This page is still being polished. If you have thoughts, please share them via the feedback form.
Data on this page is preliminary and may change. Please do not share or cite these figures publicly.
Input validation, output filtering, and content moderation classifiers.
Also in Non-Model
An essential step of the output safeguard is to detect undesirable content.
To do this, two open-source Python packages — Guard [358] and Guardrails [359], are developed to check for sensitive information in the generated content. Additionally, Azure OpenAI Service [360] integrates the ability to detect different categories of harmful content (hate, sexual, violence, and self-harm) and give a severity level (safe, low, medium, and high). Furthermore, NeMo Guardrails [223] — an open-source software developed by NVIDIA, can filter out undesirable generated texts and restrict human-LLM interactions to safe topics. Generally, the detectors are either rulebased [361], [362] or neural network-based [363]–[365], and the latter can better identify cryptic harmful information [366]. In practice, developers of GPT-4 leverage the LLM itself to construct a harmful content detector [367]. The user guide of LLaMA2 [368] suggests building the detectors with block lists and trainable classifiers. For the untruthful generated content, the most popular detectors are either fact-based or consistencybased. Specifically, the fact-based methods resort to external knowledge [369]–[371] and given context [372], [373] for fact verification, while the consistency-based methods generate multiple responses for probing the LLM’s uncertainty about the output [374]–[378]. We suggest readers refer to the surveys [107], [379] for more comprehensives summarization.
Reasoning
Output filtering detects and blocks harmful content using rule-based and neural classifiers before user delivery.
Mitigation in Input Modules
Mitigating the threat posed by the input module presents a significant challenge for LLM developers due to the diversity of the harmful inputs and adversarial prompts [209], [210].
1.2.1 Guardrails & FilteringMitigation in Input Modules > Defensive Prompt Design
Directly modifying the input prompts is a viable approach to steer the behavior of the model and foster the generation of responsible outputs. This method integrates contextual information or constraints in the prompts to provide background knowledge and guidelines while generating the output [22].
1.2.1 Guardrails & FilteringMitigation in Input Modules > Malicious Prompt Detection
Different from the methods of designing defensive prompts to preprocess the input, the malicious prompt detection method aims to detect and filter out the harmful prompts through the input safeguard
1.2.1 Guardrails & FilteringMitigation in Language Models
This section delves into mitigating risks associated with models, encompassing privacy preservation, detoxification and debiasing, mitigation of hallucinations, and defenses against model attacks.
1.1 ModelMitigation in Language Models > Privacy Preserving
Privacy leakage is a crucial risk of LLMs, since the powerful memorization and association capabilities of LLMs raise the risk of revealing private information within the training data. Researchers are devoted to designing privacypreserving frameworks in LLMs [226], [227], aiming to safeguard sensitive PII from possible disclosure during humanmachine conservation
1.1 ModelMitigation in Language Models > Detoxifying and Debiasing
To reduce the toxicity and bias of LLMs, prior efforts mainly focus on enhancing the quality of training data and conducting safety training.
1.1 ModelRisk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems
Cui, Tianyu; Wang, Yanling; Fu, Chuanpu; Xiao, Yong; Li, Sijia; Deng, Xinhao; Liu, Yunpeng; Zhang, Qinglin; Qiu, Ziyi; Li, Peiyang; Tan, Zhixing; Xiong, Junwu; Kong, Xinyu; Wen, Zujie; Xu, Ke; Li, Qi (2024)
Despite their impressive capabilities, large lan- guage models (LLMs) have been observed to generate responses that include inaccurate or fabricated information, a phenomenon com- monly known as “hallucination”. In this work, we propose a simple Induce-then-Contrast De- coding (ICD) strategy to alleviate hallucina- tions. We first construct a factually weak LLM by inducing hallucinations from the original LLMs. Then, we penalize these induced hallu- cinations during decoding to enhance the fac- tuality of the generated content. Concretely, we determine the final next-token predictions by amplifying the predictions from the orig- inal model and downplaying the induced un- truthful predictions via contrastive decoding. Experimental results on both discrimination- based and generation-based hallucination eval- uation benchmarks, such as TruthfulQA and FACTSCORE, demonstrate that our proposed ICD methods can effectively enhance the factu- ality of LLMs across various model sizes and families. For example, when equipped with ICD, Llama2-7B-Chat and Mistral-7B-Instruct achieve performance comparable to ChatGPT and GPT4 on TruthfulQA, respectively.
Build and Use Model
Training, fine-tuning, and integrating the AI model
Developer
Entity that creates, trains, or modifies the AI system
Measure
Quantifying, testing, and monitoring identified AI risks
