Perspicere at SemEval-2026 Task 2: Modeling Longitudinal Valence and Arousal via Dense Embeddings and Agentic Reasoning

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

Perspicere's system for SemEval 2026 Task 2, Subtask 1, addresses longitudinal affect assessment by modeling Valence and Arousal trajectories. The system employs a tripartite framework, integrating zero-context feature extraction, latent temporal modeling with LSTM, and explicit semantic reasoning via a Teacher-Guided Clinical Reasoning Agent. Evaluation revealed that robust static feature extraction, specifically Matryoshka-distilled Jasper embeddings combined with XGBoost, achieved superior performance over explicit sequence modeling. This configuration yielded a Valence composite r of 0.654, representing a 17.4% improvement over the baseline, while mitigating overfitting. Although the reasoning agent struggled with high-frequency fluctuations, it demonstrated a distinct advantage in user-level affective profiling, achieving the highest Between-User Valence correlation of r = 0.725 through SOTA psychological profiling. The study also identified a persistent "arousal bottleneck," confirming text-only modeling limitations for physiological activation.

Key takeaway

For machine learning engineers developing longitudinal affect assessment systems, prioritize robust static feature extraction over complex temporal models for general accuracy. Your focus should be on high-quality embeddings like Matryoshka-distilled Jasper with XGBoost, which demonstrated a 17.4% Valence improvement. However, if user-level psychological profiling is critical, integrate agentic reasoning, as it achieved a 0.725 Between-User Valence correlation. Be aware that text-only models inherently limit physiological arousal detection.

Key insights

Robust static feature extraction can outperform complex temporal models for longitudinal affect assessment.

Principles

• Static embeddings with XGBoost offer speed-accuracy balance.
• Agentic reasoning excels in user-level affective profiling.
• Text-only models limit physiological arousal detection.

Method

The system combines zero-context feature extraction, LSTM-based temporal modeling, and explicit semantic reasoning using a Teacher-Guided Clinical Reasoning Agent with in-context learning for affect assessment.

In practice

• Employ Matryoshka-distilled Jasper embeddings for affect modeling.
• Use XGBoost for efficient, accurate static feature classification.
• Consider agentic reasoning for user-specific psychological profiling.

Topics

• SemEval 2026
• Affect Assessment
• Longitudinal Modeling
• Valence and Arousal
• Matryoshka Embeddings
• XGBoost
• Agentic Reasoning

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Correlation Is Not Enough: Embedding Human Metadata for Individual Causal Discovery — Biomedical language models require explicit causal embedding separation to prevent false causal links in LBMs.
• PPG-Based Affect Recognition with Long-Range Deep Models: A Measurement-Driven Comparison of CNN, Transformer, and Mamba Architectures — CNNs remain highly effective for PPG-based affect recognition, often outperforming newer long-range models.
• Bayesian Spectral Emotion Transition Discovery from Multi-Annotator Disagreement — Preserving annotator uncertainty in multi-rater judgments is crucial for understanding emotion dynamics and aligning with psychological theory.
• Emotion-Attended Stateful Memory (EASM):The Architecture for Hyper-Personalization at Scale — Emotion-attended stateful memory significantly enhances language model personalization by integrating user history and emotional signals.
• Deep Temporal Modeling and Ensemble Fusion for Multimodal Emotion Recognition from Physiological Signals — Combining deep temporal models with sensor and ensemble fusion significantly boosts physiological emotion recognition accuracy.
• Multimodal Hidden Markov Models for Persistent Emotional State Tracking — Sticky factorial HDP-HMMs effectively track persistent emotional regimes in multimodal conversations, outperforming standard HMMs and enhancing LLM responses.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.