ost agent evaluations operate under a naive assumption: an AI system should always generate an answer. In real-world enterprise deployments, knowing when to keep quiet is just as crucial as knowing when to speak. Duke University researchers Shreya Mendi and Brinnae Bent from the Pratt School of Engineering introduced the Why2Speak framework to investigate whether an agent's rationales faithfully reflect its decision to act or abstain. Unlike standard question-answering benchmarks, real-time interventions in multi-party dialogues force models to navigate asymmetric error costs and massive class imbalances.
Evaluating the Qwen3-8B hybrid model across direct policies, explicit chain-of-thought (CoT) reasoning, supervised fine-tuning, and reinforcement learning revealed an unavoidable capability-auditability tradeoff. Direct decision policies yielded the highest raw decision quality, yet they left zero visible trace for safety audits. Conversely, forcing the model to emit explicit CoT explanations generated an audit trail at the expense of decision performance, sharply reducing recall when valid intervention opportunities appeared. Supervised fine-tuning failed to bridge this divide—either suppressing reasoning traces or stalling performance—while reinforcement learning faltered because group-relative objectives generated zero gradient signal on confidently incorrect prompts where rollouts reached unanimous consensus.
Auditing these reasoning traces with activation probes and behavioral ablations proved that standard faithfulness metrics routinely mislead ML teams. Probability scores saturate during high-confidence failures, and activation probes remain susceptible to textual leakage and class skews. Forcing an AI assistant to articulate its inner monologue fundamentally shifts its underlying policy rather than merely opening a window into it. Enterprise architects cannot rely on surface-level reasoning traces to certify that an agent's silence is safe.