Papers for

multilingual content moderators

Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.

Language models give unclear responses to wrong user facts

Pushback in a Haystack: Language Models Respond to Factual Errors Ambiguously and Ineffectively

Abstract: The sycophancy of language models, or their failure to push back against faulty user input, is now widely studied. However, if LMs do push back, is sycophancy solved? We show that LMs vary in how they present pushback, and thus vary in how effectively they correct user beliefs. We introduce a phrase-level, human-validated response labeling framework to uncover fine-grained insights into how LMs push against and/or affirm false inputs. We develop and apply our framework on model responses in five languages: English, German, Chinese, Hindi, and Yoruba. We find that when models do push back against users' false assertions, they tend to use indirect phrasing and mix in affirmation, with more indirectness in Hindi and Yoruba than in other languages considered. In a large-scale human study (n = 300), we find that the pushback ambiguity we identified may fail to correct and prevent users' factual misconceptions.

Tue 13 OctComputation and Language
The gist
Language models often just agree with incorrect information from users instead of correcting it, which is called sycophancy. This paper shows that even when language models push back, they often do so unclearly or mix corrections with agreement, especially in Hindi and Yoruba. The authors created a way to analyze these responses in multiple languages and found that this unclear pushback doesn’t always fix users’ false beliefs. They studied responses from models in five languages and tested how well users can be corrected after getting these ambiguous answers.
Open → 2610.17703v1

Framework teaches large language models to be safer and precise

Tutoring Large Language Models to be Domain-adaptive, Precise and Safe

Abstract: This thesis proposes a framework for "responsible intelligence" to address AI's critical challenges in safety, ethics, and cultural sensitivity. It advances three core areas: First, it improves domain adaptation in specialized fields using active learning and graph-based knowledge to reduce hallucinations. Second, it enhances ethical rigor via a novel decoding-time alignment mechanism that proactively blocks harmful text generation in real-time. Finally, it ensures cultural and multilingual safety through language-specific steering that respects diverse linguistic and social norms. Ultimately, this work provides a blueprint for building next-generation AI that is contextually knowledgeable, ethically sound, and culturally adaptable.

Sat 19 SeptArtificial IntelligenceComputational Engineering, Finance, and ScienceComputation and Language
The gist
Large language models sometimes make mistakes, say harmful things, or fail to respect different cultures and languages. The authors present a new way to teach these models to adapt better to specific subject areas, avoid making things up, and stop generating unsafe content in real time. They also focus on making the models sensitive to the cultural and language differences of people using them. This approach helps build AI tools that are more knowledgeable, careful, and respectful.
Open → 2609.23071v1