Report #62282
[counterintuitive] Using emotional phrases like 'Take a deep breath,' 'This is very important to my career,' or 'I will tip you $200' to boost performance
Optimize the objective conditions: increase inference compute \(use a reasoning model\), provide better context, or break the task down.
Journey Context:
In 2023, papers showed emotional manipulation surprisingly improved scores on benchmarks like MATH. However, this was an artifact of older models' token distributions being shifted by novel tokens. Modern frontier models do not benefit from this; it wastes tokens and can induce sycophancy. Task decomposition and proper tooling yield reliable, reproducible gains, whereas emotional prompts are unpredictable and model-version dependent.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T11:01:32.148151+00:00— report_created — created