Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Yes FLoReNce, I Will Do Better Next Time! Agentic Feedback Reasoning for Humorous Meme Detection

Olivia Shanhong Liu, Pai Chet Ng, De Wen Soh, Konstantinos N. Plataniotis

TL;DR

The paper tackles humor understanding in multimodal memes by framing reasoning as a closed-loop process modulated by critique. It introduces FLoReNce, a framework that couples a frozen vision–language reasoning agent with a Judge, a PID controller, and a feedback-informed non-parametric KB to adapt prompts at inference without fine-tuning. Through PrideMM experiments, FLoReNce achieves competitive predictive performance and enhanced reasoning quality, demonstrating that memory-guided, feedback-driven prompting can stabilize and improve subjective meme interpretation. The work presents a principled approach to controllable, adaptive multimodal reasoning applicable to nuanced social content analysis.

Abstract

Humorous memes blend visual and textual cues to convey irony, satire, or social commentary, posing unique challenges for AI systems that must interpret intent rather than surface correlations. Existing multimodal or prompting-based models generate explanations for humor but operate in an open loop,lacking the ability to critique or refine their reasoning once a prediction is made. We propose FLoReNce, an agentic feedback reasoning framework that treats meme understanding as a closed-loop process during learning and an open-loop process during inference. In the closed loop, a reasoning agent is critiqued by a judge; the error and semantic feedback are converted into control signals and stored in a feedback-informed, non-parametric knowledge base. At inference, the model retrieves similar judged experiences from this KB and uses them to modulate its prompt, enabling better, self-aligned reasoning without finetuning. On the PrideMM dataset, FLoReNce improves both predictive performance and explanation quality over static multimodal baselines, showing that feedback-regulated prompting is a viable path to adaptive meme humor understanding.

Yes FLoReNce, I Will Do Better Next Time! Agentic Feedback Reasoning for Humorous Meme Detection

TL;DR

The paper tackles humor understanding in multimodal memes by framing reasoning as a closed-loop process modulated by critique. It introduces FLoReNce, a framework that couples a frozen vision–language reasoning agent with a Judge, a PID controller, and a feedback-informed non-parametric KB to adapt prompts at inference without fine-tuning. Through PrideMM experiments, FLoReNce achieves competitive predictive performance and enhanced reasoning quality, demonstrating that memory-guided, feedback-driven prompting can stabilize and improve subjective meme interpretation. The work presents a principled approach to controllable, adaptive multimodal reasoning applicable to nuanced social content analysis.

Abstract

Humorous memes blend visual and textual cues to convey irony, satire, or social commentary, posing unique challenges for AI systems that must interpret intent rather than surface correlations. Existing multimodal or prompting-based models generate explanations for humor but operate in an open loop,lacking the ability to critique or refine their reasoning once a prediction is made. We propose FLoReNce, an agentic feedback reasoning framework that treats meme understanding as a closed-loop process during learning and an open-loop process during inference. In the closed loop, a reasoning agent is critiqued by a judge; the error and semantic feedback are converted into control signals and stored in a feedback-informed, non-parametric knowledge base. At inference, the model retrieves similar judged experiences from this KB and uses them to modulate its prompt, enabling better, self-aligned reasoning without finetuning. On the PrideMM dataset, FLoReNce improves both predictive performance and explanation quality over static multimodal baselines, showing that feedback-regulated prompting is a viable path to adaptive meme humor understanding.
Paper Structure (27 sections, 12 equations, 3 figures, 2 tables)

This paper contains 27 sections, 12 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Humorous Meme Detection
  4. LLM-Based Multi-Agent Frameworks
  5. FLoReNce Framework
  6. Problem Statement.
  7. Closed-Loop Learning
  8. Agent reasoning.
  9. Judge feedback and control.
  10. Experience-based KB update.
  11. Open-Loop Inference
  12. Query embedding and retrieval.
  13. Compact KB signal and control.
  14. Final reasoning and inference.
  15. Experiments Setup
  16. ...and 12 more sections

Figures (3)

  • Figure 1: Comparison of our feedback-based reasoning correction approach (bottom) with prior approach (upper).
  • Figure 2: FLoReNce framework. In the closed-loop learning (top), the VLM reasons over memes, receives judge feedback, and stores judged embeddings in the Knowledge Base ($\mathcal{K}$). In the open-loop inference (bottom), $\mathcal{K}$ provides retrieval-based guidance for adaptive prompting on unseen memes, distinguishing humorous from non-humorous cases.
  • Figure 3: Examples of retrieved memes with feedbacks. Correct prediction in green. Incorrect prediction in red.