Research map / Large language models / AI and society
Position: TrustLLM: Trustworthiness in Large Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TrustLLM, a position paper with eight trustworthiness principles, a six-dimension benchmark of 30+ datasets, and an evaluation of 16 mainstream LLMs on trustworthiness.
Problem
Trustworthiness of large language models lacks a comprehensive, unified set of principles and a practical benchmark for assessing how trustworthy mainstream LLMs are across dimensions.
Approach
Defines eight trustworthiness dimensions (truthfulness, safety, fairness, robustness, privacy, machine ethics, transparency, accountability), builds a benchmark over six of them with existing and new datasets, and evaluates 16 proprietary and open-weight LLMs, separating utility from trustworthiness.
Key results
- Proprietary LLMs such as GPT-4 generally outperform most open-weight LLMs in trustworthiness, yet Llama2 surpasses some proprietary models in many tasks.
- GPT-4 achieves only about 65% overall accuracy on stereotype recognition, and the best model shows a 0.5% agreement rate with stereotypes.
- Llama2-7b shows a 57% refusal rate on harmless prompts, indicating over-alignment that hurts utility.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Llama2-7b rather than a less over-aligned model | LLMs responding to benign prompts · Benign prompts | refusal rate: 57% | Hurt |
| Use GPT-4 rather than other evaluated LLMs | LLMs on stereotype recognition · Stereotype recognition | accuracy: 65% | Helped |
| Use a top-performing model rather than the least effective model | LLMs under perturbation · Open-ended tasks | average semantic similarity: 88% -> 97.64% | Helped |
| Use GPT-4 rather than the least effective models | LLMs on OOD detection · OOD detection | RtA rate: 0.4% -> over 80% | Helped |
| Use GPT-4 rather than the least effective models | LLMs on OOD generalization · OOD generalization | F1 score: around 30% -> over 92% | Helped |
| Use ChatGPT rather than Oasst-12b | LLMs judging privacy information use · Privacy use judgment | Pearson correlation with humans: 0.665; Oasst-12b less than zero | Helped |
| Apply privacy-policy instructions | LLMs responding to private-information queries · Privacy awareness | likelihood of refusing to respond | Helped |
| Use proprietary LLMs rather than most open-weight LLMs | LLMs across trustworthiness tasks · TRUSTLLM benchmark | trustworthiness | Helped |
| Increase model parameter size | Llama2 models · OOD performance | OOD performance | No clear effect |
| Use Llama2-series models rather than models with lower refusal rates | LLMs in high-ambiguity ethical scenarios · Implicit ethics, high-ambiguity scenarios | RtA: 99.9%; others less than 70% | Mixed |
Surprising findings
- Llama2 open-weight models surpass several proprietary LLMs in trustworthiness without external moderators.
- Many LLMs are over-aligned, refusing benign prompts (e.g., Llama2-7b refuses 57% of harmless prompts).
Limitations
- Evaluation is solely in English, limiting cross-lingual generalizability.
- Transparency and accountability dimensions are not benchmarked.
- No rigorous certification of worst-case LLM trustworthiness; prompt sensitivity and instruction-following affect results.
Experimental setup
Datasets: TruthfulQA AdvGLUE ETHICS Do-Not-Answer XSTest Enron Email Dataset StereoSet CrowS-Pairs SQuAD 2.0 HotpotQA
Benchmarks: TruthfulQA AdvGLUE ETHICS Do-Not-Answer XSTest Enron Email Dataset StereoSet CrowS-Pairs
Metrics: accuracy refuse-to-answer rate attack success rate micro F1 Pearson correlation semantic similarity
Baselines: ChatGPT GPT-4 PaLM 2 ERNIE Llama2-70b Vicuna-33b
Builds on: TruthfulQA AdvGLUE ETHICS Do-Not-Answer XSTest
Closest papers by meaning
- Decoding Compressed Trust: Scrutinizing the Trustworthiness of Efficient LLMs Under Compression (ICML 2024 · similarity 0.90)
- HonestLLM: Toward an Honest and Helpful Large Language Model (NeurIPS 2024 · similarity 0.89)
- Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference (ICML 2024 · similarity 0.89)
- Enhancing Multiple Dimensions of Trustworthiness in LLMs via Sparse Activation Control (NeurIPS 2024 · similarity 0.89)
- MultiTrust: A Comprehensive Benchmark Towards Trustworthy Multimodal Large Language Models (NeurIPS 2024 · similarity 0.89)
- To trust or not to trust: Attention-based Trust Management for LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.