Finding

Paper: "The Reasoning Trap: How Enhancing LLM Reasoning Amplifies Tool Hallucination"
arXiv: https://arxiv.org/abs/2510.22977

Core Idea

Reinforcement-learning-enhanced reasoning models (CoT, o1-style) are more prone to tool hallucinations than standard models, not less. Hallucination types identified:

• Fabricating tool existence (calling a tool that doesn't exist)
• Invoking tools with incorrect parameters
• Misreading or fabricating tool output

Counterintuitive finding: deeper reasoning increases confidence in hallucinated tool interactions, making the agent harder to correct.

Applicability to Zeph

High (5/5). Zeph uses OpenAI models including o-series and reasoning-capable models. If a planner or orchestrator uses a reasoning model for tool decisions, this amplification effect could manifest as:

• Tool calls with fabricated parameter values
• Tool output interpretation errors escalating through the orchestration DAG
• Self-reinforcing incorrect plans that survive multiple turns

Relates to:

• NabaOS receipts (research(tools): NabaOS tool receipt layer — 94.2% hallucination detection at <15ms overhead (arXiv:2603.10060) #2266) — tool receipt verification would catch these post-hoc
• AgentErrorTaxonomy (research(tools): AgentErrorTaxonomy — 5-domain failure classification + domain-aware recovery reduces cascading errors (arXiv:2509.25370) #2253) — adds a causation pathway
• Tool reliability taxonomy (research: tool invocation reliability taxonomy — 12 categories, model-size threshold for reliable tool use (arXiv:2601.16280) #2234) — adds a model-type dimension

Implementation Sketch

1. Add reasoning-model-specific heuristic in : flag calls where parameter values appear hallucinated (not present in context)
2. Consider using non-reasoning models for tool routing/dispatch while reserving reasoning models for planning-only steps
3. Instrument tool_execution anomaly detector with a 'reasoning_amplification' category