3. The empirical study is broad, covering ten datasets and two complementary metrics. MAND is consistently competitive or best among the reported methods, and the comparison of stopping criteria is especially helpful for understanding the efficiency trade-off.
Critical weaknesses
  1. The task definition is not formally identifiable from the stated input. Section 3.1 says that known category names constrain a unique classification rule, but names alone can still support multiple granularities, attributes, or naming conventions. The experiments construct known and novel classes from a chosen dataset hierarchy, which effectively supplies the target rule to the evaluator. The paper should state an explicit identifiability assumption, test cases with ambiguous known lists, and explain how performance should be interpreted when several rules are consistent with the provided names.
  2. The evaluation protocol leaves important sources of bias and uncertainty unresolved. It is unclear whether the ten runs vary only the image sampling order or also the known/novel class split; no standard deviations or confidence intervals are reported. In addition, Table 4 selects the FiNDR discovery-set factor using the evaluation datasets, while the selection procedure for c, theta, and other hyperparameters is not fully specified. Results over multiple random class partitions, unequal known/novel proportions, and a validation-only hyperparameter protocol are needed to establish robustness.
Questions for authors

1 What is the formal condition under which C_known determines a unique target taxonomy and naming convention? Can you give a counterexample where the condition fails, and report how MAND behaves when the known names are compatible with multiple rules? 2 For the main results, are the known classes fixed once per dataset, resampled for each run, or sampled at the episode level? Please report mean +/- standard deviation (or confidence intervals) over both class partitions and image orders, and include experiments with imbalanced known/novel class proportions. 3 Was the FiNDR factor k in Table 4 selected using a held-out validation set? Please state how c, theta, alpha, and any stopping or prompt settings were selected, and rerun the comparison without tuning on the reported test datasets if necessary. 4 How many LLM calls, generated tokens, GPU seconds, and total wall-clock cost does MAND require per episode, compared with each baseline? Is the use of all evaluation images for prototype refitting intended as a transductive setting, and how does performance change under a held-out inductive deployment split? 5 Please specify the sentence transformer, text templates, and aggregation used for semantic accuracy. A direct name-level audit or a contamination-controlled benchmark would help establish that sAcc reflects correct discovery rather than semantic similarity learned from the same web data.

2 comments

Thank you for your review comments, we will give a general response in the main rebuttal comment, while addressing specific points here.

W1, Q1: Thank you for this point, it is indeed correct that the task cannot always be identified from the given known classes, we have revised the paper to include a detailed discussion on task identifiability in Section 3.1. For this, we limit the class of rules to those that divide the data along attribute lines. We show that under some assumptions, the known categories narrow down the space of compatible rules considerably, depending on attribute variation among the known categories. The remaining ambiguity is limited to whether constant attributes are part of the rule or not, and can be resolved from the data if every image has a label under the rule. One example for rules that cannot be identified is mixed-granularity settings, e.g. given Persian cat, Siamese cat, and Dog as known categories it is impossible to infer if an image of a horse should be classified according to species or breed. Experiments with ambiguous known class lists will follow in the second revision.

W2, Q2, Q3: The main experiments are aimed at capturing the mean performance for a single task, so they vary only the sampling order. We have added additional experiments to also show the performance across different known/novel splits (Table 7 and Table 8 in the revised paper). We found a slight accuracy loss on average compared to the main split, which may be due to the main split selecting known and novel by class index. This in some cases results in slightly easier tasks as some class lists are alphabetical, which results in a slight tendency for pairs of similar classes to end up in the same part of the split, reducing difficulty a bit. In addition, we measured the behavior of MAND for different novel class percentages, the results are shown in Figure 4. We also updated Table 5 and Table 6 to include the standard deviations for the main experiments.

Regarding the hyperparameter selection, following FiNDR we report one set of hyperparameters that offers a good trade-off between accuracy and episode steps. To offer a more systematic way of selecting them, we are currently working on a leave-one-out selection study, where we fit c and α\alpha on nine datasets and evaluate the selected configuration on the tenth. To achieve a compromise between accuracy and step count, we select the lowest c that still achieves accuracy within one standard deviation of the optimum. After completing four seeds, we currently observe a mean overall cluster accuracy of 74.7% for the pretrained model, and 76.9% for the finetuned one, compared to 75.0% and 77.0% for the main result on the same seeds. The full results will be included in the second revision. θ\theta on the other hand cannot be fitted in this way as our evaluation suite contains only one dataset with low CLIP recall. We do note that there is a wide band of values for which the accuracy is stable, any value between 0.23 and 0.60 will obtain the same result, showing the selection's robustness.

Q4: We will provide the requested performance benchmarks with our second revision. As for the evaluation protocol, OWIC is indeed intended as a transductive setting, but we will also provide results for an inductive variant as part of the second revision.

Q5: We have added the specifics of the used semantic accuracy to Appendix D.1. In short, we follow the same implementation as FiNDR, meaning we use the model at https://huggingface.co/sentence-transformers/all-mpnet-base-v2, and encode just the category text without a prompt template. Similarities are computed per-image, we then report the mean over all/known/novel category samples. Regarding the name-level audit, we are not quite sure what you mean by this, could you please clarify?

Here is the second part of our response:

Q1: We now provide results for a setting that violates our assumptions in appendix D.9. The task we examine is based on the Stanford Cars dataset and mixes two granularities in its ground-truth labels, and provides examples of either one or both granularities in the known classes. The results show that in this case, when presented with only one granularity in the known classes, MAND will follow it, but when presented with both granularities the method commits to one of them and follows it, disregarding the other. In this case MAND consistently preferred the car model level over the make level.

Q3: We provide the results and the selected hyperparameters for a leave-one-out selection in appendix D.8. We tune both c and α\alpha, while keeping θ\theta constant as discussed previously. To select a good compromise between accuracy and step count, we select the lowest c whose accuracy value is still within one standard deviation of the best. The results show little change in accuracy compared to the main result, showing MAND's robustness to parameter choices. The main difference is in step efficiency, with MAND pre requiring more steps, while MAND ft is much faster under the selected hyperparameters.

Q4: We provide various performance metrics for MAND and FiNDR in appendix D.10 (CaSED uses no LLM and as a result has negligible cost). The results show that while MAND requires more wall-clock time than FiNDR, the time spent running LLM calls is very similar. A detailed inspection of LLM usage patterns reveals that MAND makes more LLM calls, but generates fewer tokens with them owing to the decision stage that discards known-class samples with a single-token response. Our efficiency claim concerns the stopping criterion relative to alternative criteria and to processing all images. MAND is not cheaper than FiNDR in wall-clock time, but uses comparable LLM time and about 30× fewer generated tokens. As the discovery stage is a one-time cost, the observed wall clock times are not problematic in practice as inference remains cheap regardless of dataset size.

We also provide results for an alternative inductive setting in appendix D.7. In this setting, we hold out 50% of samples for each class during discovery, and later evaluate on the held-out data. The results show no change in accuracy compared to the standard transductive setup, showing that MAND does not overfit on the discovery set.

Reviewer_11ioOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a new technical problem called Open-World Zero-Shot Image Classification (OWIC), which is an image classification setting where some of the class vocabulary is known. The core idea here is to be able to use the existing class names to understand the class taxonomy structure and use it to discover new classes.

The paper proposes a cascaded method which first identifies whether the image has a novel category. If so, the category is named, and duplicates are filtered. Over time, the number of undiscovered categories would go down, and therefore, the paper proposes a stopping criterion which encourages the process of discovery to end so that the model does not hallucinate class names.

The proposed method is evaluated over several datasets like Caltech 101, FGVC-Aircraft etc. and the method with a pretrained encoder can outperform some of the adapted baselines.

Critical strengths
  1. The paper comprehensively ablates its design choices which helps understand the overall effect of each component.
  2. The proposed method is intuitive and simple, while being quite effective.
Critical weaknesses
  1. The motivation of the problem is weak. The paper argues that because supervised image classification is solved, research moved to open world settings and then arbitrarily mentions that full knowledge of class names is not always known. The paper tries to position itself between methods where classes are known and vocabulary free methods, but the reasoning isn’t strongly grounded.
  2. How is the LLM prior defined? The paper assumes knowledge of this prior. It is hard to determine what a prior probability of 0.5 means.
  3. The training process seems to make some choices which are difficult to follow: 3a. The training seems to depend on balanced sampling between known and novel categories (lines 308-310). Including the knowledge of novel categories does not stay true to the overall task. 3b. During finetuning, the top 3 neighbors of the query image ar eincluded as additional references in prompt (lines 313-315). This seems arbitrary and not true to this setup.
Questions for authors
  1. The approach relies heavily on being able to separate the similarity distribution between known and unknown classes to build class priors (202-212) and duplicate filtering (lines 225-230). This potentially makes this method weak on fine grained classification where the concepts heavily overlap with each other and CLIP similarity distributions are potentially hard to discriminate between known and unknown classes. However, the experiments include several fine grained datasets which show improvement, so I am curious to understand why this may be the case.
  2. The method uses custom accuracy metrics - cluster and semantic accuracy. For fine grained datasets, how do these metrics disambiguate between similar GT classes for each predicted class name?
2 comments

Thank you for your review comments, we will give a general response in the main rebuttal comment, while addressing specific points here.

W1: We would like to clarify our motivation. We see applications for OWIC in any setting where zero-shot image classification could be applied, but not all categories are known, e.g. surveying the species that occur in an area from a wildlife camera database, tracking failure types in a factory, or categorizing media libraries. For all these tasks, we will usually have knowledge of an initial set of category names, but simply applying existing zero-shot classification methods will assign all images to the known categories, even those belonging to undetected novel classes. On the other hand, VIC methods will discover novel categories, but they may not be aligned with our intended task description and our existing category names. We thus believe that OWIC is a valuable addition to the existing zero-shot classification tasks that solves problems that can currently not be mapped to any other zero-shot setting. The introduction was revised to better reflect the motivation and highlight practical use-cases more clearly.

W2: The LLM's prior is its bias towards declaring a sample known or novel acquired during its training. A prior probability of 0.5 means that the model has no bias either way. For the pretrained model, we do not know the prior so, as explained, our prior correction may be off by an additive constant. During finetuning, we explicitly set it to 0.5, thus ensuring that the prior correction works correctly.

W3: Explicitly setting the novel rate during training helps improve the accuracy of the prior correction, and ensures the model sees diverse training samples. This does not mean that we give the model access to any additional knowledge during evaluation, it is a purely training-side decision.

Giving the finetuned model access to the query sample's nearest neighbors does not conflict with the task setup in any way, it only utilizes data that is available under OWIC, notably the model is not given the labels for any sample. We also evaluated giving the pretrained model the same references, but found that it does not work as well since it lacks the necessary training (see Table 2).

Q1: While individual samples might be hard to discriminate, our prior estimation relies on population-level similarity statistics that are less vulnerable to sample-level ambiguity. As long as there is a global tendency for the correct class name to be closest to an image in CLIP space, the estimation works well (see Table 9). We see on Aircraft that there exist cases where this is not the case though. As for the duplicate filtering, the filtering threshold is set adaptively, so fine-grained tasks will receive a very strict threshold that will only discard extremely similar class names.

Q2: We would like to clarify that our metrics are not custom and the choice of metrics follows previous work such as FiNDR and CaSED. Cluster accuracy completely disregards category names, it will simply evaluate based on the cluster IDs, with the assignment between predicted and ground-truth classes decided by Hungarian matching. Semantic accuracy relies on a pretrained SentenceTransformer to capture the semantic differences between fine-grained categories. In that sense, the metric is limited by the capability of the transformer.

Thank you again for your review, we have now published a second revision of our paper that contains several new results that may be of interest to you. These include a new inductive setting, a detailed investigation of an example case in which the classification rule violates our attribute model, and various performance metrics. If you have any follow-up questions, we will be happy to answer them.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Rebuttal Comment

Thank you to all reviewers for the valuable feedback. We appreciate the comments calling our newly proposed setting and method well-motivated, and our experiments comprehensive. We also thank the reviewers for their constructive criticism, especially regarding our claims about task identifiability. When making our claim, we implicitly assumed certain properties of both the task and the known classes, and failed to consider scenarios where they do not hold. In our revised paper, we explicitly state those assumptions, and show that when they are fulfilled the known categories can substantially reduce the number of candidates for the classification rule, or even identify it outright. We also took note of all requested experiments and are working on providing all results before the end of the discussion period.

Due to the scope of the reviews and the time required to run all the requested experiments, we will split our response into two parts and publish two revisions of the paper, one now and the second once all experiments are complete. The changes in the first revision are:

  • A more precise discussion of task identifiability in Section 3.1, and under which assumptions the known categories can be used to identify the classification rule. Claims elsewhere in the paper were adjusted to match the analysis.
  • We reworked the introduction to better convey the motivation.
  • We also report the results for the first of the requested experiments, a sweep over the novel class percentage, in Appendix D to investigate MAND's performance in cases where fewer (or more) than 50% of categories are known.
  • We report results for varying the known/novel split while keeping the novel percentage constant in Table 7 and Table 8.
  • We add the standard deviations for the existing main results in Table 5 and Table 6 to help gauge the impact of different sample orderings.
  • We provide details on the implementation of semantic accuracy in Appendix D.1.

General Comment:

We now publish the second part of our rebuttal, and the second revision of our paper. The changes in the second revision are:

  • We add an inductive setting where half of all images for each class are held out and then later used for evaluation. The results can be found in appendix D.7.
  • We conduct a leave-one-out hyperparameter selection process and rerun our evaluation under the selected parameters. The results and the hyperparameters selected under this scheme can be found in appendix D.8.
  • We investigate MAND's behavior in cases where the classification rule violates our attribute model, the results can be found in appendix D.9
  • We add a new baseline based on the suggestions of reviewer DHwn, the details of the baseline can be found in appendix C.2, the results in appendix D.2.
  • We report computational cost (LLM calls, generated tokens, GPU and wall-clock time) in appendix D.10.

This concludes our initial response, we are looking forward to further questions and discussions.

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

Related papers

Discussion (0)

Sign in to comment.