FairCompressAgent (FCA): agentic framework for fairness-aware model compression targeting FPGA deployment

KEY POINTS

1. The paper introduces FairCompressAgent (FCA), an agentic framework that composes fairness-aware pruning, incremental quantization, and sparse low-rank factorization under a common operator interface; a language-model planner uses model profiles and measured outcomes to select compression configurations while an execution layer handles compression, fine-tuning, evaluation, and constraint-based selection.
2. Evaluated on Fitzpatrick-17k with VGG-11, FCA under an accuracy-constrained request produced a compressed model with 59.54% less inference tensor storage, increased validation average precision from 0.5141 to 0.5233, and reduced equalized opportunity from 0.2251 to 0.2168, while reaching the same final selection as one-shot planning with fewer candidate evaluations (7.33 vs 12 on average).
3. FCA shows how measured feedback and an LM-based planner can automate trade-offs among accuracy, fairness, and deployment cost for constrained targets (e.g., FPGA), enabling interactive, constraint-aware compression workflows.