DART: Mitigating Harm Drift in Difference-Aware LLMs via Distill-Audit-Repair Training

• URL: http://arxiv.org/abs/2604.16845v1
• Date: Sat, 18 Apr 2026 05:28:53 GMT
• Title: DART: Mitigating Harm Drift in Difference-Aware LLMs via Distill-Audit-Repair Training
• Abstract summary: Large language models (LLMs) tuned for safety often avoid acknowledging demographic differences.<n>This identity-blindness yields incorrect responses, unnecessary refusals, or generic "equal-treatment" defaults.<n>We introduce DART (Distill--Audit-Repair Training), which distills label-conditioned reasoning from a teacher.
• Score: 18.22030439716779
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large language models (LLMs) tuned for safety often avoid acknowledging demographic differences, even when such acknowledgment is factually correct (e.g., ancestry-based disease incidence) or contextually justified (e.g., religious hiring preferences). This identity-blindness yields incorrect responses, unnecessary refusals, or generic "equal-treatment" defaults. We study this via difference-awareness classification: given a question involving demographic groups, the task is not to answer directly, but to classify whether a correct answer requires recognizing group differences (yes) or whether groups should be treated identically (no). Crucially, fine-tuning for accuracy triggers harm drift: model-generated explanations become increasingly harmful as decision accuracy improves, whether by elaborating harmful content, introducing problematic assumptions, or failing to flag harms the baseline identified. To mitigate this, we introduce DART (Distill--Audit--Repair Training), which distills label-conditioned reasoning from a teacher, audits outputs for harm drift cases relative to baseline, and repairs problematic cases via severity-weighted fine-tuning. On eight benchmarks, DART improves Llama-3-8B-Instruct accuracy from 39.0% to 68.8%, with largest gains on equal-treatment prompts (11.3% -> 72.6%), while reducing harm drift cases by 72.6%. It also transfers to 280 open-ended real-world queries across medical, legal, policy, and educational domains, improving difference-appropriate responses from 39.8% to 77.5% while reducing refusals from 34.3% to 3.0%. Our results demonstrate that accuracy and safety need not conflict when explicit detection and repair mechanisms are in place.

Related papers

• IdeaAMBIG: Benchmarking Implementation-Critical Gaps in Research-Idea Specifications [52.570108663867046]
  A research idea may be novel, coherent, and scientifically plausible, yet its proposed method may remain insufficiently specified for faithful implementation.<n>We study the codification readiness of implementation-facing research-method specifications, defined by whether they provide sufficient methodological information for a competent implementer or coding agent to construct the intended method without unsupported assumptions.<n>We introduce IdeaAMBIG, a benchmark of 660 evidence-grounded instances: 163 real-world gaps from reports and GitHub issues, and 497 controlled synthetic gaps injected into codification-ready references.
  arXiv  Detail & Related papers  (2026-09-09T17:59:04Z)
• Safety for Whom? Boundary-Aware Self-Distillation for Controlled LLM Safety Refusal [1.8531140061496822]
  A civics tutor and a public-sector assistant may share a base model yet need different boundaries inside the same topic.<n>We formulate this as narrow-boundary safety and introduce an offline self-generated framework combining controlled topic generation, coverage repair, in-distribution compensation data, and harmful-benign pairs for training and evaluation.
  arXiv  Detail & Related papers  (2026-09-03T21:03:45Z)
• Instruction Duplication as an Inference-Time Control Primitive [0.0035960465625249592]
  Procedural instruction following is a basic requirement for controllable language-model systems.<n>We introduce instruction duplication, a minimal black-box inference-time control that repeats only the procedural instruction.
  arXiv  Detail & Related papers  (2026-09-03T16:05:21Z)
• Cybersecurity Detection Classification with Reasoning-enabled Language Models [50.72675435043546]
  A major issue in Security Operations Centers (SOCs) is alert fatigue.<n>Prior work prompts or fine-tunes large language models (LLMs) to emit a triage label directly, but does not train them to reason about whether a detection is a genuine threat.<n>We train a chain-of-thought (CoT) reasoning-enabled triage classifier on real, human-labeled Windows endpoint detections.
  arXiv  Detail & Related papers  (2026-07-30T16:22:13Z)
• LLM Doesn't Know What It Doesn't Know: Detecting Epistemic Blind Spots via Cross-Model Attribution Divergence on Clinical Tabular Data [2.1443570696048906]
  Large language models (LLMs) are increasingly applied to structured clinical data.<n>We compare Qwen 2.5 7B and XGBoost on a prediction task via attribution divergence analysis.
  arXiv  Detail & Related papers  (2026-06-17T18:49:44Z)
• Better Accuracies, Worse Reasoning: A Step-Level Audit of Medical Chain-of-Thought Distillation [14.101636730819175]
  Chain-of-thought (CoT) distillation trains a smaller model to imitate a teacher's reasoning trace.<n>We ask whether gains in answer quality are accompanied by improvements in the trace.
  arXiv  Detail & Related papers  (2026-05-27T10:55:24Z)
• Step-wise Rubric Rewards for LLM Reasoning [72.17879367869503]
  Reinforcement Learning with Verifiable Rewards (RLVR) is widely used to improve reasoning in large language models.<n>We find 18.2% of steps in correct responses are wrong yet positively rewarded.<n>Across six mathematical reasoning benchmarks, SRaR improves average accuracy over RaR by 3.57 points.
  arXiv  Detail & Related papers  (2026-05-17T07:08:14Z)
• Knowing but Not Correcting: Routine Task Requests Suppress Factual Correction in LLMs [26.062372963777452]
  LLMs reliably correct false claims when presented in isolation, yet when the same claims are embedded in task-oriented requests, they often comply rather than correct.<n>We term this failure mode emphcorrection suppression and construct a benchmark of 300 false premises to systematically evaluate it across eight models.<n> Suppression rates range from 19% to 90%, with four models exceeding 80%, establishing correction suppression as a prevalent and severe phenomenon.
  arXiv  Detail & Related papers  (2026-05-07T10:04:39Z)
• Measuring the Permission Gate: A Stress-Test Evaluation of Claude Code's Auto Mode [9.116800340266066]
  Claude Code's auto mode is the first deployed permission system for AI coding agents.<n>Anthropic reports a 0.4% false positive rate and 17% false negative rate on production traffic.<n>We evaluate 253 state-changing actions at the individual action level against oracle ground truth.
  arXiv  Detail & Related papers  (2026-04-04T17:56:30Z)
• Interpretability without actionability: mechanistic methods cannot correct language model errors despite near-perfect internal representations [1.6676380665893165]
  Language models encode task-relevant knowledge in internal representations that far exceeds their output performance.<n>Current mechanistic interpretability methods cannot reliably translate internal knowledge into corrected outputs.
  arXiv  Detail & Related papers  (2026-03-18T23:31:05Z)
• Residual Stream Analysis of Overfitting And Structural Disruptions [26.444437808784482]
  We show that safety data exhibits substantially lower token entropy and 2-gram diversity compared to general instruction data.<n>To uncover the root cause, we introduce FlowLens, a stable PCA-based tool for residual-stream geometry analysis.<n>We propose Variance Concentration Loss, an auxiliary regularizer that penalizes excessive variance concentration in mid-layer residuals.
  arXiv  Detail & Related papers  (2026-03-04T11:13:58Z)
• Reviewing the Reviewer: Graph-Enhanced LLMs for E-commerce Appeal Adjudication [0.7136933021609076]
  We propose a minimal representation for adjudication reasoning that prevents hallucination through operational grounding.<n>We develop a conflict-aware graph reasoning framework that performs top-down deductive reasoning for new cases.<n>We evaluate the framework in large-scale e-commerce seller appeal adjudication.
  arXiv  Detail & Related papers  (2026-02-27T00:43:59Z)
• High Accuracy, Less Talk (HALT): Reliable LLMs through Capability-Aligned Finetuning [84.52940628494879]
  Large Language Models (LLMs) currently respond to every prompt.<n>LLMs can produce incorrect answers when they lack knowledge or capability.<n>We propose post-training an LLM to generate content only when confident in its correctness and to otherwise abstain.
  arXiv  Detail & Related papers  (2025-06-04T15:16:21Z)
• Medical Hallucinations in Foundation Models and Their Impact on Healthcare [71.15392179084428]
  Hallucinations in foundation models arise from autoregressive training objectives.<n>Top-performing models exceeded 97% accuracy when augmented with chain-of-thought prompting.
  arXiv  Detail & Related papers  (2025-02-26T02:30:44Z)
• CausalDiff: Causality-Inspired Disentanglement via Diffusion Model for Adversarial Defense [61.78357530675446]
  Humans are difficult to be cheated by subtle manipulations, since we make judgments only based on essential factors.<n>Inspired by this observation, we attempt to model label generation with essential label-causative factors and incorporate label-non-causative factors to assist data generation.<n>For an adversarial example, we aim to discriminate perturbations as non-causative factors and make predictions only based on the label-causative factors.
  arXiv  Detail & Related papers  (2024-10-30T15:06:44Z)
• Uncertainty-Based Abstention in LLMs Improves Safety and Reduces Hallucinations [63.330182403615886]
  A major barrier towards the practical deployment of large language models (LLMs) is their lack of reliability. Three situations where this is particularly apparent are correctness, hallucinations when given unanswerable questions, and safety. In all three cases, models should ideally abstain from responding, much like humans, whose ability to understand uncertainty makes us refrain from answering questions we don't know.
  arXiv  Detail & Related papers  (2024-04-16T23:56:38Z)
• Cost-Sensitive Machine Learning Classification for Mass Tuberculosis Verbal Screening [0.0]
  The sensitivity of our identifier increased by 1.26% while specificity increased by 13.19% in absolute value compared to the traditional score-based method. Results indicate that even with limited data we can actually devise a better method to identify TB suspects from verbal screening.
  arXiv  Detail & Related papers  (2020-11-14T21:41:29Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.