VARM-Bench provides an auditable and reproducible benchmark for evaluating verifiable moderation rationales in Chinese abusive-speech moderation, and shows that strong label-level performance can conceal substantial errors in complete moderation records.
Abstract
The widespread circulation of abusive online content has increased the need for reliable moderation of Chinese social-media text. Existing Chinese benchmarks support label classification, fine-grained toxicity categorization, and target-aware extraction, but do not provide a unified representation for deterministically verifying the stated basis of a moderation decision. We introduce VARM-Bench, a benchmark for field-anchored chain-of-thought rationales in Chinese abusive-speech moderation. Each instance contains a concise natural-language rationale with explicit anchors for six decisions: target, target type, target explicitness, author stance, harmfulness label, and fine-grained category. Our deterministic protocol evaluates field correctness, target alignment, output validity, complete-record agreement, and hidden record errors conditioned on correct final decisions, without relying on an LLM judge. Under a common structured-output protocol, we evaluate language models across multiple model families using zero-shot prompting, taxonomy guidance, and structured CoT supervision, and analyze lexical-cue sensitivity and field-level errors. Results show that strong label-level performance can conceal substantial errors in complete moderation records. VARM-Bench provides an auditable and reproducible benchmark for evaluating verifiable moderation rationales in Chinese abusive-speech moderation.
LLM-based evaluators of natural language generation (NLG) quality are widely deployed as scoring tools and as automated training signals, yet the internal procedure by which they assign a rating remains poorly understood. We investigate this procedure mechanistically through an eight-attack perturbation taxonomy across the Readability and Adequacy dimensions of NLG quality, a generation pipeline that produces paired clean and corrupt summaries with controlled error intensity and explicit token-level modification maps, and a four-experiment battery of causal tracing, logit-lens vocabulary projection, and attention-head knockout applied to Themis (Llama-3-8B) and Prometheus (Mistral-7B). Both evaluators implement a structured, coherent evaluation pipeline operating in two stages: below layer 15, attention performs local error comparison and routes the result to the final input position; above it, the MLP cascade integrates the signal and writes the rating, with the decision crystallizing in the residual stream at a sharp late layer (L = 26 on Themis, L = 25 on Prometheus). Furthermore, a base-model control at the same scale (Llama-3-8B) reproduces the routing architecture and crystallization but not the stage separation, isolating the two mechanisms that fine-tuning specifically installs, suppression of below-L15 MLP contribution at the last position and a two-layer advance of the crystallization depth, indicating that fine-tuning sculpts an existing substrate rather than building the pipeline from scratch. We release the source code and data at https://github.com/himil-v/judge-mech
The wide adoption of conversational AI systems necessitates urgent and interpretable safety moderation, especially given that Large Language Models (LLMs) continue to exhibit vulnerabilities despite alignment efforts, posing significant risks to individual users, organisations, and society. The ideal AI safety moderation system must be transparent and structurally interpretable. However, current moderation approaches typically rely on coarse classifications that offer limited interpretability and fail to capture the nuanced intent and contextual dependencies present in real-world user inputs. To advance moderation beyond these coarse labels, we present REMEDy, a novel dataset specifically built for extracting fine-grained rationales from user prompts. REMEDy features span-level annotations covering a broad taxonomy of safety-relevant categories, allowing for overlapping and nested textual spans to reflect complex prompt structures. Using REMEDy, we fine-tune multiple LLMs and evaluate their performance across two tasks: (i) rationale extraction, assessing their ability to accurately localise and classify harmful or ambiguous content; and (ii) prompt moderation, measuring improvements over state-of-the-art safety detectors. Our experiments demonstrate that REMEDy-trained models achieve competitive or superior moderation outcomes while simultaneously providing structured, human-readable rationales. REMEDy thus offers a valuable resource for developing safer, more transparent, and context-sensitive moderation systems.
L. piano, Claudia Battistin, Jeriek Van den Abeele et al.· Neural computing & applicati...· 0 citations
Hate speech research has moved from coarse-grained classification towards structured parsing, where systems jointly identify targets, supporting arguments, and target-level labels. Documents with multiple targets, conflicting local readings, or culturally coded language make these bindings difficult to recover. SPAR-Hate is an auditor-guided multi-perspective role-reasoning framework for bilingual hate speech parsing. It decomposes each document into local focus units, elicits evidence-grounded candidates from Victim, Moderator, and Cultural Bystander perspectives, resolves candidate conflicts under grounding and schema constraints, and reassembles sample-level predictions. Experiments on STATE-ToxiCN and a controlled TBO split show gains across local and API backbones, concentrated on strict joint target-argument-label metrics. Full-test integrated-prompt controls, component ablations, and bounded-arbitration diagnostics identify the contribution of separated perspective generation and arbitration. Structured teacher traces also support training a smaller student model.
Yifan Lyu, Dianqing Lin, Xinran Li et al.· 0 citations
This work introduces MMDS-Bench, a diagnostic benchmark for multimodal dynamic stance classification in social media parent-reply interactions, and evaluates 12 closed-source and open-source multimodal large language models and proposed reference-grounded LLM-judge protocol for assessing reasoning quality.
Transformer-based models such as BERT, RoBERTa, DistilBERT, and DeBERTa have achieved remarkable performance across a wide range of natural language processing (NLP) tasks. However, their decision-making processes remain difficult to interpret, particularly in high-risk applications such as hate speech detection, where unreliable explanations may undermine model transparency, trust, and accountability. This study investigates whether explainability methods remain faithful and stable under domain shift in transformer-based text classification. Four transformer architectures were fine-tuned and evaluated on two linguistically distinct datasets: IMDb Movie Reviews and Hate Speech Offensive. Model performance and explanation quality were assessed using classification accuracy, macro F1-score, top-k token-removal faithfulness analysis, and cross-domain Spearman rank correlation. Experimental results show that DeBERTa achieved the highest classification performance, reaching accuracies of 95.6% on IMDb and 91.3% on Hate Speech. Across all evaluated models and datasets, SHAP consistently produced higher faithfulness scores than attention-based explanations. Cross-domain analysis further revealed reduced agreement between SHAP and attention-based explanations under domain shift, indicating lower explanation consistency across linguistically distinct domains. Qualitative error analysis further showed that implicit sentiment, sarcasm, and domain-specific slang remain major sources of prediction errors. Overall, the results demonstrate that superior predictive performance does not necessarily correspond to higher explanation faithfulness or stronger cross-domain stability. These findings highlight the importance of jointly evaluating predictive performance, explanation faithfulness, and explanation robustness when developing trustworthy transformer-based NLP systems.
Dony Bahtera Firmawan, B. Darnoto· Journal of Computing Theorie...· 1 citation
A survey and comparative analysis of NLP-based Automatic Deception Detection focusing on the legal domain and the evolution from feature-based machine learning to Large Language Model (LLM) approaches are presented, showing strong domain sensitivity.
T. Samaradiwakara, Nisansa de Silva, George C. Lobb· arXiv.org· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
