NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·May 21·5 min read
To Call or Not to Call: Diagnosing Intrinsic Over-Calling Bias in LLM Agents
TL;DR
LLM agents often over-call tools when they are not necessary, leading to inconsistent performance. The research utilized the When2Call benchmark to evaluate six models from three families, revealing a significant bias towards making calls.
✦ Why It Matters
Engineers can refine LLM decision-making processes to minimize unnecessary tool calls, enhancing overall model performance.
Key Takeaways
- Evaluate LLMs using the When2Call benchmark.
- Investigate intrinsic biases in decision-making processes.
- Adjust models to reduce unnecessary tool calls.
- Aim for higher overall accuracy in LLM applications.
Full Summary
Large Language Model (LLM) agents have been observed to frequently invoke tools even when unnecessary, which can hinder their effectiveness. To investigate this issue, researchers employed the When2Call benchmark, assessing six different models from three distinct families.
They discovered that these models exhibited high accuracy in making calls but significantly lower accuracy in deciding when not to call, leading to an overall accuracy range of 55% to 70%. The study proposes an Intrinsic Bias Hypothesis (IBH), suggesting that the decision-making process for calls is influenced by an activation-independent call offset, causing a preference for calling.
This finding indicates that LLMs may require adjustments to their decision-making frameworks to reduce unnecessary tool invocation. The implications of this research are critical for improving the reliability and efficiency of LLM agents in practical applications.