acceptodds
Under review as a conference paper at ICLR 2027

Learn Before You Judge: Progressive Knowledge-to-Decision Alignment for Explainable Hateful Meme Detection

Abstract

Hateful memes spread abusive content through implicit interactions between images and text, posing serious threats to the safety of online communities. In recent years, multimodal large language models have been widely used for hateful meme detection and are increasingly adopted to generate explainable detection results. However, we find that existing explain-then-detect methods often couple explanation generation and label prediction within the same training process. This coupling causes interference between task objectives, leading to limited detection performance and even worse results than simple SFT baselines. To address these challenges, we propose ProKDA, a progressive knowledge-to-decision alignment method for explainable hateful meme detection. Inspired by the human annotation training process, ProKDA first uses an agentic background knowledge construction pipeline to obtain external knowledge related to meme understanding. It then adopts a three-stage training strategy that sequentially performs background knowledge learning, hatefulness detection learning, and hatefulness boundary alignment. Unlike prior explain-then-detect methods that jointly optimize both tasks, ProKDA focuses on a single training objective at each stage. This design reduces interference between the two tasks and progressively transforms background knowledge into robust detection decisions. Experiments on three public hateful meme benchmarks show that ProKDA achieves state-of-the-art detection performance and provides accurate, explainable, and evidence-supported decisions for hateful meme moderation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Hateful Meme Detection

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_n3XROpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

Summary

The paper studies explainable hateful meme detection. The initial observation is that existing 'explain-then-detect' methods works on explanation generation followed by corresponding label prediction. Both the stages are optimized together. The authors argue that it causes task-objective interference and results worst performance than simple SFT.

Author proposed, ProKDA (Progressive Knowledge-to-Decision Alignment) that mimics how human annotators are trained. It learns the background, annotate, then re-review hard cases through a pipeline with one dedicated objective per stage:

  1. An agentic knowledge-construction step (GPT-4o agents + Wikipedia retrieval) builds a background-knowledge corpus for each meme.
  2. Stage 1 (SFT): It learns that background knowledge.
  3. Stage 2 (SFT): It adapts to generate a 'Yes' or 'No' in hatefulness detection.
  4. Confidence filtering + Stage 3 (DPO): It keeps only low-confidence “boundary” samples and refine the decision boundary.

The study attempts to show that decoupling them into separate stages improves the classification task.

Critical strengths

Strengths

  1. It has a clean ablation of stage ordering. Table 2 genuinely supports the compounding story, i.e., stage 1+2 > stage 2 alone, and adding Stage 3 helps further.
  2. The 'learn before you judge' framing is intuitive, and the ordered-SFT evidence is the paper’s most convincing part.
  3. It consists of a set of honest failure analyses. Figures 14 to 16 and the error section openly show ProKDA failing on highly implicit memes.
  4. The outputs are likely reproducible. The required LoRA config, LLaMA-Factory version, hardware, and full prompts are given.
Critical weaknesses

Weakness:

  1. Check the claim of SFT Explain-then-Detect < SFT Direct Detection at line number 309. Prior works [1] highlighted it too. Please cite them. [1] EXPO-HM: LEARNING TO EXPLAIN-THEN-DETECT FOR HATEFUL MEME DETECTION

  2. The real gap mentioned in the study is that the existing explain-then-detect methods combine (a) explanation generation and (b) label prediction in the same training process. It causes disturbance between the two task objectives that hurt both detection performance and explanation quality. The proposed method ProKDA is formulated around the interference hypothesis. The empirical gap is reported in Table 1 but the causal “interference” gap is not investigated enough. We need to be convinced that interference is the root cause behind the fall. Kindly add one or two gradient-conflict or representation-level analyses to endorse the claim.

  3. Distillation vs. Decoupling: From Table 2, i.e., the Ablation study of ProKDA components, we see that

  • Stage 2 only (= Direct Detection SFT): W-F1 75.22

  • Stage 1+2: 81.79, i.e., adding the knowledge gives a gain of +6.57.

  • Full ProKDA (+DPO): 83.44, i.e., boundary alignment adds only +1.65

    So, the teacher-synthesised external knowledge genuinely helps, but 1.65 from the “boundary alignment” piece that is most distinctive to the decoupling story contributes least. It seems that a large share of ProKDA’s advantage is due to teacher distillation, not progressive decoupling.

  1. Narrow pipeline for knowledge quality analysis: It considers a total of 100 samples from HMC hateful memes and compares ProKDA’s knowledge via LLM judges and two human annotators on informativeness and consistency. But the knowledge base spans all three datasets and both classes. So, other datasets and knowledge related to benign memes remained unvalidated. Note that benign-meme knowledge matters for confounded cases.

    In addition, the evaluation pipeline measures preference instead of correctness. Evaluators chose the better explanation or marked a tie, which is relative, but nothing in the pipeline verifies the knowledge is factually accurate or that the Wikipedia retrieval was on target. There is a high chance that hallucinated or wrong retrievals have propagated into the Stage-1 SFT (Eq. 4).

  2. No variance, no significance testing: Many important deltas w.r.t. ProKDA are +1 to +4 points (Tables 1 and 2) with apparently single runs. Kindly include seeds/CIs/mean±std.

Questions for authors

Kindly consider the issues mentioned in the 'Weakness Section' to improve the work.

3 comments

Dear Reviewer #n3XR,

We would like to kindly follow up on our rebuttal. We would greatly appreciate it if you could let us know whether our responses have adequately addressed your concerns or if any issues remain unresolved.

Thank you for your time and consideration!

Thank you for addressing the important points. I will raise my score.

Thank you for your reconsideration and feedback. We really appreciate it !!!

Reviewer_pureOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ProKDA, a three-stage approach to explainable hateful meme detection. An agentic pipeline uses GPT-4o and Wikipedia retrieval to construct background knowledge for training memes. The model first learns to generate this knowledge together with the annotations, then learns direct binary classification, and finally undergoes DPO on low-confidence training examples using gold and flipped labels as preference pairs. Experiments cover HMC, MAMI, and PrideMM with Qwen2.5-VL-3B and 7B. The main contribution is an effective training recipe that separates explanation supervision from subsequent decision optimization.

Critical strengths
  1. The problem is relevant and the approach is easy to follow. Implicit interactions between images, text, and cultural context make hateful meme detection difficult. Learning contextual knowledge before specializing in classification is a sensible design. The comparison between explanation-based and direct SFT provides a useful empirical motivation for investigating training order.

  2. The classification improvements are substantial relative to the authors' direct baselines. In Table 1, the 7B model improves HMC macro-F1 from 75.15 with direct SFT to 83.41 with ProKDA, MAMI from 78.37 to 83.08, and PrideMM from 77.71 to 80.27. Improvements over the corresponding direct SFT and DPO baselines occur at both model sizes and across all three datasets. The 7B version also has the highest macro-F1 among the methods reported in the table.

  3. The ablations help establish the value of the stages. Table 2 shows benefits from background knowledge learning and from the final alignment stage on HMC. For the 7B model, adding Stage 1 to Stage 2 increases weighted-F1 from 75.22 to 81.79, and adding Stage 3 further increases it to 83.44. The threshold analysis also helps explain the choice to focus alignment on a subset of examples.

  4. The evaluation includes explanation quality. Section 5.6 combines two LLM judges with three human evaluators on hateful HMC memes with HatReD reference explanations. The improvements over both the original model and the Stage 1 checkpoint suggest that the later classification stages preserve useful explanation capability on this subset. The appendices provide prompts, training settings, qualitative error cases, and an anonymous code-repository link.

Critical weaknesses
  1. The experiments do not fully isolate why progressive training helps. The explain-then-detect baseline already uses the same constructed background knowledge and gold labels (Section 5.2), and the authors report consistent hyperparameter settings. This is a useful control for the supervision source. Nevertheless, its lower performance does not establish task interference as the cause: response length, loss weighting, output format, total training budget, and the final emphasis on short labels could also contribute. A comparison that additionally matches the total optimization budget while varying the training schedule would strengthen that interpretation. Table 2 supports the benefit of the complete Stage 1, and Appendix F compares the quality of the agentic pipeline's explanations with direct GPT-4o explanations. A remaining question is whether Wikipedia retrieval improves downstream detection when the teacher, prompts, and training schedule are otherwise held fixed; the existing comparisons do not isolate that contribution.

  2. The specific benefit of DPO needs a stronger control. Gold-versus-flipped-label preferences are natural for binary classification. The unresolved question is whether DPO provides an advantage over additional SFT on the same low-confidence examples with a matched training budget. The existing stage ablation demonstrates a benefit from the final stage, but does not distinguish the preference objective from additional optimization on selected examples. A size-matched random subset would help assess the selection rule as well.

  3. The smaller gains over the strongest reported methods lack uncertainty estimates. For example, 7B ProKDA exceeds ExPO-HM by 0.78 macro-F1 points on MAMI. Repeated runs or uncertainty estimates would help establish whether such differences are reliable. Appendix C also states that most external baseline results come from their original papers. These comparisons provide useful context, but a stronger claim of superiority requires attention to identical splits and evaluation procedures. Appendix C specifies validation-based checkpoint selection; the procedure used to select the confidence threshold should be stated separately.

  4. The human explanation evaluation omits annotator agreement. Section 5.6 reports ratings from three student annotators on informativeness, readability, soundness, and persuasiveness, and Appendix F reports pairwise preferences from two annotators. Neither evaluation reports inter-annotator agreement or uncertainty around the aggregate scores. Without these, it is difficult to tell how consistently the criteria were applied. The explanation metrics assess the usefulness of the generated text, but do not establish that it faithfully reflects the evidence responsible for the classifier's decision. Appendix E.5 does provide a prompt that requests both a hatefulness judgment and an explanation, so the paper demonstrates explanation generation after the later stages train on label outputs. The remaining concerns are evaluation reliability and scope: the reference-based evaluation covers hateful memes with HatReD explanations, and does not assess explanation quality on benign cases or on MAMI and PrideMM. The acknowledged failures on highly implicit hate also deserve a more systematic breakdown, since those examples are central to the motivation.

  5. The offline cost of constructing background knowledge is not quantified. The method uses GPT-4o agents and Wikipedia retrieval to construct supervision for training memes, but reports no retrieval latency, total preprocessing time, or cost per example. These are costs of preparing training data; the described detection inference uses the trained model directly and does not require a Wikipedia lookup for each new meme. Reporting preprocessing time, API cost, and total training time would help assess the resources required to reproduce the method. Appendix C identifies the two-GPU hardware and specifies three epochs per training task, but does not provide a runtime comparison with the SFT baselines.

  6. The discussion of related staged training is incomplete. MemeIntel (Kmainasi et al., 2025), Section 5.3 already motivates staged optimization through conflicts between classification and explanation objectives, training classification first and then jointly learning classification and explanation. DR-HM (Cheng et al., Findings ACL 2026) combines teacher-generated structured reasoning, two-stage SFT, and subsequent adaptive GRPO for harmful meme detection. Neither appears in the submitted manuscript's bibliography, despite their relevance to the central motivation and training strategy. The authors should explain what ProKDA adds through its different training order, retrieved knowledge, and confidence-filtered DPO. MemeGraphs (Kougia et al., 2023) is also relevant background for external knowledge augmentation: it combines visual scene graphs with knowledge retrieved from Wikidata for meme classification. Its connection is to knowledge construction rather than staged training. These overlaps warrant clearer positioning; they do not establish that ProKDA's complete pipeline has already been proposed.

  7. Generalization to unseen templates and knowledge is not established. ProKDA incorporates retrieved knowledge through training, while its described inference procedure uses the trained model without external retrieval. It may therefore struggle when a meme requires facts about a recent event or an unfamiliar cultural reference that are absent from the model's learned knowledge and cannot be inferred from the input. Updating Wikipedia alone would not update the trained model. The paper also provides no analysis of image, text, template, or cultural-reference similarity between its training and test splits. Held-out examples can still share these elements with training examples, so the reported improvements do not establish that the advantage persists on unfamiliar templates or knowledge. This is an evaluation gap, not evidence that the splits contain duplicates or leakage. Characterizing train–test similarity and reporting comparisons with the baselines on a subset with minimal overlap would help establish the scope of the method's generalization.

Questions for authors
  1. How were the confidence thresholds chosen, and were all choices made using validation data? How many examples are retained for each dataset and model?
  2. Does ProKDA retain its advantage when the direct and joint-training baselines receive the same total training budget? What happens when the same knowledge supervision is mixed rather than presented progressively?
  3. How does Stage 3 compare with continued SFT on the identical selected examples and with DPO on a size-matched random subset?
  4. Can you report run variability and inter-annotator agreement for both the four-dimension explanation ratings and the pairwise background-knowledge evaluation?
  5. What are the per-example and total time and monetary costs of GPT-4o knowledge generation and Wikipedia retrieval? How much additional training time does ProKDA require over the SFT baselines?
  6. How does ProKDA differ from MemeIntel's staged treatment of classification and explanation objectives and DR-HM's distill-then-reinforce pipeline? Which contribution is specific to ProKDA's training order, Wikipedia knowledge construction, or confidence-filtered DPO?
  7. How similar are the train and test memes in their images, text, templates, and cultural references? Does ProKDA retain its advantage over the baselines on a subset with minimal overlap or with unseen templates, events, or references? Given that retrieval is used only to construct training supervision, how would new knowledge be incorporated after training?
3 comments

Dear Reviewer #pure,

We would like to kindly follow up on our rebuttal. We would greatly appreciate it if you could let us know whether our responses have adequately addressed your concerns or if any issues remain unresolved.

Thank you for your time and consideration!

Thanks for the additional experiments. While the reviews addressed my concerns I agree with the other reviewers that the novelty is limited. But I think the paper presents a strong case for acceptance. I will not raise the scores given the novelty concerns.

Thank you for your thoughtful consideration and constructive feedback. We sincerely appreciate your support for our work !!!

Reviewer_4LKGOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ProKDA for explainable hateful meme detection, combining annotation-conditioned knowledge generation, SFT on background knowledge, SFT for short-label detection, and standard DPO on low-confidence samples into a progressive training schedule. GPT-4o query and answer agents use Wikipedia retrieval to generate textual, visual, and cross-modal background question–answer pairs offline; the fine-tuned Qwen2.5-VL handles inference. The paper reports classification improvements on HMC, MAMI, and PrideMM at two model scales, and evaluates generated explanations using human and model judges.

Critical strengths
  1. The reported gains with the same backbone have practical value. In Table 1, the 7B model achieves macro-F1 scores of 83.41, 83.08, and 80.27 on HMC, MAMI, and PrideMM, respectively, improving over direct-detection SFT by 8.26, 4.71, and 2.56 percentage points. The 3B results show a consistent direction of improvement. Although these results have not been independently reproduced and their stability is not quantified, they provide useful preliminary empirical support for the progressive training schedule.
  2. The paper includes stage ablations and evaluates explanations with human judges and two model judges. Table 2 compares combinations of the knowledge-learning and preference-optimization stages using the same backbone, while Table 3 provides several perspectives on the quality of generated explanations. These analyses help identify components that merit further investigation and provide more insight than final performance alone.
Critical weaknesses
  1. Originality is limited, and the distinctive contribution of the combination has not been isolated. Knowledge retrieval and teacher-generated data, SFT, standard DPO, and confidence-based training-sample selection are established techniques. The paper introduces neither a new learning objective nor a new optimization algorithm; the agents construct supervision offline and therefore do not constitute a new inference architecture. The contribution consequently needs to be established through controlled empirical findings about the training combination. Tables 1–2 do not include joint training, sequential training, and repeated detection-SFT controls matched for teacher data, total updates, and target-token budget. Appendix C allocates three epochs to each training task, leaving additional training as an alternative explanation for the gains. Moreover, the Stage 1 target, Format(h_i,b_i), still contains the label and the target of the hateful content. A single loss function does not imply a single semantic task, and the current results do not establish that task interference has been disentangled. Knowledge generation is also conditioned on gold labels and the targets of the hateful content, without separate controls for teacher generation without retrieval, annotation-only supervision, or generation without label conditioning. Stage 3 lacks comparisons against continued SFT on exactly the same selected samples, equally sized random subsets, and DPO under matched budgets. It is therefore unclear whether the gains arise from external knowledge, additional annotations and teacher distillation, retraining on difficult samples, or the specific stage design. Confidence analysis on the training set does not establish probability calibration on held-out data. My main reason for rejection is that the paper presents an effective training recipe as a new mechanistic contribution without evidence at that level. Additional equations or stronger emphasis on stage names would not resolve this concern.
  2. The evaluation does not sufficiently support the full claims about explainability and superior performance. Table 3 compares only the original Qwen model, Stage 1, and ProKDA, omitting the explain-then-detect SFT baseline that the paper primarily challenges. The HatReD evaluation covers only 246 hateful examples. Agreement with reference explanations and persuasiveness do not establish that the explanations faithfully reflect the evidence used by direct Yes/No detection, and benign examples and explanations for misclassifications are not covered. The paper should add blinded comparisons on the same examples, sample-level checks of label–explanation consistency, and interventions on the evidence cited in the explanations, or narrow its conclusions to improvements in generated rationale quality. The classification results also contain a specific, checkable inconsistency. Table 4 lists the HMC test set as 500 benign and 500 hateful examples. Under standard definitions, macro-F1 and F1 weighted by true class support must be identical for the same evaluated predictions on such a balanced set. Yet Table 1 reports 71.07/71.18 for 3B direct SFT and 83.41/83.44 for 7B ProKDA. Please clarify the actual sample IDs, class supports, handling of invalid outputs, and metric definitions, and provide per-example predictions and the metric script. Most strong external baselines are taken from the literature without reruns under the same protocol or sufficient explanation of protocol alignment. Random seeds, variability across repeated runs, and paired confidence intervals for performance differences are also absent, making the reliability of the claimed margins difficult to assess. This inconsistency requires clarification; it does not justify an inference of fabrication or leakage.
Questions for authors

The answers to the following questions could materially affect the rating. The first two concern attribution of the contribution, and the last two concern evaluation validity.

  1. With the same backbone, teacher corpus, total number of updates, and target-token budget, does ProKDA still outperform joint explanation-and-detection SFT, repeated detection SFT, and alternative training orders? What are the complete Stage 1 output template and loss mask, and are label tokens included in the optimization?
  2. How does performance change when Wikipedia retrieval is removed while retaining the same teacher and annotations? On exactly the same low-confidence samples selected for Stage 3, what is the difference between continued SFT and DPO with matched update counts? How does DPO on an equally sized random subset compare?
  3. What are the actual HMC test-set class supports and the total number of valid evaluated examples, and why do macro-F1 and weighted-F1 differ? Please provide the metric code and per-example predictions, describe the validation split, and specify which splits are used to select the checkpoint and confidence threshold. Are multiple-seed results or paired confidence intervals available for the key comparisons?
  4. How are the explanations in Table 3 generated, and are gold labels or existing predictions supplied as inputs? Does the explanation advantage persist when explain-then-detect SFT is included and benign examples and misclassified examples are covered? When textual or visual evidence identified as crucial in an explanation is removed or counterfactually modified, does the direct detection decision change accordingly?
3 comments

Dear Reviewer #4LKG,

We would like to kindly follow up on our rebuttal. We would greatly appreciate it if you could let us know whether our responses have adequately addressed your concerns or if any issues remain unresolved.

Thank you for your time and consideration!

Thank you for addressing my main concerns! Based on your clarifications, I will raise my score.

Thank you very much for your recognition and support of our work. We sincerely appreciate your time and consideration !!!

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

This paper proposes ProKDA, a progressive knowledge-to-decision alignment framework for explainable hateful meme detection using multimodal large language models (MLLMs). The authors identify that existing explain-then-detect methods suffer from task interference when explanation generation and label prediction are jointly optimized. To address this, ProKDA adopts a three-stage training strategy: (1) background knowledge learning via supervised fine-tuning on an agentic-constructed knowledge corpus, (2) hatefulness detection learning via SFT to adapt the model to binary classification, and (3) hatefulness boundary alignment via DPO on confidence-filtered boundary samples. The method is evaluated on three hateful meme benchmarks (HMC, MAMI, PrideMM) and reports competitive detection performance alongside improved explanation quality compared to several baselines.

Critical strengths

The paper's primary strength is the well-motivated progressive training paradigm that decouples knowledge acquisition from decision learning, with ablation studies (Table 2) providing clear empirical evidence that each stage contributes incrementally to detection performance. Building on this empirical foundation, the agentic background knowledge construction pipeline effectively incorporates external context (e.g., historical events, cultural symbols) into the detection process, yielding measurable improvements in both classification accuracy and explanation quality (Table 3).

Critical weaknesses

While the pipeline is well-constructed, I have concerns regarding its novelty. ProKDA largely combines established methods (RAG, SFT, DPO) without significant algorithmic innovation. The "agentic" construction and DPO stages feel like standard applications rather than new contributions. To meet ICLR standards, the authors should better articulate the theoretical insights behind this combination.

Additionally, the practicality of the method needs clarification. The reliance on multiple GPT-4o calls per sample introduces significant overhead compared to retrieval-free baselines. A detailed analysis of the computational cost (latency, API costs) versus accuracy gains would strengthen the paper's argument for deployability.

Lastly, the evaluation of explanation quality could be more rigorous. Relying primarily on LLM judges is risky due to potential biases, and the current human evaluation subset is too small to be conclusive. Expanding the human study and reporting inter-annotator agreement would greatly enhance the credibility of the results.

Questions for authors
  1. How does ProKDA's detection performance and explanation quality change when the agentic knowledge construction pipeline uses a smaller, open-source LLM (e.g., Qwen2.5-VL-7B itself) instead of GPT-4o? Providing such results would clarify whether the performance gain stems from the pipeline design or from the capability of the underlying knowledge-generation model.

  2. Could you report a detailed cost analysis of the full ProKDA pipeline, including the number of API calls, token consumption, and wall-clock time required for background knowledge construction per training sample? This would help the community assess the practical trade-off between accuracy gains and computational overhead.

  3. The DPO stage constructs preference pairs by contrasting gold-label responses with flipped-label responses on confidence-filtered boundary samples. Has the authors considered whether this objective is merely teaching the model to memorize the training labels more confidently on hard examples, rather than learning genuinely discriminative features? A comparison with a baseline that applies DPO directly on the original training set (without confidence-based filtering) would help isolate the contribution of the filtering mechanism from the DPO objective itself.

2 comments

Dear Reviewer #XmaW,

We would like to kindly follow up on our rebuttal. We would greatly appreciate it if you could let us know whether our responses have adequately addressed your concerns or if any issues remain unresolved.

Thank you for your time and consideration!

Dear Reviewer #XmaW,

As today is the final day of the rebuttal period, we would sincerely appreciate it if you could kindly revisit our responses and let us know whether your concerns have been addressed.

Thank you very much for your time and consideration!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.