@inproceedings{dinh-2026-one,
    title = "One and Only at {S}em{E}val-2026 Task 2: Evaluating Zero-Shot Autonomous {LLM} Agents and Heuristic Proxies in Ecological Affect Forecasting",
    author = "Dinh, Nam",
    editor = "Kochmar, Ekaterina  and
      Ghosh, Debanjan  and
      North, Kai  and
      Komachi, Mamoru",
    booktitle = "Proceedings of the 20th {I}nternational {W}orkshop on {S}emantic {E}valuation (2026)",
    month = jul,
    year = "2026",
    address = "San Diego, California, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl-workshops/2026.semeval-1.162/",
    pages = "1205--1211",
    ISBN = "979-8-89176-414-9",
    abstract = "This paper presents team One and Only{'}s sys-tem for SemEval-2026 Task 2: PredictingVariation in Emotional Valence and Arousalover Time (Soni et al., 2026). We investigatewhether zero-shot LLM reasoning can replacefine-tuning for ecological affect forecasting bycombining deterministic statistical priors withfrozen LLMs (Gemini 3 Pro, Claude Opus4.6, GPT-5.2). For short-term state changes(Subtask 2A), an OLS mean-reversion anchoris paired with LLM-generated impulses; forlong-term disposition changes (Subtask 2B),a Chain-of-Thought prompt drives direct nu-meric prediction. Our system underperformsfine-tuned approaches on both subtasks. How-ever, post-submission ablation across threeLLMs reveals a task-dependent pattern: CoTreasoning substantially improves dispositionforecasting (rV : {\ensuremath{-}}0.185 {\textrightarrow} +0.129; MAEV :0.899 {\textrightarrow} 0.422), while uncalibrated LLM im-pulses degrade state-change prediction due tovariance collapse ({\ensuremath{\sigma}}pred = 0.41 vs. {\ensuremath{\sigma}}gold =1.73). We provide a detailed diagnostic anal-ysis of these failure modes and release allprompts and outputs for reproducibility."
}