CSE-UOI at SemEval-2026 Task 6: A Two-Stage Heterogeneous Ensemble with Deliberative Complexity Gating for Political Evasion Detection

1. Problem Definition

The paper addresses SemEval-2026 Task 6, which focuses on Political Interview Response Clarity Classification. The goal is to automatically classify question-answer pairs from political interviews into three categories:

1. Clear Reply: The politician directly answers the question.
2. Ambivalent: The answer is evasive, vague, or indirect.
3. Clear Non-Reply: The politician explicitly refuses to answer or claims ignorance.

This task is challenging due to subtle pragmatic variations in political discourse, where evasion tactics (e.g., dodging, deflection, generalities) are often indistinguishable from clear answers without deep contextual reasoning.

2. Methodology

The authors propose a two-stage heterogeneous ensemble system that combines cross-model self-consistency with a novel post-hoc correction mechanism.

Stage 1: Heterogeneous Dual-Model Ensemble

Instead of directly predicting the three clarity classes, the system employs an "evasion-first" strategy:

• Models: Two distinct Large Language Models (LLMs) are used in parallel: Grok-4-1-fast-reasoning and Gemini-3-flash-preview.
• Self-Consistency (SC): Each model generates $k=5$ independent responses using Chain-of-Thought (CoT) prompting. The models are instructed to classify the response into 9 fine-grained evasion categories (e.g., Dodging, Deflection, Explicit, Implicit) rather than the final 3 clarity classes.
• Mapping: The 9 evasion labels are deterministically mapped to the 3 target clarity classes. This acts as an "error buffer," as 62.8% of errors at the evasion level still map to the correct clarity bucket.
• Aggregation: An asymmetric weighted voting scheme aggregates the predictions:
  • Grok contributes 5 votes (one per SC sample).
  • Gemini contributes a single "block vote" with a weight of $w=4$ based on its majority evasion label.
  • The final Stage 1 prediction is the clarity class with the highest total weighted vote.

Stage 2: Deliberative Complexity Gating (DCG)

To address the limitations of Stage 1 on "hard" or borderline samples (where models disagree), the authors introduce DCG, a novel, unsupervised, post-hoc correction layer.

• Core Insight: The system exploits the finding that LLM response length correlates with ambiguity. Specifically, longer responses from Gemini often indicate higher deliberative difficulty or ambiguity.
• Signal Discovery: The system uses two behavioral signals to detect uncertainty:
  1. Gemini Response Length ($L_i$): The average length of the 5 SC samples.
  2. Grok Self-Consistency ($s_i$): The consistency score of Grok's 5 samples.
• Gate Logic: A gate triggers an override only if both conditions are met (cross-model confirmation):
  1. The average Gemini response length exceeds the first quartile ($Q_{25}$) of the current batch ($L_i > \theta_1$).
  2. Grok's self-consistency is low ($s_i < 1.0$).
• Correction Mechanism: If the gate triggers, the system overrides Gemini's contribution to the weighted vote, forcing its label to Ambivalent. This is based on the hypothesis that high length + low consistency indicates an ambiguous sample that likely belongs to the "Ambivalent" class.
• Transductive Adaptation: The threshold $\theta_1$ is batch-adaptive (calculated dynamically on the inference batch) rather than fixed, allowing the system to adjust to distribution shifts between development and evaluation sets.

3. Key Contributions

• Heterogeneous Ensemble with Evasion-First Strategy: Moving from direct 3-class classification to a 9-class evasion taxonomy reduces error propagation and leverages the distinct strengths of different model architectures.
• Deliberative Complexity Gating (DCG): A novel, lightweight, post-hoc mechanism that uses cross-model behavioral signals (length and consistency) to adaptively gate reasoning. Unlike multi-agent debate, which increases computational cost and agent count, DCG increases deliberative capacity without extra API calls or training.
• Empirical Validation of Length as a Proxy: The paper provides strong evidence that response length in LLMs serves as a robust proxy for sample ambiguity and model uncertainty in political discourse.
• Comparison with Debate: The authors explicitly compare their static fusion approach against multi-agent debate. They found that debate (iterative peer review) did not improve performance over static weighted voting and actually degraded results, suggesting that model diversity is more effective than conversational consensus for this specific task.

4. Experimental Results

• Performance: The system achieved a Macro-F1 score of 0.8505 on the blind evaluation set.
• Ranking: This secured 3rd place out of 41 submissions.
• Ablation Studies:
  • Stage 1 vs. Stage 2: Adding DCG improved the Evaluation Macro-F1 from 0.8122 (Stage 1) to 0.8505 (Stage 2), a significant gain of ~3.8 points.
  • Debate vs. Static: Multi-agent debate reduced performance (0.7656) compared to the static ensemble (0.7941), confirming that diverse models with weighted voting outperform homogeneous debate loops.
  • Threshold Sensitivity: The batch-adaptive threshold ($Q_{25}$) outperformed fixed thresholds, proving the necessity of transductive adaptation.
• Error Analysis: The system struggled most with the Clear Reply $\leftrightarrow$ Ambivalent boundary (86.8% of remaining errors). Performance was highly correlated with human annotator agreement; samples with low human agreement were significantly harder for the model.

5. Significance and Limitations

• Significance: The paper demonstrates that adaptive deliberation using behavioral proxies (like response length) can significantly boost LLM performance on complex classification tasks without the overhead of fine-tuning or multi-agent debate. It offers a practical framework for detecting political evasion that is robust to model-specific biases.
• Limitations:
  • Domain Dependency: The reliance on response length as an ambiguity proxy may not generalize perfectly to all domains or models.
  • Batch Requirement: The DCG mechanism requires batch inference to calculate the dynamic threshold ($Q_{25}$), making it less suitable for strictly single-sample, real-time streaming scenarios without buffering.
  • Proprietary Models: The system relies on proprietary APIs (Grok and Gemini), which limits reproducibility and deployment in resource-constrained environments, though the authors note the approach could be adapted to open-weight models.

In conclusion, the CSE-UOI system successfully combines heterogeneous model diversity with a smart, data-driven gating mechanism to achieve state-of-the-art results in political response clarity classification, highlighting the potential of "behavioral signal" engineering in LLM ensembles.