Summary: Patil et al. (2023) — Gorilla: LLM Connected with Massive APIs¶
Full citation: Patil, S.G., Zhang, T., Wang, X., and Gonzalez, J.E. (2023). Gorilla: Large Language Model Connected with Massive APIs. arXiv:2305.15334.
Key Contribution¶
This UC Berkeley paper introduces APIBench, the first large-scale benchmark for evaluating LLM tool-calling accuracy across 1,645 real APIs from HuggingFace, TorchHub, and TensorFlow Hub. It documents severe API hallucination rates across frontier models and demonstrates that fine-tuning (Gorilla) and retrieval-augmented approaches can improve tool-calling reliability. The paper provides the foundational hallucination rate figures cited throughout the tool-calling wiki page.
APIBench and Evaluation¶
The benchmark comprises 1,645 APIs with 16,450 instruction-API pairs generated via self-instruct. Evaluation uses AST sub-tree matching: the generated API call is parsed into an abstract syntax tree, and functional correctness is determined by whether the generated call's AST is a sub-tree of the reference API's AST. This captures functional equivalence (correct API with valid parameters) while allowing syntactic variation.
API Hallucination Rates (Zero-Shot, 2023)¶
| Model | TorchHub | HuggingFace | TensorFlow Hub |
|---|---|---|---|
| GPT-4 | 36.6% | 43.1% | 78.7% |
| GPT-3.5 | 18.8% | 35.7% | 47.9% |
| Claude | 66.0% | 72.7% | 88.5% |
| LLaMA | 93.1% | 71.7% | 83.8% |
Hallucination types include: inventing nonexistent API endpoints, fabricating repository names (e.g., "AutoModel.from_pretrained('dir_name')" with arbitrary names), and generating syntactically valid but semantically wrong calls.
Key Findings¶
- Retrieval can help or hurt: Adding a retriever (BM25, GPT-Index) sometimes reduces hallucination but can also increase errors. Non-optimal retrievers misguide the model — accuracy dropped 21.5% on TorchHub and 47.6% on HuggingFace with sub-optimal retrieval compared to no retrieval.
- RLHF reduces tool hallucination: GPT-3.5 (RLHF-tuned) has fewer hallucination errors than GPT-4 on most benchmarks, suggesting alignment training helps tool-calling reliability.
- Fine-tuning works: Gorilla (fine-tuned LLaMA-7B) surpasses GPT-4 on API accuracy with retrieval, demonstrating that domain-specific fine-tuning can dramatically improve tool-calling.
- Adapts to documentation changes: Retrieval-aware training enables the model to adapt to test-time API documentation changes — a critical property for production systems where APIs evolve.
- Constraint handling: Models struggle with API calls that have constraints (e.g., "accuracy >= 80%, parameters < 10M"). Gorilla handles constraints better than GPT-4 in zero-shot settings.
Relevance to This Wiki¶
This paper provides the baseline empirical data for tool-calling reliability. The hallucination rates (36-93% depending on model and API domain) establish that tool calling is not a solved problem — it is a probabilistic process with high failure rates, especially for less common APIs. The finding that retrieval can worsen performance connects to the over-retrieval problem in retrieval-augmented-generation. For safety-critical applications where tool calls query plant data systems, invoke simulations, or retrieve regulatory text, these hallucination rates translate directly into operational risk.