Guocheng He, Ziyan An, Meiyi Ma — Journal of Artificial Intelligence Research Vol. 86 (2026). Full title: Formal Logic Inference Guided Uncertainty Quantification for Personalized Federated Learning. Published 8 Jul 2026. Verified via the JAIR abstract page (jair.org/index.php/jair/article/view/21429). Note: “alignment” in this paper means semantic alignment of client time-series behavior for clustering, not LLM safety alignment.
Core Contribution
Personalized federated learning (FL) that combines Signal Temporal Logic (STL) — to cluster clients by temporal semantic behavior rather than raw statistical distance — with decentralized Conformal Prediction (CP) inside each cluster — to produce distribution-free prediction intervals with coverage guarantees. Supports runtime assignment of new clients to clusters without retraining. Empirically evaluated on three real forecasting domains; reports up to ~95% improvement in client-level MSE vs BNN / clustering / CP baselines while staying scalable.
Method
Problem setting
FL across heterogeneous clients (sensors / sites) where:
- Data never leave the client (privacy).
- Client distributions differ (non-IID).
- Predictions need uncertainty quantification, not just point forecasts.
- New clients may appear at deployment time and must be slotted into an existing personalization structure.
Bayesian NN and naïve clustering baselines struggle with scale and consistent personalization under those constraints (authors’ claim).
Stage 1 — STL property inference & semantic clustering
Signal Temporal Logic provides a formal language for temporal properties over signals (e.g., bounds, eventually/always patterns, response constraints). LogiCP:
- Extracts / infers temporal logical properties that characterize each client’s local time series.
- Measures semantic similarity between clients in that logical property space (not only Euclidean distance on raw features or embeddings).
- Forms clusters with controlled intra-cluster heterogeneity — clients whose STL-described behaviors align share a personalized model / UQ apparatus.
This is the “formal logic inference guided” half: logic is used as a representation for grouping, not as a verifier bolted on after training.
Stage 2 — decentralized conformal prediction per cluster
Inside each cluster, apply Conformal Prediction in a decentralized fashion:
- CP yields prediction intervals with finite-sample, distribution-free coverage guarantees under exchangeability assumptions (standard CP theory).
- “Decentralized” = the conformal calibration procedure is organized to respect FL constraints (client data stay local; exact communication pattern is PDF-level).
- Guarantees are that intervals cover the true label/value at a target rate — UQ you can state mathematically, unlike many heuristic variance heads.
Stage 3 — dynamic runtime client assignment
New clients can be assigned to an existing cluster at runtime from their observed STL properties without retraining the whole federation — practical for streaming sensor networks where sites come and go.
Dataset / evaluation protocol
Confirmed from the JAIR abstract:
| Dataset domain | Role |
|---|---|
| Traffic forecasting | Real-world distributed sensors / flow prediction |
| Temperature prediction | Real-world environmental time series |
| Electricity demand forecasting | Real-world smart-grid-style load prediction |
Baselines: Bayesian Neural Network approaches, clustering-based personalized FL, and conformal-prediction-based methods.
Primary metric highlighted: client-level MSE, with up to 95% improvement in the best reported comparison; scalability emphasized alongside accuracy.
PDF still needed for: number of clients per dataset, train/test split, STL formula templates, CP miscoverage level \(\alpha\), communication rounds, and per-dataset tables.
Related prior work by the same line (AAAI 2024 Formal Logic Enabled Personalized Federated Learning through Property Inference, An, Johnson, Ma) used traffic data across fifteen U.S. states plus a synthetic smart-city multi-task setting — useful context for what “traffic” likely looks like, but do not assume identical corpora without PDF confirmation.
Limitations
- Wrong “alignment” for my survey if skimmed carelessly. This will false-positive on keyword alerts for AI alignment; it is FL semantic clustering.
- Domain = tabular/sensor forecasting, not LLMs; no red-teaming content.
- STL property design still requires domain-meaningful predicates — the formal layer doesn’t remove modeling choices, it structures them.
- CP coverage guarantees depend on exchangeability; strong non-stationarity in traffic/electricity may stress that assumption (discuss in PDF limitations).
Relevance to My Niche
None of Directions 1–3 directly. Logged because it cleared the venue whitelist (JAIR) in the Week 2 Scholar firehose and teaches a triage lesson: “alignment” ≠ safety alignment. Keep in the completeness bucket; do not let it dilute the quantization / multilingual / attribution narrative.