# TLDR: Token Loss Dynamic Reweighting for Reducing Repetitive Utterance Generation

arXiv: https://arxiv.org/abs/2003.11963 · HTML: https://mlsanity.com/paper/2003.11963/

## Key facts

- Cited by 6 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 26 Mar 2020.

## Built on by (most cited first, 6 of 6)

- [Straight to the Gradient: Learning to Use Novel Tokens for Neural Text Generation](https://mlsanity.com/paper/2106.07207/) (2021)
- [A Simple Contrastive Learning Objective for Alleviating Neural Text Degeneration](https://mlsanity.com/paper/2205.02517/) (2022)
- [Implicit Unlikelihood Training: Improving Neural Text Generation with Reinforcement Learning](https://mlsanity.com/paper/2101.04229/) (2021)
- [Learning to Insert [PAUSE] Tokens for Better Reasoning](https://mlsanity.com/paper/2506.03616/) (2025)
- [Influence Scores at Scale for Efficient Language Data Sampling](https://mlsanity.com/paper/2311.16298/) (2023)
- [Multi-level Adaptive Contrastive Learning for Knowledge Internalization in Dialogue Generation](https://mlsanity.com/paper/2310.08943/) (2023)

## Abstract (arXiv)

Natural Language Generation (NLG) models are prone to generating repetitive utterances. In this work, we study the repetition problem for encoder-decoder models, using both recurrent neural network (RNN) and transformer architectures. To this end, we consider the chit-chat task, where the problem is more prominent than in other tasks that need encoder-decoder architectures. We first study the influence of model architectures. By using pre-attention and highway connections for RNNs, we manage to achieve lower repetition rates. However, this method does not generalize to other models such as transformers. We hypothesize that the deeper reason is that in the training corpora, there are hard tokens that are more difficult for a generative model to learn than others and, once learning has finished, hard tokens are still under-learned, so that repetitive generations are more likely to happen. Based on this hypothesis, we propose token loss dynamic reweighting (TLDR) that applies differentiable weights to individual token losses. By using higher weights for hard tokens and lower weights for easy tokens, NLG models are able to learn individual tokens at different paces. Experiments on chit-chat benchmark datasets show that TLDR is more effective in repetition reduction for both RNN and transformer architectures than baselines using different weighting functions.