王子睿
restructure + add files
8c10cf2
|
Raw
History Blame Contribute Delete
5.56 kB

AI-NativeBench: Processed Results & Analysis Artifacts

This directory contains processed experiment outputs, aggregated statistics, and analysis/visualization scripts for the paper "AI-NativeBench: An Open-Source White-Box Agentic Benchmark Suite for AI-Native Systems".

Compared with the raw dataset (see README-RAW.md), this RESULTS/ folder additionally includes:

  • Derived artifacts such as execution_path.md (extracted trace trees), per-run metadata, and task outputs.
  • Aggregated tables (CSV) and paper-ready figures (PDF) for RQ1–RQ3.
  • Scripts to reproduce the aggregation and plots.

Repository layout

  • Model-level run outputs (per-model folders):

    • GPT-5/
    • GPT-4o-mini/
    • DeepSeek-V3-1/
    • DeepSeek-R1/
    • Gemini-2.5-flash/
    • Gemini-2.5-flash-nothinking/
    • Qwen3-235b/
  • RQ-level analysis artifacts:

    • RQ1/: trajectory/success/retry evaluations and radar charts
    • RQ2/: agent-time vs. LLM/tool share breakdowns + plots
    • RQ3/: performance summaries (time/tokens), comparisons, and violin plots
  • Trace processing utilities:

    • extract_trace_tree.py: generates execution_path.md from langfuse_trace.json
    • batch_extract_trees.sh: batch-generate execution_path.md for all runs
    • README-extract_trace_tree.md: detailed usage and output format

Model folders (per-run outputs)

Each model folder contains multiple application/architecture subfolders (e.g., BookWriter-MCP, SQLAssistant-A2A).

Architecture naming conventions

  • No suffix: pure framework baseline (e.g., EmailResponder/, GameBuilder/, MarkdownValidator/)
  • *-A2A: A2A protocol variant
  • *-H_A2A: hard-coded / heterogeneous A2A baseline (in some plots/scripts, this may be treated as an A2A-mix / H-A2A variant)
  • *-MCP: tool calling via MCP

Typical per-run structure

Under each application folder, test_results/ stores multiple runs (commonly run_*/ or session_*/ directories). A typical run contains:

  • langfuse_trace.json: raw distributed tracing data for the run
  • execution_path.md: extracted execution-path tree (generated by extract_trace_tree.py)
  • metadata.json: run metadata
  • Task outputs: generated artifacts such as books/chapters, reports, validation results, logs, etc.

Example:

RESULTS/
├── GPT-5/
│   └── BookWriter-MCP/
│       └── test_results/
│           └── run_YYYYMMDD_HHMMSS/
│               ├── langfuse_trace.json
│               ├── execution_path.md
│               ├── metadata.json
│               └── ... (task outputs)

RQ1: trajectory/success/retry evaluation

Location: RQ1/

  • What it contains:

    • per-task evaluation scripts and reference trajectories (YAML)
    • aggregated CSV outputs (e.g., trajectory metrics, success rates, retry summaries)
    • radar charts under RQ1/RadarCharts/
  • Entry points:

    • per-task scripts (run from a specific RQ1/<Task-Variant>/ folder):
      • evaluate_trajectory.py
      • evaluate_success.py
      • analyze_retry_patterns.py
    • plot generator (run from RQ1/):
      • generate_radar_charts.py

See RQ1/README.md for details.

RQ2: agent-time / LLM-share breakdown

Location: RQ2/

  • What it contains:

    • per-task aggregation outputs (CSV) and mapping files (agent_map.md, FRAMEWORK_map.yaml)
    • consolidated summary tables (e.g., llm_share_summary.csv)
    • paper-ready figures (PDF), including heatmaps and bar charts
  • Typical reproduction flow (from RQ2/):

    • python aggregate_llm_share.py
    • python plot_llm_share_heatmap.py
    • python plot_model_overhead_bars_a2a.py
    • python plot_model_overhead_bars_mix.py
    • python plot_framework_overhead_ratio_mix_vs_a2a.py
    • python plot_agent_time_share_bars_unified_y_by_series.py
    • python plot_total_classified_ecdf.py

See RQ2/README.md for details.

RQ3: performance analysis (time/tokens)

Location: RQ3/

  • What it contains:

    • parsers/aggregators that scan execution_path.md across all runs
    • performance reports (CSV) under RQ3/performance_reports/
    • comparison reports under RQ3/agent_time_reports/
    • violin plots (PDF) under RQ3/Violin/
  • Typical commands (from repository root):

    • python RQ3/analyze_performance.py
    • python RQ3/analyze_agent_time_comparisons.py
    • python RQ3/generate_summary.py
    • python RQ3/plot_total_tokens_violin.py

See RQ3/README.md for details.

Trace extraction: generating execution_path.md

execution_path.md is a key intermediate artifact consumed by RQ analyses.

  • Single file:
python3 extract_trace_tree.py <path/to/langfuse_trace.json>
  • Batch (scan the entire RESULTS/ tree and generate execution_path.md next to each trace):
./batch_extract_trees.sh

See README-extract_trace_tree.md for the output format and filtering rules.

Requirements

  • Trace extraction (extract_trace_tree.py):

    • Python 3
    • standard library only
  • RQ analyses and plotting (RQ1/, RQ2/, RQ3/):

    • Python 3
    • common scientific stack (e.g., numpy, pandas, matplotlib)

Notes

  • Some scripts assume the default directory layout under RESULTS/. If you rename/move folders, update the corresponding path configuration inside the scripts.
  • Trace files (langfuse_trace.json) can be large; many PDFs/CSVs in RQ1/, RQ2/, and RQ3/ are generated analysis artifacts.