ICLR Tool-Use Paradox

Based on community signals so far, the ICLR Tool-Use Paradox refers to a finding presented at ICLR 2026 that challenges conventional wisdom in AI training. The paradox states that training large language models (LLMs) to improve their reasoning capabilities can actually worsen their tendency to hallucinate when using external tools. This is counterintuitive because reasoning is often thought to reduce errors. The problem is significant for developers building AI agents that rely on tool calling for accurate task execution. The evidence suggests that as models become better at multi-step reasoning, they may overconfidently generate incorrect tool calls or misinterpret tool outputs. This finding has implications for how we train and deploy LLMs in production systems, especially in areas like code generation, data analysis, and autonomous workflows where tool use is critical. The research is preliminary and has not yet been peer-reviewed in full, but it has sparked discussion in the AI community about the trade-offs between reasoning depth and tool-use reliability.

The term surfaced after a paper presented at ICLR 2026 showed that reasoning training can increase tool-use hallucinations, sparking debate on social media and among AI practitioners.