DLLG: Dynamic Logit-Level Gating of LLM Experts

In three linesDLLG introduces dynamic logit-level gating to ensemble multiple specialized LLMs. A lightweight gating module predicts token-level fusion weights from response-level supervision alone, without token-level labels or expert retraining. Outperforms routing, heuristic ensembling, and parameter merging baselines on reasoning and code benchmarks.

Summary generated by Claude — human-verified