Oct 2026· Information· Vol 17, pp. 981· 0 citations· 39 references
TL;DR
A run-level translation-quality advantage and task-specific human-outcome associations under the evaluated conditions support a run-level translation-quality advantage and task-specific human-outcome associations under the evaluated conditions.
Abstract
Professional translators may resist highly directive AI assistance when it threatens perceived autonomy, authorship, and control. This study introduces the Recursive Control and Alignment Protocol (ReCAP), a reactance-aware framework for interactive machine translation. ReCAP combines a reactance-potential model with a recursive loop of Intent Inference, Reactance-Guided Generation, and Critical Feedback Integration. Its Adversarial Reactance Minimization objective balances translation quality against estimated autonomy costs, and the task-specific Critical Automation Capability Index (CACI) provides an exploratory behavioral summary of critical engagement. In an evaluation with 156 professional translators across four English-based bidirectional language-pair settings, ReCAP achieved a mean COMET × 100 score of 36.8, 2.8 points above QE-guided MT (34.0), the strongest baseline, across five seed-matched runs per method–language-pair configuration. Mixed-effects estimates showed lower reported reactance for ReCAP than QE-guided MT (β = −0.55, 95% CI [−0.71, −0.39]) and higher exploratory CACI scores than CHORUS (β = 0.64, 95% CI [0.44, 0.84]). Post hoc CACI checks remained task-specific and exploratory. These findings support a run-level translation-quality advantage and task-specific human-outcome associations under the evaluated conditions.
Generative workflows improve structural reliability and reduce judged defects on short passages, while their large recovery advantage on long documents disappears after chunking, suggesting that reference recovery alone understates alignment quality.
Large Language Models (LLMs) can explain sensitive topics across domains, but harmful goals may be concealed behind educational, fictional, historical, or research-oriented framing. We study Human Indirect Prompting (HIP), a conversation-level behavior in which a user steers an LLM toward restricted information without...
Unais Sait, Saravanakumar Duraisamy· Proceedings of the 35th ACM...· 0 citations
Graph databases are increasingly queried through natural language, yet every existing benchmark evaluates isolated single-turn queries rather than the multi-turn sessions through which analysts actually work. We introduce CypherTurn, the first benchmark for conversational Text-to-Cypher evaluation, comprising 721 sessi...
Yu-Zhe Zhang, Wei-Jie Zhu, Hao-Lin Yang et al.· Proceedings of the 2026 Conf...· 0 citations
Large Language Models (LLMs) are increasingly accessible to children through publicly available conversational AI systems, creating new challenges for AI safety and content moderation. While existing jailbreak research has primarily focused on adult-oriented adversarial prompting, considerably less attention has been p...
Szandra Anna Laczi, Valéria Póser· Symposium on Intelligent Sys...· 0 citations
The results point to a fidelity gap in AI-mediated communication, with implications for journalism, policy deliberation, scientific communication, and other domains where opinion-laden messages pass through language models.
Ling-Chong Liu, Yan-Fei Zhou, Jacob Bien et al.· 0 citations
As AI systems are increasingly deployed in high-stakes domains, regulatory and societal demands for transparency and interpretability continue to grow. While classical Explainable AI (XAI) methods produce structured technical artifacts (e.g., feature attributions and counterfactual examples), these remain difficult to...
Gabriele Tolomei, Zi-Heng Chen, Flavio Giorgi et al.· Proceedings of the 35th ACM...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.