Search papers, labs, and topics across Lattice.
The paper introduces BLooP (Bigram Lookahead Promotion), a training-free decoding intervention for LLMs that encourages the generation of tokens forming bigrams present in the source document to improve abstractive summarization. BLooP uses a hash table lookup at each decoding step to promote relevant bigrams, requiring no training or model modification. Experiments on CNN/DM, CCSum, Multi-News, and SciTLDR datasets using Llama-3.1-8B-Instruct, Mistral-Nemo-Instruct-2407, and Gemma-2-9b-it show improvements in ROUGE and BARTScore, with human evaluations confirming enhanced faithfulness without sacrificing readability.
LLMs can generate more faithful abstractive summaries without any training, simply by biasing the decoder towards source-document bigrams.
Abstractive summarization requires models to generate summaries that convey information in the source document. While large language models can generate summaries without fine-tuning, they often miss key details and include extraneous information. We propose BLooP (Bigram Lookahead Promotion), a simple training-free decoding intervention that encourages large language models (LLMs) to generate tokens that form bigrams from the source document. BLooP operates through a hash table lookup at each decoding step, requiring no training, fine-tuning, or model modification. We demonstrate improvements in ROUGE and BARTScore for Llama-3.1-8B-Instruct, Mistral-Nemo-Instruct-2407, and Gemma-2-9b-it on CNN/DM, CCSum, Multi-News, and SciTLDR. Human evaluation shows that BLooP significantly improves faithfulness without reducing readability. We make the code available at https://github.com/varuniyer/BLooP
