All projects

Fed-BLEND — Federated Conformal Prediction for VLMs

Novel federated conformal prediction method that mitigates hallucinations in federated fine-tuned vision-language models via abstention.

Python
PyTorch
LoRA
Qwen2.5-VL
Federated Learning
Conformal Prediction

Problem

Federated fine-tuned VLMs hallucinate confidently, and naive conformal calibration breaks under non-IID client distributions.

Approach

Bin calibration data by predictive entropy, then shrink local per-bin quantiles toward a global anchor with a sample-size-aware weight (Fed-BLEND).

Architecture

5 non-IID clients · LoRA fine-tuned Qwen2.5-VL-3B · per-client entropy-binned calibration · global anchor aggregation · abstention head.

Results

Hallucination 13.05% → 3.79%, useful-answer rate 76.63%, Precision@Commit 95.29%.

Lessons learned

Entropy binning + anchor shrinkage is more robust than per-client conformal scores under non-IID heterogeneity.