acceptodds
Withdrawn from ICLR 2027

BEAST: Benchmarking Evolving Abuse Streams

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Online abusive language evolves continually, so classifiers deployed in real-world moderation systems must continually adapt to text that may no longer resemble what they were trained on. Testing whether adaptation methods are effective requires a benchmark that preserves real-world task constraints—naturally occurring content, a fixed prediction task, a continuous stream, and meaningful volume—while methodologically isolating the challenge of continuous linguistic evolution. No existing continual-learning benchmark for abuse detection combines these properties. Instead, they manufacture distribution shift by sequencing tasks, datasets, bias dimensions, or adversarial perturbation types, which tests adaptation to one specified form of change rather than to language that evolves on its own. We introduce Benchmarking Evolving Abuse Streams (BEAST), a domain-incremental benchmark that holds the task and moderation policy fixed while only the incoming language changes. BEAST draws millions of naturally occurring posts from a single 3.5-year source and orders them so that incoming text grows progressively less familiar relative to the initial training period. Models are scored on an interval of the stream before that interval becomes available for training. We measure and report predictive performance, adaptation cost, and forgetting: across encoder and decoder models, increasing model scale and periodic retraining both improve average performance yet leave substantial end-to-end degradation, while retrieval augmentation performs far worse despite substantially higher compute. BEAST exposes a persistent gap between current adaptation methods and the robustness that evolving abuse detection demands.

Primary area:
datasets and benchmarks
Keywords:
Continual Learning, Benchmarks, Domain-Incremental Learning, Abuse Detection, Moderation, Streaming, Trust, Safety

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ABh3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose BEAST a large scale benchmark for evaluating robustness and continual adaptation of abusive language classifier under linguistic distribution shift. The benchmark comprises of 6.04 million naturally occuring posts , collected from a 3.5 years single online source. The authors order them as by sorting them according to their negative log-likelihood under a 5-gram language model trained on the initial data. That leads to 20 evaluation intervals from relatively familiar to increasingly unfamiliar. The authors utilize LLMs to extract semantic attributes from the posts. These are the mapped to binary abuse labels using a deterministic moderation policy.The benchmark is evaluated on BERT-Large, ModernBERT-Large, and Qwen3.5 models under static training, periodic retraining, and retrieval-augmented adaptation. The experimental results show that increasing model size and periodic retraining improve average classification performance. However, it does not prevent substantial degradation in performance as linguistic novelty increases. The retreival augmented configuration performs worse than its static counterpart, while requiring more computational overhead.

Critical strengths

Strength:

  1. This paper addresses an important problem of ensuring reliable abusive-language detection under evolving linguistic distributions.

  2. It provides a large scale benchmark comprising 6.04 million naturally occurring posts collected over 3.5 years, providing much larger evaluation setting than existing continual abusive-language detection benchmarks.

  3. The paper uses a prequential protocol(test-then-adapt); so every model is tested on each new step before it can train on it. This lets static and adaptive models be compared fairly as the posts become less familiar.

  4. Besides classification performance, the authors also report computational costs and backward transfer, providing useful insights into the efficiency, and forgetting behaviour of adaptation methods. The experiments include both encoder and decoder-based language models and compare static training, periodic retraining, and retrieval-augmented adaptation, providing useful baseline results for future research.

Critical weaknesses
  1. The benchmark does not actually measure how abusive language changes over time. The paper's own chronological test (Appendix D) shows no drop in performance over the 3.5 years. The authors rather sort posts by how unfamiliar they are to a 5-gram model trained on the initial batch. It does not order them as per timeline; rather it's an out-of-distribution ranking. E.g., an unusual post from 2022, could appear at the end of the benchmark; while a fairly common post from 2024, could appear at the beginning of the benchmark. High NLL does not suggest linguistic evolution or newly emerging abuses; rather they might suggest rare words, unusual syntax, or uncommon spelling variations. This raises concerns regarding whether the degradation is due to chronological linguistic evolution or due to the artificially constructed novelty.

  2. The authors use Gemini 3.5 Flash to generate semantic annotations, which are then converted into binary labels using a deterministic policy. However, the authors do not validate these labels against human annotations. An annotation model may also experience difficulty in reliably interpreting linguistically uncommon expressions and coded abusive language. Therefore, annotation errors could also increase with novelty and contribute to the performance degradation.

  3. The RAG performs substantially worse than its static counterpart; with mean PR-AUC decreasing from 0.7093 to 0.4141. However, the paper does not sufficiently investigate the reasons for this degradation. The experiments use a single embedding model and a fixed number of retrieved examples. Therefore, it is difficult to conclude whether the poor performance of RAG is an inherent limitation of retrieval-based adaptation or shortcomings of the particular implementation.

  4. Although the paper proposes a benchmark for continual learning; the evaluation is limited to static models, periodic retraining, and a single RAG configuration. Established continual-learning approaches, such as experience replay, regularization-based methods, and cumulative retraining, are not evaluated in this paper. Given the paper's emphasis on continual adaptation, the absence of these baselines limits the strength of its experimental conclusions.

  5. There are also some issues related to formatting and presentation: a)Table 1 (page 2) extends beyond the normal text margins. b) The mathematical expression defining the abuse policy in Appendix B (page 13) extends beyond text margin. c)Some tables, particularly Table 8 (page 13), are difficult to read due to small font sizes and densely packed information.d) In Figure 7(b) (page 17), the legends refer to Mean Perplexity and Median Perplexity; however, both the y-axis and figure caption refer to negative log-likelihood (NLL). The terminology should be consistent throughout the paper.

Questions for authors
  1. Given that the chronological evaluation shows no clear performance drop, how do the authors justify that the NLL-based ordering captures evolving abusive language rather than simply unfamiliar or difficult examples?

  2. Can the authors validate the LLM-generated labels against human annotations? Does annotation quality remain consistent as linguistic novelty increases?

  3. Why RAG perform significantly worse than the static model? Is this due to the choice of embedding model, retrieval quality or the number of retrieved examples?

  4. How would known continual-learning methods (e.g. experience replay, regularization-based methods) perform on BEAST?

Reviewer_Tf7yOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces BEAST, a benchmark for continual learning in abusive-language detection. It uses 6.04 million naturally occurring posts from a 3.5-year corpus. Labels are generated by extracting semantic attributes with Gemini 3.5 Flash and applying a deterministic moderation policy. The benchmark does not preserve chronological order. Instead, the authors use a 5-gram language model to rank posts by negative log-likelihood (NLL), downsample the middle of the NLL distribution, and construct 20 pseudo-timesteps of increasing linguistic unfamiliarity. Experiments with BERT, ModernBERT, and Qwen models show substantial performance degradation across this constructed stream, even with periodic retraining.

Critical strengths

Abuse-detection systems need to handle changing language, and evaluating adaptation over a large collection of naturally occurring posts is potentially valuable.

Critical weaknesses
  1. The benchmark construction is fundamentally unsound. The authors report no systematic performance degradation under the original chronological ordering. They then reorder the data using 5-gram NLL and discard 90% of the middle 60% of the NLL distribution to construct an artificial difficulty gradient. High NLL does not necessarily represent evolving abusive language. The paper provides no convincing validation of this assumption. Showing that classifiers degrade on increasingly unfamiliar text does not establish that they fail to adapt to evolving abuse. This undermines the central motivation and conclusions of the paper.

  2. The benchmark labels are not independently validated, making the reported degradation difficult to interpret. The labels rely entirely on semantic attributes extracted by Gemini 3.5 Flash, without an independent human-annotated reference set. This is especially problematic because the evaluation deliberately progresses toward increasingly unfamiliar language, where annotation reliability may also deteriorate. The observed performance decline could therefore partly reflect annotation errors rather than classifier failures. The paper provides no evidence ruling out this confound.

These are fundamental problems with benchmark validity, not minor limitations that can be addressed by adding more models or baselines.

Questions for authors

1 Why is 5-gram NLL an appropriate measure of evolving abusive language? What evidence shows that higher NLL corresponds to newly emerging abusive expressions rather than general linguistic unfamiliarity or classification difficulty?

2 Given that the original chronological stream shows broadly stable performance, what independent evidence supports the claim that the constructed stream captures an important failure of adaptation to real-world language evolution?

3 How can the authors establish that annotation errors do not increase with NLL? Without independent human validation across the novelty spectrum, how can the observed degradation be separated from potential changes in label reliability?

Reviewer_vBitOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces BEAST, a domain-incremental benchmark for abuse detection. Models are evaluated using a prequential test-then-adapt protocol, with predictive performance, compute cost, and forgetting measured. The experiments find that larger models and periodic retraining improve average performance but do not prevent substantial degradation.

Critical strengths

The problem is practically relevant. Abuse classifiers deployed in moderation systems face changing language, and the paper appropriately asks whether they can adapt as the input distribution changes.

The benchmark offers a useful and clearly motivated evaluation setup.

Critical weaknesses

The benchmark’s novelty gradient is constructed rather than observed in deployment. BEAST sorts examples by their negative log-likelihood under an early-period language model and subsamples the middle of that distribution. This limits how directly its main results can be interpreted as estimates of real-world temporal degradation.

The experimental evidence is narrow for broad conclusions about continual adaptation. The paper evaluates one primary policy, a limited set of models and adaptation strategies, and one training seed per configuration.

The data and labels limit external validity and label validity. The corpus comes from one community and primarily one language, and its labels were generated by an LLM-based annotator without validation against an independent human-annotated subset.

The length limit is not sufficiently used.

Questions for authors

Does the novelty-ordered benchmark predict real chronological failure?

What is the intended scope of generalization? Since BEAST comes from one community and primarily one language, what evidence would support applying its conclusions to other platforms, populations, and moderation policies?

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

Related papers

Discussion (0)

Sign in to comment.