模型架構與代理工具使用
Recurrent Inference Improves Multi-Tool Composition, Raising the 1B Llama’s Average BFCL Score from 21.4 to 32.9
Researchers repeatedly update hidden states through the same set of Transformer layers before generating each token, comparing recurrent and conventional models using identical data and LoRA configurations. Gains are concentrated in parallel calls and cross-call dependencies, while API-Bank, which primarily involves single API calls, shows no consistent advantage.
A conventional autoregressive model passes data through a Transformer of fixed depth once for each token it generates. A looped language model instead reuses the same parameters to iteratively update its hidden state before outputting a token. New research examines whether this approach—adding depth without adding weights—can improve an agent’s ability to select functions, compose multiple calls, and correctly bind the output of one API to the parameters of the next.
The team compared the natively recurrent Ouro-1.4B and Ouro-2.6B architectures with recurrent variants adapted from Llama-3.2-1B and OLMo-2-1B. All controlled models were trained using the Hermes function-calling dataset, the same LoRA rank of 32, two epochs, a 4,096-token limit, and an A100 80GB GPU. Across the average results for BFCL’s Simple, Multiple, Parallel, and Parallel-Multiple categories, Llama-3.2-1B scored 21.4 after conventional supervised fine-tuning (SFT), compared with 32.9 for the recurrent version. In the Parallel category, its score rose from 14.0 to 31.0. OLMo-2-1B’s average increased more modestly, from 39.1 to 41.8.
NESTful, which requires models to process the output of a previous call, reveals the effects of additional depth more clearly. The SFT version of Ouro-2.6B achieved a win rate of 0.371, exceeding the 0.345 recorded by Qwen3-8B-Instruct in the comparison. When the model was held constant and the number of recurrent iterations was increased, scores on nested workflows generally continued to rise. An adaptive gate that lets the model decide at each token whether to stop iterating also maintained the same NESTful score with fewer than four iterations on average, compared with a fixed four-iteration configuration.
However, Ouro has no non-recurrent counterpart pretrained under exactly the same conditions, so the entire performance gap cannot be attributed to recurrence. The adapted models also remained well behind the natively recurrent models on deeply nested tasks. All three benchmarks are static, single-turn evaluations and do not yet cover tool failures, retries, or state drift. Future work should measure real-world per-token latency, KV cache usage, and memory costs, and verify in executable multi-turn environments whether adaptive depth remains more cost-effective than simply using a larger model.