TOC_Proposal.MD

Fairness Pruning: Bias Mitigation through Activation-Guided MLP Width Pruning in Large Language Models

Author: Pere Martra
Program: Master in Research in Artificial Intelligence (UIMP)
Format: IEEE double column, 12-14 pages
Target: UIMP defense (July 2026) · TMLR submission (post-defense)

---

Table of Contents

• Abstract (~half column)

• 1. Introduction (~1 page)

  • 1.1 Motivation: bias in LLMs and limitations of existing mitigation methods
  • 1.2 Gap: no prior work on bias localization via activation-guided MLP pruning
  • 1.3 Contributions
  • 1.4 Document structure https://github.com/peremartra/fairness-pruning/blob/main/TOC_Proposal.MD

• 2. Related Work (~1.5 pages)

  • 2.1 Demographic Bias & Mitigation strategies
  • 2.2 Pruning & Interpretability

• 3. Methodology (~2 pages)

  • 3.1 Pipeline overview (flow figure)
  • 3.2 Prompt pair datasets — construction, token constraint, HuggingFace
  • 3.3 Neuron bias detection — OptiPFair, FairnessPruningScore formula
  • 3.4 Fairness pruning — selective layer pruning with layer_indices
  • 3.5 Evaluation protocol — capabilities benchmarks, bias benchmarks, metrics

• 4. Experimental Setup (~0.5 pages)

  • Models: Llama-3.2-1B (primary), Llama-3.2-3B and Salamandra-2B (validation)
  • Hardware, dtypes, reproducibility

• 5. Results (~4-5 pages)

  • 5.1 Baseline capabilities and bias (BBQ + EsBBQ)
  • 5.2 Bias Localization: Depth, Neural Structure, and Circuit Specificity (RQ1 + RQ2)
  • 5.3 Pruning results — bias reduction vs. capability retention (RQ3)
  • 5.4 Cross-model validation — Llama-3B and Salamandra

• 6. Discussion (~1 page)

  • 6.1 Two encoding strategies: distributed vs. sparse bias
  • 6.2 Limitations — Western-centric scope, models <4B, MLP-only pruning

• 7. Conclusions (~0.5 pages)

  • 7.1 Answers to RQ1, RQ2, RQ3
  • 7.2 Implications and future work

• References (~1-1.5 pages)