Source-linked AI summary

Trusting Your Evidence: Hallucinate Less with Context-aware Decoding

Weijia Shi, Xiaochuang Han, Mike Lewis, Yulia Tsvetkov, Luke Zettlemoyer, Scott Wen-tau Yih

arXiv:2305.14739v1cs.CL

TL;DR

Language models may underweight supplied context, causing unfaithful or hallucinated outputs, especially when context conflicts with pretrained knowledge. The paper introduces context-aware decoding, which contrasts context-conditioned and context-free probabilities without additional training. CAD improves summarization faithfulness across model families and substantially helps resolve knowledge conflicts, although performance can decline when evaluation data appeared in instruction tuning.

  • Problem

    Language models can insufficiently attend to supplied context, producing hallucinations and failing to override outdated prior knowledge when the context contradicts it.

  • Method

    Context-aware decoding contrasts output probabilities with and without context to downweight prior knowledge and promote context-supported generation without additional training.

  • Results

    CAD improves faithfulness across diverse language models and summarization tasks, including a 14.3% factuality gain for LLaMA and a 2.9x improvement on a knowledge-conflict QA dataset.

  • Takeaways & Limitations

    CAD provides more reliable and factual outputs across language models of varied sizes and can override prior knowledge when trusted contextual information conflicts with it.

  • Takeaways & Limitations

    A slight decline occurred on one NQ dataset because it had been included in FLAN-T5’s instruction-finetuning data.

Abstract

from arXiv · show

Language models (LMs) often struggle to pay enough attention to the input context, and generate texts that are unfaithful or contain hallucinations. To mitigate this issue, we present context-aware decoding (CAD), which follows a contrastive output distribution that amplifies the difference between the output probabilities when a model is used with and without context. Our experiments show that CAD, without additional training, significantly improves the faithfulness of different LM families, including OPT, GPT, LLaMA and FLAN-T5 for summarization tasks (e.g., 14.3% gain for LLaMA in factuality metrics). Furthermore, CAD is particularly effective in overriding a model's prior knowledge when it contradicts the provided context, leading to substantial improvements in tasks where resolving the knowledge conflict is essential.

1 Introduction

Language models must balance learned prior knowledge with information supplied in context, but may underweight new or contradictory evidence. Context-aware decoding addresses this by amplifying context-sensitive probabilities, improving faithfulness across models and especially in knowledge-conflict tasks.

  • Language models combine pretrained prior knowledge with context knowledge supplied in the input, but their balance during generation remains unresolved.
  • Insufficient attention to newly introduced context can produce summaries containing facts absent from the input document.
  • Contradictory context may fail to override outdated model knowledge, as LLaMA predicts two World Cups despite a document stating Argentina won three.
  • Context-aware decoding amplifies probability differences with versus without context, downweighting prior knowledge without additional training.
  • CAD improves faithfulness across OPT, GPT-Neo, LLaMA, and FLAN summarization models, including a 14.3% gain in summary factuality metrics for LLaMA-30B.
  • CAD is especially effective on knowledge-conflict tasks, producing a 2.9x improvement for LLaMA-30B and larger gains as model size increases.

2 Method

CAD contrasts context-conditioned and context-free model outputs to reduce reliance on prior knowledge when context is unfamiliar or conflicting. It normalizes a PMI-weighted product-of-experts distribution and samples from the adjusted probabilities.

  • CAD addresses cases where out-of-distribution or conflicting context is underweighted and the model instead relies on encoded prior knowledge.
  • The method factors out prior knowledge contrastively by comparing output probabilities conditioned on context with those conditioned only on the query and prefix.
  • CAD forms a product-of-experts distribution combining the original output probability with context-generation PMI weighted by α, favoring context-supported outputs.
  • The unnormalized expression is normalized across possible next tokens before sampling from the adjusted distribution.
  • Larger α increases adjustment strength, while α = 0 recovers regular decoding; standard sampling strategies can then operate on the adjusted distribution.
  • CAD is a contrastive ensemble of logits from context-conditioned and context-free model evaluations.

3 Experimental Setup

The evaluation tests CAD with prompting on context-dependent summarization and knowledge-conflict tasks across pretrained and instruction-finetuned language models. Experiments use standard task-specific datasets, metrics, baselines, and fixed adjustment settings.

  • The experiments evaluate whether language models read and reason over contexts while producing outputs faithful to those contexts.
  • Summarization: Summarization uses CNN-DM and XSUM, with ROUGE-L for quality and BERT-Precision plus FactKB for factual consistency.
  • Knowledge Conflicts: Knowledge-conflict evaluation uses MemoTrap, which tests memorization traps by changing proverb endings, and NQ-Swap, a question-answering dataset.
  • Task Inputs: Across datasets, CAD is intended to upweight source documents in XSUM and NQ-Swap and instructions in MemoTrap.
  • Models: Models include OPT, GPT-Neo, LLaMA, and instruction-finetuned FLAN-T5 variants spanning approximately 2.7B to 30B parameters.
  • Settings: The adjustment uses α = 0.5 for summarization and α = 1 for knowledge-conflict tasks, while CAD retains the baseline sampling strategies.

4 Results

CAD improves summary quality and factuality across diverse language models and substantially helps resolve knowledge conflicts. Its gains remain consistent across model sizes in CNN-DM and increase with size on conflict-focused tasks.

  • Main Results: CAD outperforms regular decoding across eight models on CNN-DM and XSUM for both summary quality and factuality.The evaluated metrics are ROUGE-L, factKB, and BERT-P.
  • Main Results: 21% increase in ROUGE-L, 14.3% increase in factKB, and 7.8% increase in BERT-P result from applying CAD to LLaMA-30B on CNN-DM.
  • Main Results: CAD significantly outperforms regular decoding on knowledge-conflict datasets, except for a minor FLAN-T5 decrease on non-conflict NQ.The reported exception is attributed to FLAN-T5 having seen NQ during instruction-finetuning.
  • Main Results: 54.4% improvement on MemoTrap and 128% on NQ-SWAP occur for GPT-Neo 20B with CAD.These gains are reported for knowledge-conflict datasets.
  • Analysis: CAD’s gain stays consistent across OPT sizes on CNN-DM but increases with model size on Memotrap and NQSWAP.The pattern suggests larger models rely more on prior knowledge in knowledge-conflict tasks.
  • Analysis: λ = 0.5 consistently provides robust improvements across three datasets, while larger adjustment values further improve knowledge-conflict tasks.The adjustment level α controls how strongly CAD changes the original next-token distribution.

5 Related Work

Prior work studies hallucination reduction, knowledge conflicts, and contrastive decoding, but context addition alone may not override parametric knowledge. CAD addresses this gap with a decoding strategy applicable across language models.

  • Summarization Factuality: Summarization research has pursued factual consistency through source-document fact triples, factuality objectives, and post-editing correction.
  • Knowledge Conflicts: Knowledge-conflict settings evaluate whether models follow updated contextual information instead of outdated parametric knowledge.This setting is especially relevant to retrieval-augmented language models using external documents.
  • Knowledge Conflicts: Existing prompting-based approaches are limited to large instruction-finetuned models, whereas CAD is presented as a decoding strategy applicable to any LM.
  • Contrastive Decoding Methods: Contrastive decoding methods compare differently conditioned model outputs to improve generation properties such as coherence or diversity.CAD applies the same contrastive intuition to knowledge-conflict scenarios.

6 Conclusion

The paper presents CAD as a simple inference-time method for improving models’ attention to supplied context over learned prior knowledge. Across summarization and knowledge-conflict tasks, CAD produces more reliable and factual outputs across language models of various sizes.

  • CAD downweights output probabilities associated with prior knowledge to promote attention to contextual information.
  • CAD is evaluated on summarization and knowledge conflicts tasks, both requiring strong attention to context.
  • CAD provides more reliable and factual outputs across different language models of various sizes.
Loading 2305.14739v1…