Learning to Break the Loop: Analyzing and Mitigating Repetitions for Neural Text Generation

1 Introduction

• Language models prefer consecutive sentence-level repetition more strongly than word- or phrase-level repetition, unlike human language.Greedy decoding on Wikitext-103 gets stuck in sentence-level repetition, whereas human corpora contain substantially fewer such repetitions.Source: §1 · p. 1, Figure 1 · p. 2
• Repetition probability rises with prior repetitions, and sentences with higher initial probabilities usually show stronger self-reinforcement.The probability can increase almost monotonically and eventually stabilize near a ceiling.Source: §1 · p. 2, Figure 2 · p. 3
• Self-reinforcement creates sentence-level loops because each generated repetition further increases the probability of continuing that sentence.The mechanism links high-likelihood generated sentences to progressively stronger copying behavior.Source: §1 · p. 3
• DITTO trains on pseudo-repetitive data to make repetition probabilities decay as sentence-level repetitions accumulate.The method targets both the tendency to repeat previous sentences and their self-reinforcement effect.Source: §1 · p. 3
• DITTO reduces repetitions without sacrificing perplexity and improves generation quality across Wikitext-103 generation and CNN/DailyMail summarization.The reported gains include lower perplexity and higher MAUVE than strong competitors in open-ended generation, with best overall performance on summarization.Source: §1 · p. 3

2.1 Experiment Design

• The experiment repeats each sentence N = 100 times and computes average token probability, increased-probability rate, and winner rate across three corpora.The corpora are Drandom, Dbook, and Dwiki, each containing 1,000 sentences.Source: Figure 3 · p. 4, §2.1 · p. 4
• A winner is a token whose probability exceeds its initial value and is maximal under the current context, making winner rate relevant to greedy decoding.The setup defines sentence-level context repetition by matching token positions and preceding within-sentence context across repetitions.Source: §2.1 · p. 4

2.2 Results and Analyses

• IP1 exceeds 90% across corpus types, showing that one sentence-level context repetition usually raises the next repetition’s probability.The model can copy the next token after recognizing the same sentence-level context, even before the token itself has repeated at the current step.Source: §2.2 · p. 4
• TP, IP, and WR increase monotonically with repetition count before converging near ceiling values, demonstrating self-reinforcement.More historical sentence repetitions make the model more likely to continue generating that sentence.Source: §2.2 · p. 5
• Sentences with high initial probability grow faster in TP and WR and reach high ceilings after fewer repetitions.Such sentences are therefore more likely to be repeatedly generated by maximization-based decoding.Source: §2.2 · p. 5
• The analysis attributes repetition loops to high-likelihood prior sentences, context-based copying, and subsequent self-reinforcement during maximization-based decoding.Stochastic sampling may reduce the risk by lowering prior-sentence likelihoods and avoiding maximum-probability token selection.Source: §2.2 · p. 5

Sentence-level Repetition Penalization

• The method explicitly teaches the model to avoid redundant sentences by applying a sentence-level repetition penalization loss to tokens in repeated sentences.This training targets the tendency to repeat previous sentences and the resulting self-reinforcement effect.Source: §3 · p. 5

Pseudo Repetitive Data Construction

• DITTO constructs pseudo-repetitive data by repeating a randomly sampled training sentence until the model’s maximum input length is reached.The method may use sequences such as 1,536 tokens for open-ended generation.Source: §3 · p. 5

Sentence-level Repetition Penalization

• When λ < 1, DITTO makes token probabilities decay across repetitions, whereas λ = 1 prevents self-reinforcement without requiring probability decay.The penalization factor controls how strongly later repetitions are suppressed.Source: §3 · p. 5

Training Procedure

• DITTO fine-tunes a standard maximum-likelihood baseline by mixing repetition-penalization updates with normal MLE updates.Appending the repetitive sentence’s previous context as a prefix produced better and more stable performance than omitting it.Source: §3 · p. 5

Discussion

• Unlike hard repetition constraints, DITTO treats necessary token- and phrase-level repetitions as useful positives while targeting over-repetition as negative behavior.This objective is intended to reduce over-penalization of repetitions that are appropriate in context.Source: §3 · p. 6

4.2 Results of Open-ended Generation

• DITTO reduces phrase- and sentence-level repetition while achieving the best MAUVE score of 0.77 on Wikitext-103 open-ended generation.Compared with standard MLE, DITTO also achieves lower perplexity and higher accuracy rather than sacrificing language-modeling quality.Source: §4.2 · pp. 6–7, Table 1 · p. 7
• DITTO achieves the highest MAUVE of 0.96 with top-k or nucleus sampling and produces repetition metrics closest to human text.The method reduces repetitions relative to the MLE baseline under both stochastic decoding strategies.Source: §4.2 · p. 7, Table 2 · p. 7
• DITTO consistently and significantly outperforms competing methods in human evaluations across different decoding strategies.Crowdworkers compare grammaticality, relevance, coherence, and informativeness on generations from shared test prefixes.Source: §4.2 · p. 7, Table 3 · p. 7
• DITTO overcomes self-reinforcement: TP, IP, and WR drop rapidly as repetitions grow, unlike UL and SG where they increase almost monotonically.These analyses directly test whether training-based methods address the repetition mechanism identified earlier.Source: §4.2 · pp. 7–8, Figure 4 · p. 8

4.3 Results of Directed Generation

• On CNN/DailyMail summarization, DITTO consistently outperforms UL, SG, and other competitive baselines on ROUGE scores while remaining compatible with n-gram blocking.The directed-generation setup uses BART-large and standard beam search with tri-gram blocking.Source: §4.3 · p. 8, Table 4 · p. 8

4.4 Analyses

• DITTO remains effective across decoding lengths, consistently and significantly outperforming MLE under the tested length constraints.The decoding-length analysis reports robustness across the evaluated settings.Source: §4.4 · p. 8, Figure 5 · p. 9
• The recommended mix ratio is γ = 0.5, while the preferred penalization factor differs by task: λ = 0.5 for Wikitext-103 and λ = 0.9 for summarization.The authors conjecture that repetition is more severe in the more unconstrained open-ended generation task.Source: §4.4 · p. 9, Figure 5 · p. 9

• Neural language models generate consecutive word-, phrase-, and sentence-level repetitions, with sentence-level repetition producing redundant, non-informative text unlike human language.Source: §5 · p. 9
• Unlike self-recovery results obtained with ancestral sampling, this paper studies why maximization-based decoding increases repetition when tokens share the same sentence-level context.Source: §5 · pp. 9–10
• Earlier work proposes training- and decoding-based mitigation methods, including unlikelihood training, straight to gradient, n-gram blocking, top-k, and nucleus sampling.Source: §5 · p. 9
• Prior analyses identify positive feedback loops and long-range contextual relationships, but provide limited detailed numeric and statistical evidence across sentences.Source: §5 · p. 10

• The analysis identifies previous-sentence preference and self-reinforcement as the mechanisms underlying consecutive sentence-level repetition.Source: §6 · p. 10
• DITTO constructs pseudo-repetitive data and trains models to gradually decay repetition probability as sentence repetitions increase.Source: §6 · p. 10
• Experiments on Wikitext-103 open-ended generation and CNN/DailyMail summarization demonstrate the method’s superiority.Source: §6 · p. 10
• The work does not analyze deeper causes from model embeddings, neural architectures, or intrinsic language characteristics, leaving those questions for future research.Source: §6 · p. 10

• The paper reports that it describes its contributions and scope and discusses limitations in Section 6.Source: Checklist · p. 14
• The paper marks potential negative societal impacts as not applicable and states that ethics review guidelines were followed.Source: Checklist · p. 14
• The paper states that code, data, instructions, training details, error bars, and compute information are provided or documented.Source: Checklist · p. 14
• The paper cites existing assets and marks licensing and inclusion of new assets as not applicable.Source: Checklist · p. 14
• The checklist marks consent, personally identifiable or offensive content, participant instructions, participant risks, and compensation as not applicable or addressed as specified.Source: Checklist · p. 14

A Consecutive Repetitions and Statistics of Beam Search Results

• The paper defines repetition statistics separately for adjacent words, repeated k-word phrases, and consecutive sentences, averaging each measure over the corpus.Source: A Consecutive Repetitions and… · p. 15
• Beam search and greedy decoding produce substantially more sentence-level repetition than human language, extending the phenomenon beyond greedy search.Source: A Consecutive Repetitions and… · p. 15, Figure 6 · p. 15

B Self-reinforcement in Model Generated Texts

• The probability of the token ‘located’ increases almost monotonically with the number of historical repetitions, demonstrating self-reinforcement in autoregressive sentence loops.Source: B Self-reinforcement in Model… · p. 15, Figure 7 · p. 16
• Manual sentence repetition isolates the relationship between repetitive-token probabilities and repeated context, avoiding irregular generated sequences and testing whether the effect is intrinsic to language modeling.Source: B Self-reinforcement in Model… · pp. 15–16

• DITTO minimizes its loss when the current token probability equals λ times the previous repetition's probability, while stopping gradients through the reference value.Source: C Loss Function of DITTO · p. 16
• λ < 1 makes the probability of a token in the current repetition lower than in the previous repetition, countering self-reinforcement.Source: C Loss Function of DITTO · p. 16
• MSE is an alternative loss, but it achieves inferior performance in practice on Wikitext-103.Source: C Loss Function of DITTO · p. 16, Table 5 · p. 17
• LDITTO-margin performs worse because it can reduce repetition probability nearly to zero after one repetition, regardless of the previous probability.Source: C Loss Function of DITTO · p. 16, Figure 8 · p. 17
• Over-penalizing repetitions may forbid necessary repetitions, hurt language modeling ability, and lead to inferior performance.Source: C Loss Function of DITTO · p. 16

D.1 Self-reinforcement at sentences with different initial probabilities

• Higher initial-probability sentences reach higher TP and WR as repetitions increase, indicating stronger self-reinforcement and greater repeatability under greedy decoding.Source: §D.1 · p. 17
• Figure 10 averages results across Dwiki sentences to compare training-based methods when repetitive inputs are limited to 150 tokens.Source: Figure 10 · p. 18

• Repeating two consecutive sentences as a group achieves results similar to repeating one sentence as pseudo data.Source: §D.2 · p. 18
• DITTO-short reduces TP, IP, and WR for short decoding but these metrics rise during long decoding, unlike full-length DITTO.Source: §D.2 · p. 18
• With equal 150-token penalization length, DITTO is more effective than UL-token+seq and avoids its autoregressive generation requirement.Source: §D.2 · p. 18

E Human Evaluation Details

• Three fluent English speakers compare continuations from 100 Wikitext-103 prefixes on grammar, relevance, coherence, and informativeness.Source: E Human Evaluation Details · p. 18, Figure 11 · p. 19
• The evaluation win rate is DITTO's number of wins divided by the total number of pairwise comparisons.Source: E Human Evaluation Details · p. 18

F Examples

• Examples are provided for different training methods under greedy and nucleus sampling decoding.Source: F Examples · p. 19, Table 7 · p. 20, Table 8 · p. 21