ULBS: Uniform Local Beam Survival for Semantic-ID Generative Recommendation

Generative recommendation formulates recommendation as conditional generation, offering an alternative to conventional discriminative retrieval and ranking paradigms. However, SID-based generative recommendation relies on constrained beam search at inference, while standard teacher-forcing training fails to account for competition among candidate prefixes. Under this discrepancy, target items with competitive scores can still be pruned prematurely during prefix search. Recent methods partially address this train–inference mismatch but do not directly supervise the cross-parent beam boundary that determines whether a target prefix survives. To bridge this gap, we propose Uniform Local Beam Survival (ULBS), a beam-aligned training objective that optimizes intermediate target-prefix survival against this boundary. At each semantic step, ULBS obtains the beam cutoff from a detached constrained-beam trace and applies a smooth margin loss between the cumulative target-prefix score and this cutoff. This provides uniform supervision across intermediate survival events with localized credit assignment, without adding model parameters or inference-time overhead. Extensive experiments on Amazon benchmarks with LLM backbones demonstrate that ULBS outperforms strong baselines, effectively alleviating the beam admission bottleneck while revealing within-beam ranking as a critical frontier for generative recommendation.