HybridAgent combines latent reasoning with tool use in text
Esta página aún no tiene una traducción verificada. Se muestra la versión en inglés. Leer la versión en inglés
En palabras sencillas
Language model agents can write text to explain their reasoning and use tools. Other models can reason in a more compact internal form, but that makes tool use and inspection harder.
HybridAgent switches between internal reasoning and text for tool calls, and keeps the conversation and tool results available across turns.
On one 500-problem math test, its 8B version scored slightly higher than a text-based agent trained with reinforcement learning; that comparison does not cover every test or model size.
La investigación
Text-based agents can use tools, but generate their intermediate reasoning as text. Latent reasoners can avoid decoding every step, yet lack native tool pathways and readable thoughts. How can an agent combine internal reasoning with multi-turn external interaction?
HybridAgent links latent steps to text inputs, carries hybrid context across turns, and trains agents to switch modes for reasoning and tool use.
HybridAgent switches between latent reasoning and text for tool calls, carrying text, latent steps and tool observations across turns. A proof shows its memory preserves the full history’s effect on the next output under fixed model and inference conditions.
On MATH500, a 500-problem test of math reasoning, its 8B version after reinforcement learning scored 84.6% accuracy, versus 82.8% for ARPO under matched backbone, RL data and training budgets.
The MATH500 comparison is one result at one model scale, so it does not establish the same difference across all 11 tests or model sizes. The memory result is a proof under fixed-model and consistent-inference conditions, not an empirical performance finding.
Análisis
Detalles
- Autores del paper original
- Jiaru Zou, Yunzhe Qi, Sirui Chen, Federico Bianchi, Shang Zhu
- Enviado a arXiv
- Licencia
- CC BY 4.0