An ICLR 2026 paper proved that making AI agents reason harder actually makes them hallucinate more, not less. Toggle reasoning on and watch it happen.
Toggle to see what happens when you turn on chain-of-thought / reasoning RL
Base Model · 7B params
Base Model · 8B params
Base Model · 32B params
The layer in the neural network that should restrain bad tool calls is exactly what gets trained away during reasoning RL. The model learns to reason harder, but the guardrails against hallucination collapse in the process.
This isn't a bug in one model. It's a fundamental trade-off in how we train AI agents today. Every major reasoning method — RL, distillation, chain-of-thought — amplifies the same problem.
Enterprise Impact
47%
made decisions on hallucinated content
Agents in Production
96%
of enterprises run AI agents
Worst Case
2.1x
hallucination increase with reasoning
"The Reasoning Trap: How Enhancing LLM Reasoning Amplifies Tool Hallucination"
ICLR 2026 · Rio de Janeiro · SimpleToolHalluBench
Read the full paper →We help teams ship reliable AI systems — with guardrails that actually work.