"Divide-Then-Aggregate: An Efficient Tool Learning Method via Parallel Tool Invocation"

Parallel tool calls: The key to unlocking the full potential of LLMs in real-world tasks.

Divide-Then-Aggregate, DTA-Llama improves LLM tool use by enabling parallel tool calls, boosting efficiency and performance. It transforms tree-based tool sequences into DAGs, trains on this parallel data, and uses a Process/Threads framework for parallel inferences.

-----

https://arxiv.org/abs/2501.12432

Original Problem: 😥:

→ Current LLMs struggle with efficient tool use for complex real-world tasks.

→ Existing methods like CoT/ReAct use serial tool calls, limiting their scope and efficiency.

→ Tree-based methods suffer from backtracking, increasing cost and time.

-----

Solution in this Paper: 💡:

→ DTA-Llama allows parallel tool calls within each round of tool planning.

→ It transforms tree-based tool sequences into Directed Acyclic Graphs (DAGs).

→ It creates DTA-Tool, a parallel tool invocation dataset based on ToolBench.

→ DTA-Llama is trained on this dataset to perform divide-then-aggregate tool invocation.

→ A Process/Threads framework is used for parallel tool invocation during inference. Process plans and divides tasks; Threads execute independently.

-----

Key Insights from this Paper: 🤔:

→ Parallel tool invocation can significantly improve LLM efficiency and performance.

→ DAG structure enables more efficient tool planning compared to tree-based methods.

→ The Process/Threads framework provides a robust mechanism for parallel inference.

-----

Results: 📊:

→ Llama2-7B with DTA-Llama achieves performance comparable to GPT-3.5 function calling.

→ Reduces token consumption and inference time compared to existing methods.

→ Shows improvements in solvable pass rate (SoPR) and solvable win rate (SoWR).