# Improved Natural Language Generation via Loss Truncation

arXiv: https://arxiv.org/abs/2004.14589 · HTML: https://mlsanity.com/paper/2004.14589/

## Key facts

- Cited by 75 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 30 Apr 2020.

## Built on by (most cited first, 30 of 75)

- [Survey of Hallucination in Natural Language Generation](https://mlsanity.com/paper/2202.03629/) (2022)
- [Benchmarking Large Language Models for News Summarization](https://mlsanity.com/paper/2301.13848/) (2023)
- [Trusting Your Evidence: Hallucinate Less with Context-aware Decoding](https://mlsanity.com/paper/2305.14739/) (2023)
- [The Curious Case of Hallucinations in Neural Machine Translation](https://mlsanity.com/paper/2104.06683/) (2021)
- [On the Origin of Hallucinations in Conversational Models: Is it the Datasets or the Models?](https://mlsanity.com/paper/2204.07931/) (2022)
- [Annotating and Modeling Fine-grained Factuality in Summarization](https://mlsanity.com/paper/2104.04302/) (2021)
- [Language Models with Conformal Factuality Guarantees](https://mlsanity.com/paper/2402.10978/) (2024)
- [Hallucinated but Factual! Inspecting the Factuality of Hallucinations in Abstractive Summarization](https://mlsanity.com/paper/2109.09784/) (2021)
- [Truncation Sampling as Language Model Desmoothing](https://mlsanity.com/paper/2210.15191/) (2022)
- [FaithDial: A Faithful Benchmark for Information-Seeking Dialogue](https://mlsanity.com/paper/2204.10757/) (2022)
- [Evaluating Human-Language Model Interaction](https://mlsanity.com/paper/2212.09746/) (2022)
- [Cognitive Mirage: A Review of Hallucinations in Large Language Models](https://mlsanity.com/paper/2309.06794/) (2023)
- [Language Model Evaluation Beyond Perplexity](https://mlsanity.com/paper/2106.00085/) (2021)
- [Evaluating Attribution in Dialogue Systems: The BEGIN Benchmark](https://mlsanity.com/paper/2105.00071/) (2021)
- [Improving Faithfulness in Abstractive Summarization with Contrast Candidate Generation and Selection](https://mlsanity.com/paper/2104.09061/) (2021)
- [Mitigating Large Language Model Hallucinations via Autonomous Knowledge Graph-based Retrofitting](https://mlsanity.com/paper/2311.13314/) (2023)
- [Mutual Information Alleviates Hallucinations in Abstractive Summarization](https://mlsanity.com/paper/2210.13210/) (2022)
- [Text Generation by Learning from Demonstrations](https://mlsanity.com/paper/2009.07839/) (2020)
- [Order-Agnostic Cross Entropy for Non-Autoregressive Machine Translation](https://mlsanity.com/paper/2106.05093/) (2021)
- [Faithful or Extractive? On Mitigating the Faithfulness-Abstractiveness Trade-off in Abstractive Summarization](https://mlsanity.com/paper/2108.13684/) (2021)
- [FactPEGASUS: Factuality-Aware Pre-training and Fine-tuning for Abstractive Summarization](https://mlsanity.com/paper/2205.07830/) (2022)
- [Large Language Models Hallucination: A Comprehensive Survey](https://mlsanity.com/paper/2510.06265/) (2025)
- [Detecting and Mitigating Hallucinations in Multilingual Summarisation](https://mlsanity.com/paper/2305.13632/) (2023)
- [mFACE: Multilingual Summarization with Factual Consistency Evaluation](https://mlsanity.com/paper/2212.10622/) (2022)
- [Contrastive Learning Reduces Hallucination in Conversations](https://mlsanity.com/paper/2212.10400/) (2022)
- [AI Hallucinations: A Misnomer Worth Clarifying](https://mlsanity.com/paper/2401.06796/) (2024)
- [Inference-Time Policy Adapters (IPA): Tailoring Extreme-Scale LMs without Fine-tuning](https://mlsanity.com/paper/2305.15065/) (2023)
- [Relation Classification with Entity Type Restriction](https://mlsanity.com/paper/2105.08393/) (2021)
- [Elastic Weight Removal for Faithful and Abstractive Dialogue Generation](https://mlsanity.com/paper/2303.17574/) (2023)
- [Learning to Revise References for Faithful Summarization](https://mlsanity.com/paper/2204.10290/) (2022)

## Abstract (arXiv)

Neural language models are usually trained to match the distributional properties of a large-scale corpus by minimizing the log loss. While straightforward to optimize, this approach forces the model to reproduce all variations in the dataset, including noisy and invalid references (e.g., misannotation and hallucinated facts). Worse, the commonly used log loss is overly sensitive to such phenomena and even a small fraction of noisy data can degrade performance. In this work, we show that the distinguishability of the models and reference serves as a principled and robust alternative for handling invalid references. To optimize distinguishability, we propose loss truncation, which adaptively removes high loss examples during training. We show this is as easy to optimize as log loss and tightly bounds distinguishability under noise. Empirically, we demonstrate that loss truncation outperforms existing baselines on distinguishability on a summarization task, and show that samples generated by the loss truncation model have factual accuracy ratings that exceed those of baselines and match human references.