Skip to content

harmreduction benchmarking llms in harm reduction information provision to suppo

content_hash: 437482716b1e83fa ---# HarmReduction: Benchmarking LLMs in Harm Reduction Information Provision to Support People Who Use Drugs

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://huggingface.co/datasets/RayK/Harm_Reduction_QA_dataset_basic
Area: LLM Safety
Keywords: harm reduction information provision, LLM safety benchmarking, public health evaluation, polysubstance use risk, retrieval-augmented generation

TL;DR

This paper introduces HarmReduction, the first benchmark assessing LLMs' accuracy and safety risks in providing harm reduction information to people who use drugs (PWUD), revealing critical failure modes in quantitative guidance, refusal-driven censorship, and severe underestimation of lethal polysubstance interactions.

Background & Motivation

Substance use poses severe global public health challenges, impacting the health and well-being of millions of individuals worldwide. In response, harm reduction has emerged as a cornerstone public health paradigm that prioritises pragmatic, non-judgemental, and evidence-based information to mitigate adverse health consequences and fatal risks for vulnerable groups, specifically People Who Use Drugs (PWUD). Facing pervasive social stigma, systemic marginalisation, and criminalisation, PWUD frequently cannot access timely or non-punitive support within conventional healthcare systems. Consequently, they turn to online resources such as organisational information portals and peer-led mutual-aid forums. However, static institutional websites struggle to dynamically adapt to volatile street drug market trends, while peer-led community forums suffer from volunteer fatigue, inconsistent coverage, and response delays during critical emergency windows. These persistent technical barriers suggest clear opportunities for emerging conversational AI systems.

Large language models (LLMs) with demonstrated medical reasoning capabilities hold significant promise for delivering 24/7, multilingual, and personalised harm reduction guidance. Nonetheless, deploying LLMs within this high-stakes domain uncovers profound socio-technical tensions. Commercial LLM safety guardrails are predominantly calibrated against mainstream societal norms and legal prohibitions that enforce an abstinence-only stance. Under these rigid guardrails, inquiries regarding controlled substances are frequently flagged and censored, leading models to issue outright refusals or generic moralising disclaimers. For an individual in an active substance-use scenario, withholding evidence-based safety guidance effectively cuts off lifesaving risk-mitigation pathways. Furthermore, model hallucinations regarding lethal dosage thresholds or dangerous drug combinations introduce immediate, life-threatening risks. Critically, standard medical and consumer-health QA benchmarks (e.g., MedQA, PubMedQA, MultiMedQA, HealthBench) focus on clinical licensing examinations or formal doctor-patient interactions; they assume clinical accessibility, omit non-judgemental harm reduction principles, and lack source-constrained evaluation criteria.

To bridge the gap between community health needs and model alignment audits, a dedicated benchmarking framework is required to rigorously quantify both factual reliability and asymmetric clinical risk. Core idea: construct HR-Basic, a source-linked benchmark of 2,160 question-answer-evidence triples across safety boundary checks, quantitative queries, and polysubstance risk inference, evaluating 13 state-of-the-art LLMs under Instruction and RAG schemes to uncover critical quantitative hallucinations, censorship-induced failures, and asymmetric safety risks.

Method

Overall Architecture

The HarmReduction benchmark evaluates LLMs' factual reliability (Q1) and safety risks (Q2) across basic harm reduction queries. The pipeline consists of two primary phases: (1) knowledge extraction and benchmark dataset (HR-Basic) construction from verified, expert-curated harm reduction sources; and (2) dual-scheme validation comparing models' internal parametric capabilities (Instruction Scheme) against domain-knowledge integration (RAG Scheme). Standardised natural language queries are parsed into structured outputs—binary safety decisions, numerical values with units, or multi-level risk categories—paired with source-grounded explanations.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Harm Reduction Source Acquisition<br/>DrugScience / Talk to Frank / Drugs and Me / TripSit"] --> B["Semantic Chunking & Triple Construction<br/>Cosine sim > 0.8 clustering + GPT-4o-mini extraction + dual-expert review"]
    B --> C["HR-Basic Benchmark Tasks<br/>2,160 question-answer-evidence triples"]
    C --> D["Dual Evaluation Schemes<br/>Parametric inference (Instruction) vs dense-sparse hybrid retrieval (RAG)"]
    D --> E["Task-Specific & Asymmetric Risk Metrics<br/>Response rate / tolerance-band accuracy / underestimation & severe underestimation"]

Key Designs

1. Semantic Chunking and Verbatim Triple Construction: Preserving Critical Safety Context

Standard fixed-length chunking arbitrarily fragments contextually intertwined safety guidelines, frequently separating a substance's dosage parameters from its corresponding toxic contraindications and warning signs. Injecting fragmented chunks into LLMs risks generating dangerously incomplete advice. To eliminate this issue, source documents are first segmented at the sentence level and merged into semantically coherent chunks based on embedding cosine similarity (threshold \(> 0.8\), bounded by a 350-word maximum). Next, GPT-4o-mini is deployed with highly structured prompt templates and negative constraints ("skip if information is ambiguous or not specified") to extract triplets containing a user-style question, an explicit directive answer, and exact verbatim source evidence. All 2,160 retained candidates underwent double-blind manual validation by two independent researchers—one with over four years of specialised harm reduction research experience and one non-specialist validating clarity—with any disagreement leading to immediate sample exclusion.

2. Multi-Tiered Task Formulation for PWUD Information Needs: Covering Core Risk Scenarios

HR-Basic spans three distinct query types that represent the primary information needs of PWUD: (1) Safety Boundary Check (1,638 pairs, 76%): Evaluates whether the model can establish unambiguous safety boundaries (e.g., determining whether driving after substance use is safe). Ground truth is strictly formatted as Yes/No followed by an explanation, mirroring real-world harm reduction practice where binary decisions govern immediate personal safety; (2) Quantitative Questions (372 pairs, 17%): Assesses the model's ability to provide precise numerical figures with units regarding threshold doses, onset latency, duration, or withdrawal timelines, avoiding vague estimates that could cause accidental toxicity; (3) Polysubstance Use Risks (150 pairs, 7%): Addresses lethal combination hazards derived from TripSit interaction matrices across six standardised labels (Dangerous, Unsafe, Caution, Low Risk & Synergy, Low Risk & No Synergy, Low Risk & Decrease), examining whether models can correctly infer synergistic drug interactions.

3. Dual Evaluation Schemes: Parametric Knowledge Probing and Hybrid RAG

To isolate models' inherent pre-trained knowledge from their capacity to leverage external domain references, two evaluation schemes were deployed. Under the Instruction Scheme, models rely solely on human-authored prompt instructions defining the task persona and structural output constraints:

\[R_{inst} = \arg\max_{r} p_M(r \mid Q, I)\]

Under the RAG Scheme, models are augmented with external corpora indexed via LlamaIndex (chunked into 250 tokens with 25-token overlap). Retrieval combines dense semantic search (Sentence-Transformers all-MiniLM-L6-v2) and sparse term matching (BM25 via ElasticSearch) to guarantee accurate retrieval of exact chemical nomenclature and dosage numbers. Top candidates are fused via Reciprocal Rank Fusion (RRF) with constant \(k=60\):

\[\text{score}(d) = \sum_{i} \frac{1}{k + \text{rank}_i(d)}\]

The top 3 reranked documents are supplied as context, formulating the generation target as \(R_{rag} = \arg\max_{r} p_M(r \mid Q, K, I)\).

4. Asymmetric Risk Stratification and Tolerance Metrics: Quantifying Safety Failures

In harm reduction, the consequences of classification errors are fundamentally asymmetric. Rather than relying on simple accuracy, the benchmark formalises an asymmetric risk evaluation framework. First, Response Rate measures whether an actionable directive (Yes/No, number with unit, or risk label) is generated. In this closed, source-grounded setting, refusals or generic moralising disclaimers are marked as failures because withholding vital guidance leaves PWUD unprotected. For quantitative queries, accuracy is evaluated across exact match (\(0\%\)) and relaxed tolerance bands (\(10\%\), \(25\%\), \(50\%\)). For polysubstance risks, prediction errors are stratified into Underestimation (classifying an interaction one level safer than reference), Severe Underestimation (classifying two or more levels safer, such as labelling Dangerous as Caution or Low Risk, which can cause fatal respiratory arrest), and Overestimation (overly cautious warnings).

Key Experimental Results

Main Results

The benchmark was evaluated on 13 state-of-the-art LLMs, including open-weight general models (Llama 3.3 70B, DeepSeek-R1-Distill-Llama-70B, Phi-3.5-MoE, Qwen3-32B, Gemma 3 27B), medical open-weight models (HuatuoGPT-o1-70B, Llama3-OpenBioLLM-70B), and proprietary API models (GPT-4.1, GPT-4o-mini, o3-mini, o4-mini, Gemini 3.1 Flash-Lite, Qwen3.7 Plus). Primary results for overall response rates and safety boundary checks are detailed below:

Model Overall Response Rate (Instruction) Overall Response Rate (RAG) Safety Boundary Accuracy (Instruction) Safety Boundary Accuracy (RAG) Safety Boundary F1 Score (RAG)
GPT-4.1 100.0% 99.1% (-0.9) 88.0% 93.0% (+5.0) 92.4%
GPT-4o-mini 99.8% 99.6% (-0.2) 88.4% 91.0% (+2.6) 90.2%
o4-mini 99.5% 99.8% (+0.3) 88.8% 94.6% (+5.8) 94.3%
o3-mini 99.0% 99.6% (+0.6) 87.7% 95.0% (+7.3) 94.8%
Gemini-3.1-Flash-Lite 99.6% 98.8% (-0.8) 88.7% 93.1% (+4.4) 92.4%
Qwen3.7-Plus 100.0% 99.5% (-0.5) 88.6% 94.3% (+5.7) 93.9%
HuatuoGPT-70B 99.8% 98.8% (-1.0) 88.2% 94.7% (+6.5) 94.4%
DeepSeek-R1-70B 97.4% 95.9% (-1.5) 88.0% 95.3% (+7.3) 95.0%
LLaMA-3.3-70B 99.4% 99.1% (-0.3) 86.3% 93.3% (+7.0) 92.7%
Phi-3.5-MoE 99.6% 99.9% (+0.3) 86.9% 88.2% (+1.3) 86.5%
Qwen-3-32B 96.4% 59.1% (-37.3) 81.7% 87.4% (+5.7) 89.7%
Gemma-3-27B 88.5% 71.6% (-16.9) 83.4% 83.1% (-0.3) 74.2%
OpenBio-70B 95.6% 90.6% (-5.0) 59.0% 55.1% (-3.9) 13.2%

Quantitative Questions & Polysubstance Risk Breakdown

Quantitative dosage queries and polysubstance risk assessments exposed acute model deficiencies. Below are quantitative accuracy distributions across error tolerance levels and representative findings from risk inference evaluations:

Model / Evaluation Setting 0% Tolerance (Exact) 10% Tolerance 25% Tolerance 50% Tolerance Mechanism & Safety Observation
o3-mini (Instruction) 28% 50% 66% 71% High parametric discrepancy; e.g., recommended 200 mg for ketamine K-hole, exceeding ground truth (150+ mg) by 50 mg
o3-mini (RAG) +15% gain - - 81% External context lifts 50% tolerance band to 81%, though exact numerical precision remains limited
HuatuoGPT-70B (Instruction) 27% 43% 53% 69% Medically specialised model correctly provides exact figures on barely over a quarter of queries
HuatuoGPT-70B (RAG) +26% gain - - 80% Dense-sparse retrieval delivers strong gains (+26% at 0% tolerance), demonstrating the necessity of grounding
DeepSeek-R1-70B (Instruction) 31% 48% 55% 76% Reasoning-optimised model achieves higher baseline scores, but exact match remains under one-third
DeepSeek-R1-70B (RAG) +28% gain - - 84% Attains top open-weight performance, yet still exhibits a 16% error rate at 50% tolerance
Qwen-3-32B (Instruction) 30% 39% 45% 63% Weaker parametric quantitative memory; lags across all tolerance thresholds
Qwen-3-32B (RAG) +15% gain - - 55% (50% tol.) Context injection degrades instruction following, dropping response rate significantly
o4-mini Polysubstance (Instruction) - - - - Demonstrates two-level severe underestimation (classifying opioids + ketamine as Low Risk under supervision)
OpenBio-70B Polysubstance (RAG) - - - - Response rate collapses to 0.0%, generating only empty reference placeholders

Key Findings

  • RAG mitigates severe underestimation at the cost of broader moderate underestimation: In polysubstance risk inference, external retrieval eliminated extreme Severe Underestimation instances (e.g., classifying opioid-ketamine combinations as low risk). However, this came at the expense of increasing mild Underestimation, as models frequently compromised by downgrading Dangerous/Unsafe pairings to Caution.
  • Embedded moderation mechanisms produce severe negative externalities: OpenBio-70B attained an accuracy of only 59.0% on safety boundary checks (falling to 55.1% under RAG). Qualitative inspection revealed that it answered almost all queries with "No" (unsafe). This rigid refusal mechanism renders the model functionally incapable of providing practical harm reduction guidance.
  • Model scale does not monotonically track harm reduction safety: Within the Qwen3 family, Qwen3-14B outperformed Qwen3-32B on safety boundary accuracy under both Instruction (87.8% vs. 82.2%) and RAG schemes (95.3% vs. 93.8%), proving that unguided scaling does not intrinsically resolve safety alignment in specialised domains.

Highlights & Insights

  • Reframing model refusal as a safety hazard: The authors challenge the conventional assumption that refusal constitutes a safe default. In harm reduction, a vulnerable user denied evidence-based guidance is left at severe risk of preventable overdose or toxic adulteration.
  • Asymmetric risk metrics for high-stakes AI: By formalising underestimation and severe underestimation metrics, the benchmark provides a concrete methodological template for auditing models in life-critical healthcare domains.
  • Robust semantic chunking methodology: Sentence-level semantic clustering prevented the fragmentation of interdependent dosage constraints and contraindications, establishing a high-quality, verified 2,160-item gold standard.

Limitations & Future Work

  • Author-acknowledged limitations: The HR-Basic corpus is restricted to English-language sources from the UK and US, lacking global demographic and policy diversity. Preliminary tests incorporating colloquial community slang (e.g., confusing "candy flip" with LSD + cocaine) induced factual confusion that could not be reliably controlled without extensive manual curation.
  • Unaddressed operational gaps: The current benchmark focuses on single-turn factual question answering. It does not evaluate dynamic multi-turn interactions, triage during acute overdose emergencies, or conversational empathy.
  • Promising directions: Developing specialized harm reduction guardrails that reliably distinguish between promoting illicit drug use and providing lifesaving harm reduction information, alongside evaluating multimodal capabilities such as colorimetric reagent test interpretation.
  • vs MedQA / PubMedQA: Prior medical benchmarks evaluate professional medical examination questions or biomedical research abstracts, assuming formal clinical settings. HarmReduction directly addresses marginalised individuals navigating risks outside formal healthcare systems.
  • vs HealthBench: While HealthBench evaluates open-ended conversational healthcare dialogue with physician rubrics, HarmReduction establishes closed, source-constrained directives with rigid safety boundary thresholds and tolerance intervals.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ [Pioneers the intersection of LLM safety benchmarking and public health harm reduction for marginalised populations]
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Thorough evaluation across 13 models, dual Instruction/RAG schemes, semantic explanations, and asymmetric risk stratification]
  • Writing Quality: ⭐⭐⭐⭐⭐ [Flawlessly structured, ethically rigorous, and technically precise in articulating socio-technical trade-offs]
  • Value: ⭐⭐⭐⭐⭐ [Provides a crucial benchmark dataset and clear empirical guardrails to prevent AI-induced harms in substance-use health provision]