ICLR 2026: 'The Reasoning Trap' — Reasoning RL Training Amplifies Tool Hallucination in Lockstep with Task Gains

ICLR 2026 / Asanify News·high signal

Presented at ICLR 2026 in Rio de Janeiro, this paper identifies a fundamental reliability-capability trade-off: training models for stronger reasoning through reinforcement learning disproportionately collapses tool-reliability representations in late network layers, increasing tool-hallucination rates in direct proportion to task performance gains. The authors built SimpleToolHalluBench as a diagnostic benchmark and found that neither prompt engineering nor DPO closes the reliability gap. For builders running agents with tool use, this means your reasoning-optimized model may be systematically worse at knowing when NOT to call a tool.