Learning-to-Transition Framework for Large-Scale High-Order MIMO Detection

Learning-to-Transition for Large-Scale and High-Order MIMO Detection

Authors: Yubo Zhang, Yiyao Liu, Xiaodong Wang Published: 2026-08-17 arXiv: 2508.08539

Summary

High-order multiple-input multiple-output (MIMO) detection requires efficient search over a large discrete symbol space while producing reliable soft information for channel decoding. This paper develops a learning-to-transition (L2T) framework that formulates MIMO detection as a stochastic sequence of complete-vector transitions.

Method

• A channel-coupled Transformer updates both the instance embedding and the sampling policy at each transition.
• A blockwise autoregressive factorization captures inter-stream dependence with moderate sequential complexity.

Hard-Output Detection

• The transition network is applied recursively.
• Training uses a residual-to-BER curriculum:
1. First learns the MIMO search geometry from the exact residual metric. 2. Then aligns the policy with transmitted bit accuracy.

Soft-Output Reception

• The trained hard policy is cloned at the parameter level into each layer of an untied soft-input soft-output iterative detection-and-decoding (IDD) receiver.
• This tied-to-untied transfer preserves learned zero-prior search dynamics while supporting layer- and round-specific specialization under decoder feedback.
• Within each IDD round:
• Decoder priors tilt candidate generation according to Bayes' rule.
• Likelihood-weighted terminal hypotheses produce posterior extrinsic log-likelihood ratios (LLRs) for LDPC decoding.

Training Schedule

A multi-stage strategy progressively exposes the receiver to synthetic and on-loop decoder-generated priors, further stabilizing the hard-to-soft transfer.

Key Points

• Reformulates MIMO detection as a sequence of complete-vector transitions.
• Channel-coupled Transformer jointly learns embeddings and sampling policies.
• Blockwise autoregressive factorization reduces sequential complexity.
• Residual-to-BER curriculum first captures geometry, then aligns with bit accuracy.
• Tied-to-untied parameter transfer from hard to soft detection.
• Bayesian tilting of candidate generation using decoder priors in each IDD round.
• Extrinsic LLRs fed to LDPC decoder for iterative refinement.
• Progressive exposure to synthetic and on-loop priors stabilizes training.