[Paper Review] Alleviating Hallucinations of Large Language Models through Induced Hallucinations

Yue Zhang, Leyang Cui+2 more|arXiv (Cornell University)|Dec 25, 2023

Topic Modeling4 citations

TL;DR

This paper proposes Induce-then-Contrast Decoding (ICD), a lightweight decoding method that alleviates hallucinations in large language models by first inducing hallucinations in a weak model and then penalizing them during decoding via contrastive decoding. ICD significantly improves factuality, enabling Llama2-7B-Chat and Mistral-7B-Instruct to match GPT-3.5 and GPT-4 performance on TruthfulQA.

ABSTRACT

Despite their impressive capabilities, large language models (LLMs) have been observed to generate responses that include inaccurate or fabricated information, a phenomenon commonly known as ``hallucination''. In this work, we propose a simple extit{Induce-then-Contrast} Decoding (ICD) strategy to alleviate hallucinations. We first construct a factually weak LLM by inducing hallucinations from the original LLMs. Then, we penalize these induced hallucinations during decoding to enhance the factuality of the generated content. Concretely, we determine the final next-token predictions by amplifying the predictions from the original model and downplaying the induced untruthful predictions via contrastive decoding. Experimental results on both discrimination-based and generation-based hallucination evaluation benchmarks, such as TruthfulQA and extsc{FActScore}, demonstrate that our proposed ICD methods can effectively enhance the factuality 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.

Motivation & Objective

• To address the persistent issue of hallucinations in large language models, which generate factually incorrect or fabricated content despite strong overall performance.
• To develop a decoding-level method that enhances factuality without requiring changes to model weights or extensive fine-tuning.
• To explore whether inducing hallucinations in a weak model can serve as a proxy for identifying and suppressing non-factual generations in the original model.
• To evaluate the method across diverse model sizes and families, ensuring broad applicability and robustness.
• To provide a lightweight, efficient, and effective alternative to data-intensive or training-based hallucination mitigation techniques.

Proposed method

• Construct a factually weak LLM by inducing hallucinations in the original LLM through slight fine-tuning or zero-shot prompting.
• Use contrastive decoding to amplify predictions from the original model while downplaying predictions from the induced hallucination model.
• Apply a contrastive loss mechanism that adjusts token probabilities by comparing outputs from the original and hallucination-inducted models.
• Leverage parameter-efficient fine-tuning (LoRA) to train the hallucination-inducing weak model with minimal computational overhead.
• Perform decoding by combining the original model’s output with a contrastive penalty on induced hallucinatory tokens to suppress non-factual generations.
• Utilize a two-stage inference process: first generate from both models, then apply contrastive decoding to refine the final output.

Experimental results

Research questions

• RQ1 Can inducing hallucinations in a weak model serve as an effective proxy for identifying and suppressing non-factual generations in the original LLM?
• RQ2 Does contrastive decoding between a factual and a hallucination-inducted model significantly improve the factuality of LLM outputs?
• RQ3 How does ICD perform across different model architectures and sizes, including Llama2, Mistral, and their chat variants?
• RQ4 Can ICD outperform direct fine-tuning with factual data in terms of factual precision and response quality?
• RQ5 What is the impact of ICD on model coherence and response length, and does it preserve the original helpfulness learned via RLHF?

Key findings

• ICD improves the TruthfulQA MC1 score of Llama2-7B-Chat by +8 points and Mistral-7B-Instruct by +20 points, bringing their performance on par with GPT-3.5 and GPT-4, respectively.
• On the FActScore benchmark, Llama2-7B-Chat enhanced with ICD achieves a factual precision score of 66.3, outperforming the 70B version of the same model.
• ICD reduces hallucinations more effectively than direct fine-tuning with Wikipedia biographies, which introduced new hallucinations and degraded response quality.
• The method maintains high response quality and length, avoiding the shortening and loss of helpfulness often seen in direct fine-tuning approaches.
• ICD demonstrates consistent performance gains across multiple model families and sizes, including both 7B and 70B parameter variants.
• The approach increases inference latency by only 1.6x due to double forward passes, with negligible GPU memory overhead thanks to LoRA-based fine-tuning.