{"id":"4837a85a-efc7-455a-8d1f-bfcd9f3e4936","arxiv_id":"2509.01379","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An AI agent combining retrieval, a BERT classifier, slang lookups, and policy guidelines reports macro F1 0.91 on a reannotated MetaHate test set, about 3 points above specialized baselines.","lead":"This paper presents WATCHED, a chatbot-like AI agent that helps content moderators decide whether social media posts are hate speech. It retrieves similar examples, uses a BERT classifier, looks up slang, and cites platform guidelines to explain each decision, reporting a macro F1 of 0.91 on a 2,001-post test set.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2 baseline F1 scores exactly match the original [32] values, despite the paper claiming all metrics are on the reannotated eval subset — strong evidence baselines were scored on different labels.","rationale":"The reader's weakest assumption identified the ambiguity about whether baselines were recomputed on the reannotated labels. My analysis goes further: the exact match between Table 2's baseline F1 scores and the original publication values is strong evidence that they were not recomputed. This is an internal inconsistency, not merely a missing clarification. The claim of superiority is therefore not currently supported. The verdict should remain CONDITIONAL because the issue is fixable: the authors can re-run the baselines on the released reannotated labels and report the correct numbers. If the re-run still shows WATCHED ahead, the claim stands; if not, it falls. The lack of inter-annotator agreement is a secondary concern that also needs addressing, but the primary load-bearing issue is the mismatched evaluation labels.","tokens_in":11582,"tokens_out":4053,"duration_ms":43204,"concrete_test":"Download the released reannotated labels from the GitHub repository (nulldiego/watched). Re-run MetaHateBERT, Distil MetaHate, Llama 3 8B/70B few-shot CoT, and Perspective API on the same 2,001 examples using the original [32] settings. Compare the resulting F1_MACRO values to Table 2. Also compute inter-annotator agreement if raw annotations are available. If the baseline scores differ from Table 2, the comparison is invalid; if they match exactly, that would require a detailed explanation (e.g., no labels changed, which is implausible for 2,001 examples).","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that WATCHED surpasses state-of-the-art rests on Table 2, which is headed 'Results on the reannotated eval subset.' However, the F1_MACRO scores for MetaHateBERT (0.8801) and Distil MetaHate (0.8807) are identical to the values attributed to the original publications in the text. The evaluation uses a reannotated 2,001-example subset; any label changes would almost certainly alter these scores at least in the 4th decimal. Exact equality implies these numbers were copied from [31] and [32], which used the original MetaHate labels, not the reannotated labels. Thus WATCHED's 0.9139 F1_MACRO is computed on a different ground truth than the baselines, making the comparison invalid. The sentence 'we adopt the baselines and experimental settings from [32], and we report the metrics on the reannotated test split' is ambiguous, but the numerical evidence contradicts the interpretation that all baselines were rerun. Additionally, no inter-annotator agreement is reported, so the reliability of the reannotation itself is unknown. Without re-evaluating all baselines on the same reannotated labels, the 'surpasses SOTA' conclusion is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents WATCHED, an AI-agent system for hate speech detection that combines a fine-tuned BERT classifier (MetaHateBERT), retrieval-augmented generation over a vector database of MetaHate posts, Urban Dictionary lookups, a distilled reasoning LLM (Distil MetaHate), and social-network guideline retrieval. The agent produces a binary hate-speech label, a confidence score, and an explanation grounded in retrieved examples and policies. The authors evaluate on a 2,001-example subset of MetaHate that they reannotated with two expert annotators, reporting a macro F1 of 0.9139, which they claim surpasses existing state-of-the-art methods. An ablation study shows that removing any single tool degrades performance. The paper includes open-source code and releases the reannotated labels.","tokens_in":11940,"tokens_out":3961,"duration_ms":40929,"significance":"If the evaluation were sound, WATCHED would be a practically valuable contribution: it combines several complementary evidence sources (classifier, retrieval, slang definitions, reasoning, policy) into a single transparent moderation tool, and it reports a substantial improvement over strong MetaHate-trained baselines. The open-source release and the decision to reannotate the test subset to improve label reliability are commendable. However, the central claim of superiority is not supported by the evidence as presented: the baseline numbers appear to have been taken from prior publications rather than recomputed on the reannotated labels, and the single-run evaluation lacks any measure of variability. These are load-bearing issues that must be resolved before the result can be accepted.","major_comments":[{"comment":"The F1_MACRO values for MetaHateBERT (0.8801) and Distil MetaHate (0.8807) reported in Table 2 are numerically identical to the values from the original publications [31] and [32] quoted in the same section. Since Section 2.3.1 states that the 2,001-instance evaluation sample was reannotated by two independent experts, any label changes would be expected to alter these scores at least at the fourth decimal. The exact equality indicates that the baselines were not evaluated on the reannotated labels but rather copied from prior results. Consequently, WATCHED's 0.9139 is measured against a different ground truth than the baseline scores, invalidating the 'surpasses existing state-of-the-art' claim in the abstract and Section 5. The authors must rerun all baselines on the same reannotated labels and report those numbers.","section":"Section 2.3.4, Table 2"},{"comment":"The sentence 'we adopt the baselines and experimental settings from [32], and we report the metrics on the reannotated test split' is ambiguous. It could mean that all baselines were recomputed on the reannotated split, or that the metrics are simply reported as in [32]. The ambiguity is consequential: if the baselines were recomputed, the authors should explain why the scores coincide exactly with [32]; if not, they must state that explicitly and provide a rerun. Releasing the reannotated labels makes such a rerun feasible.","section":"Section 2.3.3"},{"comment":"No inter-annotator agreement is reported for the reannotation of the 2,001-instance evaluation sample. The sentence 'Two domain experts independently annotated the data' is insufficient. The authors should report Cohen's kappa or an equivalent agreement measure, describe the annotators' expertise and the annotation guidelines, and state the number and resolution of disagreements. Without this information, the reliability of the reannotated ground truth—and hence the validity of every comparison—cannot be assessed.","section":"Section 2.3.1"},{"comment":"All results, including the main comparison and the ablations, come from a single evaluation run. The agent uses stochastic LLM inference (qwen2.5:7b and Llama-3-8B) and tool calls, so run-to-run variability is expected. No confidence intervals, significance tests (e.g., McNemar or bootstrap), or repeated runs are reported. The observed gap between WATCHED (0.9139) and Distil MetaHate (0.8807) could be within this variability. The authors should report means and standard deviations over multiple runs and apply a significance test to the key comparisons.","section":"Section 2.3.4 and Section 2.3.5"}],"minor_comments":[{"comment":"The tables are numbered 'Table 2.3.4' and 'Table 2.3.5' in the text; they should be numbered continuously (e.g., Table 2 and Table 3).","section":"Tables 2 and 3"},{"comment":"Typo: 'W ATCHEDa chatbot' should be 'WATCHED, a chatbot'.","section":"Abstract"},{"comment":"The Llama 3 8B and Llama 3 70B baselines are described as 'general-purpose LLMs', but no prompting strategy or number of few-shot examples is specified. This information should be provided in Section 2.3.3 for reproducibility.","section":"Section 2.3.2"},{"comment":"The Perspective API output is binarized with a 0.5 threshold without justification. Report sensitivity to this threshold or cite a standard choice.","section":"Section 2.3.2"},{"comment":"Typo: 'whether the input constitutes hate speech or note' should be 'or not'.","section":"Section 5"},{"comment":"The figure is difficult to read at the current resolution; a vector version would improve clarity.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about baseline scores is well-founded: the exact match to the original publications is strong evidence that the baselines were not rerun on the reannotated labels. Because the authors have released the reannotated labels, rerunning the baselines is a feasible fix and not grounds for rejection. I would want to see not only the rerun but also the inter-annotator agreement and some measure of run-to-run variability before accepting the central claim. This is a system paper; the SoftwareX scope fits but the evaluation rigor is currently below the bar."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the WATCHED architecture is a reasonable agentic integration for hate speech, and the authors deserve credit for shipping code and the reannotated labels. But the central claim—that it surpasses state-of-the-art—doesn't hold as reported. The numbers in Table 2 for MetaHateBERT and Distil MetaHate are identical to those the paper itself says 'previously led the benchmarks' (0.8801 and 0.8807). If the test set truly was reannotated, any changed label would move those scores at least slightly. Exact equality means the baselines were scored on the original labels, while WATCHED was scored on the reannotated subset. That is a different ground truth, so the comparison is apples-to-oranges.\n\nThe paper does have real strengths. Combining a BERT classifier, RAG over a large hate corpus, Urban Dictionary lookups, a distilled reasoning model, and policy-grounded explanations is a new mix for this task. The ablation study gives a clear picture of each tool's contribution—MetaHateBERT removes the most, and the reasoning model also matters. The human-in-the-loop data expansion is a sensible practical feature, and the open repository plus released labels is reproducibility-friendly.\n\nThe soft spots beyond the label mismatch: no inter-annotator agreement is reported for the 2,001-example reannotation, so we don't know how stable those labels are. There is also no variance or multiple seed runs, which matters for an LLM-driven pipeline with a single run. And the evaluation omits a simple kNN or ensemble baseline using the same RAG database, so we can't tell how much of the gain is from retrieval alone versus the agent orchestration. The circularity—the whole stack is trained on MetaHate—is not disqualifying by itself, but it makes the invalid comparison worse.\n\nWho is this for? Practitioners building moderation support tools. The architecture description and ablation are useful. The SOTA claim, as written, is not supportable. A serious referee should ask for a rerun of all baselines on the reannotated labels, IAA for the reannotation, variance across runs, and a retrieval-only baseline. If those are fixed, this would be a solid software paper. I'd send it to review, because the system itself is coherent and the artifacts are real, but this needs to be fixed before the claim stands.","headline":"The system is a sensible engineering integration, but the SOTA claim is unsupported: Table 2's baseline numbers are the original MetaHate scores, not recomputed on the reannotated labels.","tokens_in":12374,"tokens_out":3299,"would_cite":false,"duration_ms":35490,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An AI agent that combines four tools beats standalone hate-speech detectors, reaching macro F1 of 0.91.","keywords":["hate speech detection","AI agent","retrieval-augmented generation","large language models","content moderation","explainability","chain-of-thought","MetaHate"],"falsifier":"Run WATCHED and all baselines on the reannotated 2,001-example test set using the released labels and also using the original MetaHate labels, and report the inter-annotator agreement between the two experts; if the baselines tie or beat WATCHED on the original labels, or if the new labels disagree with the original labels beyond expected noise, the paper's margin disappears.","tokens_in":11524,"feed_emoji":"🛡️","tokens_out":2920,"duration_ms":33857,"temperature":0.7,"pith_summary":"This paper presents WATCHED, a chatbot-style AI agent designed to help content moderators decide whether a social media post is hate speech. Instead of relying on a single model, the agent chooses per query which tools to use: a BERT-based hate-speech classifier, retrieval of similar labeled posts from a large database, a slang-dictionary lookup, and a distilled reasoning model, then produces a label with an explanation grounded in platform policies. The authors claim this combination reaches a macro F1 of 0.91 on a reannotated 2,001-post evaluation subset, outperforming MetaHateBERT, Distil MetaHate, Llama-3 baselines, and Perspective API. The result matters because hate-speech detection is notoriously brittle on slang, implicit language, and context, and human moderators need decisions that are both accurate and explainable.","feed_headline":"AI agent beats hate-speech baselines with macro F1 of 0.91","feed_subtitle":"WATCHED combines a BERT classifier, retrieved examples, slang lookup, and reasoning to outscore specialized models—with explanations.","key_machinery":"The carrying mechanism is the agent itself: a small instruction-tuned LLM (qwen2.5:7b-instruct) that decides which tools to call for each input. The tools are a MetaHateBERT hate-speech classifier, a Qdrant vector search over 1,164,586 MetaHate posts, an Urban Dictionary lookup for informal or evolving language, a call to the distilled reasoning model Distil MetaHate, and a guidelines database used to attach policy references to explanations. The agent synthesizes tool outputs and chain-of-thought reasoning into a binary label, a confidence score, and an explanation.","core_discovery":"The central claim is that an LLM-driven agent with access to four specialized tools outperforms every standalone baseline it is compared against, including the previous state-of-the-art hate-speech models, on the same evaluation data. WATCHED reaches a macro F1 of 0.9139 versus 0.8807 for Distil MetaHate and 0.8801 for MetaHateBERT. Ablation results show that removing any single tool lowers performance: dropping the BERT classifier causes the largest drop (macro F1 0.8571), removing similar-post retrieval gives 0.8733, removing Urban Dictionary gives 0.8753, and removing the reasoning tool gives 0.8853; removing all tools drops macro F1 to 0.8053. The authors interpret this as evidence that","pith_inferences":["The reported margin may depend on the reannotation: if the new labels systematically favor WATCHED's outputs, or if baseline numbers were taken from prior publications rather than recomputed on the same labels, the 0.91 figure could overstate the gain; this can be checked directly with the released labels.","The Urban Dictionary tool is intended to reduce false positives on AAVE and informal dialects, but its effect is not measured separately by dialect; a controlled error-rate analysis by language variety would test that mechanism.","The reasoning tool takes about 30 seconds per query, which is a practical bottleneck for real-time moderation; a faster distilled reasoner or caching of repeated evidence could make the approach deployment-ready."],"forward_implications":["If the result replicates, moderator-support tools can offer higher accuracy and explanations in a single interface, rather than a black-box classifier plus a separate explanation step.","Human feedback on each decision can be written back into the retrieval database, letting the system adapt to new slang and edge cases without retraining.","The ablation ordering suggests that a strong domain-specific classifier is the largest single contributor, but retrieval and slang lookup add measurable value—so simpler pipelines can be improved by adding these components.","The modular design implies the same agent pattern can be adapted to other online harms by swapping the classifier and guideline sources."],"supporting_citations":[{"why":"Supplies the MetaHate dataset used for retrieval and for training MetaHateBERT, the classifier tool.","marker":"[31]"},{"why":"Provides the Distil MetaHate reasoning model and the experimental settings and baseline numbers the evaluation extends.","marker":"[32]"},{"why":"Defines retrieval-augmented generation, the method behind the similar-posts tool.","marker":"[33]"},{"why":"Motivates iterative, agent-style retrieval over one-shot RAG.","marker":"[34]"},{"why":"Documents racial bias in hate-speech models, motivating the slang and AAVE lookup tool.","marker":"[42]"},{"why":"Shows how datasets can encode racial bias, further justifying the Urban Dictionary component.","marker":"[43]"},{"why":"The Perspective API is the production-tool baseline the agent must beat.","marker":"[44]"},{"why":"Jina embeddings v3 is the encoder used to build the retrieval index for similar posts.","marker":"[36]"}],"fun_headline_variants":["WATCHED AI agent beats hate-speech baselines, hits 0.91 F1","Four-tool agent overpowers hate-speech models","Explainable hate-speech AI exceeds state of the art","AI agent with tools tops hate-speech macro F1 at 0.91","Retrieval, slang, BERT, reasoning: agent triumphs over hate-speech baselines"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The comparison rests on the 2,001-example reannotation: if the two experts' new labels drifted toward WATCHED's outputs, or if baseline scores were computed on different labels than the agent's, the claimed superiority would not hold.","fun_headline_variants_meta":{"raw":{"variants":["WATCHED AI agent beats hate-speech baselines, hits 0.91 F1","Four-tool agent overpowers hate-speech models","Explainable hate-speech AI exceeds state of the art","AI agent with tools tops hate-speech macro F1 at 0.91","Retrieval, slang, BERT, reasoning: agent triumphs over hate-speech baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000689,"raw_usage":{"total_tokens":2988,"prompt_tokens":802,"completion_tokens":2186,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":2099}},"tokens_in":546,"tokens_out":2186,"duration_ms":21498,"temperature":1.0,"reasoning_tokens":2099,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:35:44.758467+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run WATCHED and all baselines on the reannotated 2,001-example test set using the released labels and also using the original MetaHate labels, and report the inter-annotator agreement between the two experts; if the baselines tie or beat WATCHED on the original labels, or if the new labels disagree with the original labels beyond expected noise, the paper's margin disappears.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Distil MetaHate reasoning model and the experimental settings and baseline numbers the evaluation extends."},{"cited_title":"Lewis, E","cited_arxiv_id":null,"evidence_quote":"Defines retrieval-augmented generation, the method behind the similar-posts tool."},{"cited_title":"co/docs/smolagents/en/examples/rag, [Accessed 27-05-2025] (2025)","cited_arxiv_id":null,"evidence_quote":"Motivates iterative, agent-style retrieval over one-shot RAG."}],"review_version":1}