timchen0618/browsecomp-plus-selected-tools-analysis-v1

---
license: mit
tags:
- browsecomp-plus
- search-agent
- selected-tool-calls
---

# BrowseComp-Plus: Selected Tools Analysis

Side-by-side view of selected tool calls from a reference trajectory alongside the new agent trajectory conditioned on those steps.

**Retrieval model**: Qwen3-Embedding-8B
**Agent model**: gpt-oss-120b
**Run**: traj_summary_ext_selected_tools_gpt-oss-120b_seed0

## Columns

| Column | Description |
|--------|-------------|
| `query_id` | Query identifier |
| `rationale` | GPT rationale for why these k steps were selected from the reference trajectory |
| `selected_indices` | Step indices selected from the original reference trajectory |
| `k_requested` | Number of steps requested to select |
| `k_effective` | Number of steps actually selected |
| `excerpt` | Concatenated selected steps from reference trajectory (reasoning + tool calls + results) |
| `new_trajectory` | New agent trajectory conditioned on selected tool calls (system prompt omitted) |
| `direct_answer` | True if agent answered directly with NO tool calls |
| `tool_call_counts` | JSON dict of tool type → count for the new trajectory |
| `status` | Trajectory completion status |

## Key Statistic

**Direct-answer rate: 560/830 = 67.5%**

67.5% of trajectories had zero tool calls — the agent answered directly from the selected tool call context alone, without performing any new searches.