Abstract
Fairness in toxicity classification involves three integrated axes: ranking, calibration, and abstention. Training-time interventions and post-hoc safety mechanisms cannot be evaluated independently because the former determines the efficacy of the latter. We compare Empirical Risk Minimization (ERM), instance-level reweighting, and Group DRO across these axes, combined with temperature scaling, confidence-based abstention, and per-identity threshold optimization. Evaluation uses subgroup AUC, BPSN/BNSP AUC, error gaps, and per-subgroup Expected Calibration Error (ECE) with bootstrap CIs (n=1000). We report four findings. (1) Calibration disparity is a hidden fairness violation. ERM has near-perfect aggregate calibration (0.013) but is significantly miscalibrated across all identity subgroups (+0.029 to +0.134). (2) Training interventions reshape rather than eliminate disparity. Reweighted ERM improves ranking (BPSN AUC +0.06 to +0.12) but worsens the calibration-fairness gap by up to +0.232. Group DRO eliminates calibration disparity but only by becoming uniformly miscalibrated globally (ECE 0.118). (3) Post-hoc methods inherit training failure modes. Temperature scaling fails because miscalibration is non-uniform. Confidence-based abstention works under ERM but breaks under DRO, where the risk-coverage curve rises with deferral. (4) Abstention itself is unfair. Confidence-based deferral helps background content far more than identity-mentioning content. We argue that SRAI fairness requires a multi-axis framework: methods that differ only in aggregate ranking can differ sharply in failure modes that determine real-world harm.
Explore similar work
Jul 25, 2026cs.CL
Reducing toxicity is often framed as a global alignment problem, yet perceptions of harmful language are subjective and context-dependent. We present the first comparative evaluation of training-free methods for aligning language generation to user-specific toxicity sensitivities across three inference-time intervention stages: pre-decoding (prompt conditioning and rewriting), in-decoding (token, logit, and representation steering), and post-decoding (candidate re-ranking). Evaluated against toxicity sensitivity targets derived from the PRISM dataset, all methods reduce alignment error by 28-47%. However, the results reveal a fundamental trade-off between alignment effectiveness, personalization, and general language quality, showing how toxicity sensitivity alignment is an inherently multi-objective problem.
Rares A. C. Diaconescu, Iulia Slanina, Alina Florea +5
Jan 14, 2026cs.CL
Large language models can produce toxic or inappropriate text even for benign inputs, creating risks when deployed at scale. Detoxification is therefore important for safety and user trust, particularly when we want to reduce harmful content without sacrificing the model's generation quality. Many existing approaches rely on model retraining, gradients, or learned auxiliary components, which can be costly and may not transfer across model families or to truly black-box settings. We introduce a test-time procedure that approximates the gradient of completion toxicity with respect to the input embeddings and uses a small number of descent steps to steer generation toward less toxic continuations. This is achieved with zeroth-order optimization that requires only access to input embeddings, a toxicity scoring function, and forward evaluations of the model. Empirically, the approach delivers robust toxicity reductions across models and prompts and, in most settings, achieves the best overall toxicity-quality trade-off. More broadly, our work positions word embeddings as effective control variables and encourages wider use of black-box optimization to guide autoregressive language models toward scalable, safer text generation, without requiring any training or access to intermediate computations.
Baturay Saglam, Dionysis Kalogerias
May 27, 2026cs.CL
Large language models frequently generate toxic, hateful, or harmful content, yet existing mitigation methods rely on costly retraining or output-level filtering with no mechanistic insight into where toxicity originates internally. We introduce Meow2X and TRNE, two complementary retraining-free frameworks that localize toxicity to specific layers and neurons by analyzing activation differentials between toxic and neutral prompts, then suppress them via inference-time scaling or minimal rank-one weight edits -- without any gradient descent. Evaluations across five LMs, two benchmarks, and 90 configurations using dual safety evaluators demonstrate consistent toxicity reduction while preserving language modeling quality. Our analysis reveals that toxicity is disproportionately encoded in early MLP layers, varies across architectures, and is systematically underestimated by single-evaluator setups -- underscoring the need for multi-evaluator safety assessment. By bridging mechanistic interpretability with practical detoxification, our framework offers a principled path toward safer, more transparent language models.
Himanshu Beniwal, Mayank Singh