Zhichen Zeng

Low-rank adapters (LoRAs) are a parameter-efficient finetuning technique that injects trainable low-rank matrices into pretrained models to adapt them to new tasks. Mixture-of-LoRAs models expand neural networks efficiently by routing each layer input to a small subset of specialized LoRAs of the layer. Existing Mixture-of-LoRAs routers assign a learned routing weight to each LoRA to enable end-to-end training of the router. Despite their empirical promise, we discover, both theoretically and empirically, that the routing weights often collapse to only one LoRA even when we activate $k>1$ LoRAs during finetuning. When one LoRA has a dominantly large weight, then the computation of the other $k-1$ LoRAs are essentially **wasted** because using $k>1$ would have similar accuracy to $k=1$. This essentially limits the number of effective LoRAs and thus severely hinders the realized expressive power of existing Mixture-of-LoRAs models. How can we address this critical weakness? In this work, we attribute this weakness to the nature of learnable routing weights and rethink the fundamental design of the router. To address this critical issue, we propose a simple yet effective router design that we call *Reinforcement Routing for Mixtures of LoRAs* (ReMix). Our key idea is using **non-learnable** routing weights to ensure all active LoRAs to be equally effective, with no single LoRA dominating the routing weights. However, such non-learnable routing weights make it infeasible to directly train routers via gradient descent. In response, we further propose an unbiased gradient estimator for the router and employ the reinforce leave-one-out (RLOO) technique to reduce the variance of the estimator. Our gradient estimator also enables to scale up training compute to boost the predictive performance of our ReMix. Extensive experiments demonstrate that our proposed ReMix significantly outperforms state-of-the-art parameter-efficient finetuning methods under a small number of activated parameters.

Large language models (LLMs) exhibit diverse strengths and weaknesses across tasks, motivating recent efforts to ensemble multiple models to harness their complementary capabilities, boosting test-time performance. While model diversity and capability are known to influence ensemble effectiveness, a persistent challenge in LLM ensembling arises from mismatched tokenizer vocabularies. Existing alignment strategies typically rely on token-level embeddings or string-level heuristics of tokens, overlooking the tokenizer priors embedded during LLM pretraining. Specifically, tokenizers such as Byte-Pair Encoding (BPE) and Unigram are constructed by statistically analyzing large pretraining corpora to identify frequent subword units, and they tokenize text using greedy or probabilistic algorithms that reflect these learned subword distributions. In this work, we propose a novel and remarkably simple Token Translation (ToT) method that explicitly leverages these tokenizer priors to bridge heterogeneous token spaces. Our method is lightweight, requiring only a few lines of code, pre-computable, and highly efficient at inference. To further enhance robustness, we incorporate token-level model uncertainty to dynamically reweight each model’s contribution during decoding. Extensive evaluations across diverse model combinations and tasks demonstrate that our method consistently outperforms existing ensembling baselines.