FedMark-FM: New market framework boosts federated model accuracy by 8 points

FedMark-FM tackles a critical gap in federated learning: existing incentive mechanisms treat clients as homogeneous data providers, but foundation-model adaptation involves heterogeneous artifacts—retrieval corpora, LoRA adapters, preference data, prompts, and update sketches—each with different value, privacy constraints, and vulnerability to strategic behavior. The framework introduces S3Val, a stratified Shapley estimator that handles pipeline-ordered valuation (where contributions depend on the sequence of processing) and uses lower-confidence-bound estimates to compute budget-feasible payments. This design penalizes duplication, sybil splitting, poisoned adapters, privacy-budget gaming, and cost inflation, creating an auditable ledger of contributions.

On the FedMark-FM-Bench across three tracks (FEVER retrieval, generator-backed RAG, and trained PEFT/LoRA), the system improved downstream accuracy by 7.5-8.1 points over volume-based, leave-one-out, and FL-Shapley baselines while selecting zero strategic clients. Split-conformal calibration achieved full lower-bound coverage with a mean width of 0.0141, drastically tighter than naive intervals (0.33). The authors also prove that pipeline-ordered valuation is the unique credit rule respecting serving causality, and show that it materially changes credit assignment (Spearman 0.76, selected-set overlap 0.67) without harming task quality. At 200-1000 client scale, the market preserves rare specialists and remains auditable.