Research map / Learning theory and algorithms / Fairness
On Fairness of Task Arithmetic: The Role of Task Vectors
Laura Gomezjurado Gonzalez, Hiroki Naganuma, Kotaro Yoshida, Takafumi Horie, Yuji Naraki, Ryotaro Shimizu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use task addition (lambda>=0.3) instead of FFT/LoRA | LLaMA-2-7B · Berkeley D-Lab hate speech (gender) | DPD / EOD | Improved |
| set lambda in 0.0-0.2 range (ablation) | task addition · Berkeley D-Lab hate speech (gender) | accuracy vs DPD/EOD | Mixed |
| use task addition instead of SFT/LoRA | DistilBERT · Civil Comments (gender) | worst DPD / EOD: worst-DPD 0.0735--0.0812 (LoRA) -> 0.0454; worst-EOD 0.5024--0.5084 -> 0.3358; accuracy 0.9447--0.9453 -> 0.9395 | Improved |
| use task addition instead of SFT/LoRA | DistilBERT · Civil Comments (race) | worst DPD / EOD: worst-DPD 0.0987--0.0995 (SFT) -> 0.0580; worst-EOD 0.2568--0.3544 -> 0.1289 | Improved |
| use task addition instead of LoRA | Qwen2.5-0.5B · Civil Comments (gender) | accuracy and DPD/EOD: accuracy 0.774--0.790 -> 0.810--0.820; EOD 0.232--0.362 -> 0.130--0.143 | Improved |
| use task addition instead of SFT | Qwen2.5-0.5B · Civil Comments (gender) | accuracy and EOD: accuracy 0.884--0.886 -> 0.810--0.820; EOD 0.060--0.084 -> 0.130--0.143 | Hurt |
| inject Men or Asian subgroup task vector into FFT model (ablation) | FFT LLaMA-2-7B · Berkeley D-Lab (gender / race) | DPD / EOD | Improved |
| inject Native American subgroup task vector into FFT model (ablation) | FFT LLaMA-2-7B · Berkeley D-Lab (race) | DPD / EOD | Hurt |
| use task addition instead of FFT | LLaMA-2-7B · Berkeley D-Lab subgroups | subgroup fairness: improves in 5 of 7 gender and 3 of 8 race subgroups | Mixed |
Closest papers by meaning
- FairMT: Fairness for Heterogeneous Multi-Task Learning (NeurIPS 2026 · similarity 0.88)
- FFB: A Fair Fairness Benchmark for In-Processing Group Fairness Methods (ICLR 2024 · similarity 0.87)
- FairNet: Dynamic Fairness Correction without Performance Loss via Contrastive Conditional LoRA (NeurIPS 2025 · similarity 0.86)
- FRAPPÉ: A Group Fairness Framework for Post-Processing Everything (ICML 2024 · similarity 0.86)
- Revisiting Fairness in Multitask Learning: A Performance-Driven Approach for Variance Reduction (CVPR 2025 · similarity 0.86)
- Some Optimizers are More Equal: Understanding the Role of Optimizers in Group Fairness (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
