Research map / Multi-agent systems and games / LLM agents

SEVADE: Self-Evolving Multi-Agent Analysis with Decoupled Evaluation for Hallucination-Resistant Sarcasm Detection

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SEVADE, a multi-agent GPT-4o framework (DARE) that builds a reasoning chain via self-evolving linguistic agents, then a fine-tuned BERT rationale adjudicator classifies sarcasm from that chain alone.

Problem

LLM-based sarcasm detection suffers from single-perspective analysis, static reasoning pathways, and hallucination when synthesizing conflicting cues into a final judgment.

Approach

Decouples reasoning from judgment: a Controller Agent dynamically refines and expands a team of linguistically grounded agents to produce a structured reasoning chain, and a lightweight BERT adjudicator (last 2 layers fine-tuned) classifies based solely on that chain.

Key results

• Average Accuracy 78.14 and Macro-F1 77.90 across four datasets, improving 7.01% and 6.55% over the strongest baseline DC-Net.
• On MuSTARD and SemEval-2018, accuracy gains over the strongest baseline are 7.75% and 10.61%.
• Removing the Rationale Adjudicator (replacing with base LLM) lowers IAC-V1 accuracy from 0.7304 to 0.7147.

Surprising findings

• Smaller Qwen 2 / Llama 3 models performed better on SemEval-2018, with larger models over-analyzing and producing less coherent rationales.
• The dominant error mode is false positives on non-sarcastic text, attributed to benign disagreement among agents.

Limitations

• Higher false-positive rate on non-sarcastic samples due to agent disagreement not resolved by simple weighting.
• Relies on GPT-4o API for all agents, with web search support as an auxiliary component.

Experimental setup

Datasets: IAC-V1 IAC-V2 MuSTARD SemEval-2018 Task 3

Benchmarks: IAC-V1 IAC-V2 MuSTARD SemEval-2018 Task 3

Metrics: Accuracy Macro-F1

Baselines: MIARN SAWS DC-Net BERT RoBERTa GPT-4o GPT-4o+CoC GPT-4o+GoC GPT-4o+BoC GPT-5

Builds on: AutoGen CAMEL CAF-I SarcasmCue

Closest papers by meaning

• Well, Now We Know! Unveiling Sarcasm: Initiating and Exploring Multimodal Conversations with Reasoning (AAAI 2024 · similarity 0.86)
• Multi-Agent Undercover Gaming: Hallucination Removal Through Counterfactual Test for Multimodal Reasoning (AAAI 2026 · similarity 0.86)
• Is Sarcasm Detection a Step-by-Step Reasoning Process in Large Language Models? (AAAI 2025 · similarity 0.85)
• Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs (ICML 2024 · similarity 0.83)
• InEx: Hallucination Mitigation via Introspection and Cross-Modal Multi-Agent Collaboration (AAAI 2026 · similarity 0.83)
• iMAD: Intelligent Multi-Agent Debate for Efficient and Accurate LLM Inference (AAAI 2026 · similarity 0.83)