An ideal text summary strikes a delicate balance: it should be abstractive—synthesizing, restructuring, and paraphrasing information into a fluent, cohesive narrative—and it must be faithful—strictly adhering to the facts presented in the source text without introducing hallucinations.
However, modern sequence-to-sequence models face a stubborn Pareto frontier between these two goals. When models are given the creative freedom to paraphrase, they frequently distort semantic relationships and invent false details. Conversely, when they are heavily penalized for factual errors, they default to an "extractive bias," copy-pasting verbatim chunks of the source text. Understanding why this trade-off occurs and how to bypass it is a central challenge in contemporary natural language processing.
Why Abstractiveness Breeds Hallucination
To understand why abstractive models struggle with truthfulness, we must look at how they generate text. Unlike extractive models that merely rank and select existing sentences, abstractive models generate text token-by-token. As a model departs from the syntactic structure of the source text to create a novel paraphrase, the probability distribution over its vocabulary expands. This freedom increases the risk of misattributing actions to the wrong entities or blending unrelated facts.
As Klaus-Michael Lux et al. demonstrate in Truth or Error? Towards systematic analysis of factual errors in abstractive summaries, there is a direct correlation between the degree of abstraction a model exhibits and its likelihood of generating factual errors. When models perform extensive rewriting, they face a significantly higher risk of introducing misleading implications compared to conservative, extractive baselines.
This tension is further illuminated by the temporal behavior of these models during training. In Training Dynamics for Text Summarization Models, Tanya Goyal et al. reveal that during the fine-tuning process, models learn the mechanics of copying input text very early on. The tendency to generate novel, abstractive phrasing—and consequently, the tendency to hallucinate—develops much later in the training cycle. Hallucination is not necessarily a sign of a failing model, but rather a byproduct of a model attempting to generalize and synthesize beyond its strict factual boundaries.
The Extractive Bias as a Path of Least Resistance
Faced with the risk of hallucination, researchers have attempted to enforce factuality constraints during pre-training, fine-tuning, and decoding. However, these constraints often trigger an unintended side effect: they push the model toward purely extractive behavior.
Faisal Ladhak et al. formalize this phenomenon in Faithful or Extractive? On Mitigating the Faithfulness-Abstractiveness Trade-off in Abstractive Summarization. They show that standard abstractive models often improve their faithfulness scores simply by increasing the proportion of text copied directly from the source. When a model is penalized for factual inconsistency, the safest mathematical optimization is to collapse its generative choices and rely on verbatim strings.
This creates a dependency where higher accuracy is achieved at the expense of synthesis. For instance, David Wan et al. found in FactPEGASUS: Factuality-Aware Pre-training and Fine-tuning for Abstractive Summarization that while integrating factuality-aware objectives significantly reduces hallucinations, it consistently makes the models more extractive. The model sacrifices conciseness and structural flow, delivering summaries that feel choppy, redundant, or overly long because it lacks the confidence to rephrase the source material safely.
Algorithmic Interventions to Break the Trade-off
To break this trade-off, the field has shifted away from monolithic generation toward architectures that decouple content selection from linguistic realization.
Decoupled Decoders and Novelty Rewards
One approach is to explicitly penalize copying during training while maintaining a separate mechanism for language fluency. In Improving Abstraction in Text Summarization, Wojciech Kryscinski introduces a factorized decoder that separates content extraction from surface realization, paired with a reinforcement learning novelty reward. By explicitly penalizing the model for reproducing long n-grams from the source text, the framework forces the generator to find novel paraphrases without losing the core informational content.
Hybrid Extract-and-Decontextualize Pipelines
Another prominent strategy relies on a two-stage pipeline: first extracting factual evidence clusters, and then using a generative model to smoothly merge them. Abhilash Potluri et al. leverage this in Concise Answers to Complex Questions: Summarization of Long-form Answers, demonstrating that extracting key sentences and applying minimal, targeted rewriting preserves the strict grounding of extractive methods while achieving the linguistic polish of abstractive ones.
Similarly, Sebastian Gehrmann et al. propose a "bottom-up" approach in Bottom-Up Abstractive Summarization. Their system uses a data-efficient content selector to generate a hard mask over the source text. The abstractive generator is then restricted to attend only to these pre-selected words. This constraint bounds the search space, preventing the model from wandering into hallucinated territory while still allowing it to flexibly arrange the allowed tokens into a fluent narrative.