Consistency in Language Models: Current Landscape, Challenges, and Future Directions

Jekaterina Novikova, Carol Anderson, Borhane Blili-Hamelin, Domenic Rosati, Subhabrata Majumdar

TL;DR

Large Language Models (LLMs) frequently contradict themselves or flip their answers based on minor phrasing tweaks, posing significant risks for high-stakes deployments. This survey maps out the fragmented landscape of "consistency" research from 2019 to 2025, categorizing how it is defined, measured, and mitigated. The authors highlight critical gaps—such as a severe English-centric bias and a lack of standardized metrics—and call for a shift from treating symptoms to fixing the foundational architectures of LMs.

What Problem Does This Paper Address?

While humans naturally avoid self-contradiction and maintain meaning across different contexts, state-of-the-art LLMs notoriously struggle with consistency. An LLM might answer a question correctly but fail if the prompt is slightly paraphrased, or it might provide conflicting advice across different languages.

The core issue is that the AI research community lacks a unified definition, standardized metrics, or comprehensive benchmarks for consistency. This fragmentation makes it easy to overestimate model performance and underestimate the risks of deploying them in critical fields like medicine, law, or autonomous systems. This paper systematically reviews the current landscape to identify research gaps and provide a roadmap for future work.

Methods

As a literature review, the authors analyze peer-reviewed publications and influential preprints spanning 2019 to 2025. To organize the chaotic terminology in the field, they narrow the scope to behavioral consistency and propose a taxonomy dividing it into two main categories:

  • Logical/Formal Consistency: Adhering to strict rules of formal logic, such as negational consistency (opposite predictions for opposite meanings), symmetric consistency (invariance to input swaps), and transitive consistency (deductive reasoning chains).
  • Nonlogical/Informal Consistency: Covering semantic equivalence (maintaining meaning across different phrasings), moral consistency, norm consistency, and factual/informational consistency (avoiding hallucinations).

The authors also break down evaluation methodologies into input-based sampling (testing variations of prompts) and output-based sampling (generating multiple responses to the same prompt), alongside the metrics used to score them (e.g., BERTScore, ROUGE, Entailment, and Semantic Entropy).

Datasets

Because this is a survey paper, it does not introduce a new dataset. Instead, it reviews the dataset creation trends in the literature:

  • Size: Evaluation datasets vary wildly, ranging from a few hundred to tens of thousands of samples.
  • Approaches: Common methods include multiplying prompts in existing benchmarks using perturbation rules/templates (e.g., AdvGLUE) or enhancing datasets with human- or LLM-generated paraphrases.
  • Availability: While the majority of testing datasets are publicly shared, a notable portion of studies only describe their creation process without providing open access.

Key Insights and Ideas

  • Terminology Chaos: Researchers frequently invent narrow, task-specific definitions of consistency or omit definitions entirely, leading to studies that cover completely different aspects of model behavior and sometimes contradict one another.
  • The English Bias: An astounding 73% of popular datasets used to train LLMs are primarily or entirely English. Consequently, cross-lingual consistency is highly unstable; models regularly drop safety guardrails, express different political biases, or give conflicting medical advice when asked the exact same question in different languages.
  • The Creativity Trade-off: Perfect consistency isn't always desirable. Lower consistency can actually be beneficial for creative brainstorming, generating diverse outputs, or prompting users to critically verify AI-generated content.
  • Band-Aid Solutions: Current methods to improve consistency (like fine-tuning on synthetic datasets or using knowledge distillation) only treat the symptoms at decoding time rather than addressing the root structural flaws in how LMs represent knowledge.

Key Results

  • Model Trends: More than two-thirds of the reviewed studies evaluate transformer-based generative LMs (decoder-only or encoder-decoder architectures like GPT, OPT, BART, and T5). About a quarter focus on encoder-only models (BERT-style).
  • Task Distribution: A slim majority of studies evaluate well-established NLP tasks like Q&A, summarization, Natural Language Inference (NLI), and multi-step reasoning. Approximately one-third rely on custom tasks, such as generating sentence continuations from Wikipedia.

Limitations and Future Work

The authors identify several glaring gaps in the current literature:

  • Decoding-Time Obsession: Current research focuses heavily on generation at decoding time, largely ignoring encoder-only models and how inconsistent inputs shape downstream tasks.
  • Vulnerability to Subtle Attacks: There is a severe lack of research on how inconspicuous prompt manipulations or adversarial attacks degrade consistency.
  • Lack of Transparency: The prevalence of proprietary models (like GPT-4) and unpublished code/datasets makes reproducing and verifying consistency claims nearly impossible.

Future Directions: The paper concludes with a call to action for the research community to develop standardized taxonomies, create holistic multilingual and cross-lingual benchmarks, establish robust evaluation protocols combining automatic metrics with human judgment, and investigate consistency-oriented pre-training architectures.

Summary updated

Abstract

The hallmark of effective language use lies in consistency: expressing similar meanings in similar contexts and avoiding contradictions. While human communication naturally demonstrates this principle, state-of-the-art language models (LMs) struggle to maintain reliable consistency across task- and domain-specific applications. Here we examine the landscape of consistency research in LMs, analyze current approaches to measure aspects of consistency, and identify critical research gaps. Our findings point to an urgent need for quality benchmarks to measure and interdisciplinary approaches to ensure consistency while preserving utility.
BibTeX
@misc{novikova2025consistency,
  author = {Jekaterina Novikova and Carol Anderson and Borhane Blili-Hamelin and Domenic Rosati and Subhabrata Majumdar},
  title = {{Consistency in Language Models: Current Landscape, Challenges, and Future Directions}},
  year = {2025},
  eprint = {2505.00268},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.00268},
}
Figure