Token-Level LLM Collaboration via FusionRoute

Large language models (LLMs) exhibit strengths across diverse domains. However, achieving strong performance across these domains with a single general-purpose model typically requires scaling to sizes that are prohibitively expensive to train and deploy. On the other hand, while smaller domain-specialized models are much more efficient, they struggle to generalize beyond their training distributions. To address this dilemma, we propose FusionRoute, a robust and effective token-level multi-LLM collaboration framework in which a lightweight router simultaneously (i) selects the most suitable expert at each decoding step and (ii) contributes a complementary logit that refines or corrects the selected expert’s next-token distribution via logit addition. Unlike existing token-level collaboration methods that rely solely on fixed expert outputs, we provide a theoretical analysis showing that pure expert-only routing is fundamentally limited: unless strong global coverage assumptions hold, it cannot in general realize the optimal decoding policy. By augmenting expert selection with a trainable complementary generator, FusionRoute expands the effective policy class and enables recovery of optimal value functions under mild conditions. Empirically, across both Llama-3 and Gemma-2 families and diverse benchmarks spanning mathematical reasoning, code generation, and instruction following, FusionRoute outperforms both sequence- and token-level collaboration, model merging, and direct fine-tuning, while remaining competitive with domain experts on their respective tasks.

Today's most capable AI language models are extremely large and expensive to run. A cheaper alternative is to use several smaller models, each specialized in one area such as math, coding, or writing, but these specialists struggle with tasks outside their expertise. Our work asks: can we make a team of small specialists perform as well as one large generalist? We introduce FusionRoute, a system that coordinates multiple specialist AI models to collaboratively generate text one word at a time. At each step, a lightweight "router" decides which specialist should produce the next word, and also provides its own small correction to improve that specialist's output. This correction mechanism is key: we prove mathematically that simply switching between specialists — without any ability to adjust their outputs — has inherent limitations and cannot always produce the best possible answer. We test FusionRoute on tasks spanning math problem-solving, code generation, and instruction following using two popular model families (Llama-3 and Gemma-2). Our system consistently outperforms existing approaches for combining multiple models, including sequence-level selection algorithm, model merging, previous token-level collaboration, or fine-tuning a single model. Hence, FusionRoute can be regarded as an effective token-level collaboration paradigm that works robustly, efficiently, and automatically across all scenarios.