LLM Agent Tool-Use Reasoning Strategy

Objective

Design a better search/reasoning strategy for an LLM-based tool-use agent. Your code goes in the search() method of custom_search.py.

Background

StableToolBench evaluates LLM agents on multi-step tool use tasks. Given a user query and a set of tool APIs, the agent must decide which tools to call, with what arguments, and in what order to arrive at a final answer.

The search strategy controls how the agent explores the action space:

• Greedy chain (CoT): Call LLM, execute tool, repeat. No backtracking. Simple but gets stuck on errors.
• DFS with ranking: Generate multiple children, use LLM to rank them, expand best first. Backtracks on failure. More robust but costs extra LLM queries for ranking.
• DFSDT: Generate one child, immediately recurse depth-first. Backtrack a fixed number of steps on failure. Balance between exploration and cost.

What you can modify

The search(self, root_node) method in custom_search.py (the editable region). You have access to:

• self._step(node) -- one LLM call + tool execution, returns new leaf nodes
• self._add_diversity_prompt(node) -- encourages different actions when re-expanding
• self._rank_nodes(candidates) -- LLM pairwise ranking (costs extra queries)
• Tree state: self.query_count, self.max_query_count, self.terminal_node, etc.
• Node properties: node.is_terminal, node.pruned, node.observation_code, node.get_depth()

Evaluation metrics

• pass_rate: Fraction of queries where the agent produces a valid final answer (higher is better)
• avg_queries: Average number of LLM queries per task (lower is better for efficiency)
• give_up_rate: Fraction of queries where the agent gives up (lower is better)