SCoPE: Shift-Aware Speaker-Conditioned Priors for Emotion Recognition in Conversations

Burak Can Kaplan , Stefan Wermter

doi: 10.21203/rs.3.rs-9065619/v1 - Apr 2026

Associated documents :

<!DOCTYPE html PUBLIC "-//W3C//DTD HTML 4.01//EN" "http://www.w3.org/TR/html4/strict.dtd"> <html> <head> <meta http-equiv="Content-Type" content="text/html; charset=utf-8"> <meta http-equiv="Content-Style-Type" content="text/css"> <title></title> <meta name="Generator" content="Cocoa HTML Writer"> <meta name="CocoaVersion" content="2575.7"> <style type="text/css"> p.p1 {margin: 0.0px 0.0px 0.0px 0.0px; font: 16.0px 'Helvetica Neue'; color: #191c1f; -webkit-text-stroke: #191c1f; background-color: #f6f7f9} span.s1 {font-kerning: none} </style> </head> <body> <p class="p1"><span class="s1">Large Language Models (LLMs) are increasingly used for knowledge-based reasoning tasks, yet understanding when they rely on genuine knowledge versus superficial heuristics remains challenging. We investigate this question through entity comparison tasks by asking models to compare entities along numerical attributes (e.g., âWhich river is longer, the Danube or the Nile?â), which offer clear ground truth for systematic analysis. Despite having sufficient numerical knowledge to answer correctly, LLMs frequently make predictions which contradict this knowledge. We identify three heuristic biases that strongly influence model predictions: entity popularity, mention order, and semantic co-occurrence. For smaller models, a simple logistic regression using only these surface cues predicts model choices more accurately than the modelâs own numerical predictions, suggesting heuristics largely override principled reasoning. Crucially, we find that larger models (32B parameters) selectively rely on numerical knowledge when it is more reliable, while smaller models (7-8B parameters) show no such discrimination, which explains why larger models outperform smaller ones even when the smaller models possess more accurate knowledge. Chain-of-thought prompting steers all models towards using the numerical features across all model sizes.</span></p> </body> </html>

@Article{KW26,
 	 author =  {Kaplan, Burak Can and Wermter, Stefan},
 	 title = {SCoPE: Shift-Aware Speaker-Conditioned Priors for Emotion Recognition in Conversations},
 	 booktitle = {}
 	 journal = {}
 	 editors = {}
 	 number = {}
 	 volume = {}
 	 pages = {}
 	 year = {2026},
 	 month = {Apr},
 	 publisher = {}
 	 doi = {10.21203/rs.3.rs-9065619/v1},
 	 url = {}
 }