Reliability-Gated Cue Following in Large Language Models Consuming External Reasoning Traces

Large language models increasingly consume externally produced reasoning traces, yet the reliability of such traces is not guaranteed. We study how models integrate answer cues embedded in traces under controlled corruption. Across 24 experiments (64,004 trials) with GPT-4o and Claude, we find a reliability-gated shift toward cue following: in verifiable arithmetic, GPT-4o's wrong-cue following rises from ~0.5% to ~9% at high corruption, while Claude remains near zero. This shift is determined by the consumer model, not the trace producer, and is amplified by cue salience (up to 21%). In unverifiable knowledge tasks, both models show higher cue dependence (~14%). Simple verification prompts suppress cue following to 0%. Unlike prior work on self-generated CoT faithfulness, we focus on external trace consumption, revealing reliability as the key gating variable—with direct implications for securing retrieval-augmented and multi-agent systems.