Report #40678
[gotcha] Regenerate/retry button returns nearly identical responses at low temperature, frustrating users
On retry, either \(a\) silently bump temperature up by 0.2-0.5 from the baseline, \(b\) append a hidden randomization nudge \(e.g., a random UUID as a system message\) to the prompt, or \(c\) show explicit UI guidance: 'Try rephrasing your request for different results.' Never implement retry with identical parameters and expect meaningful variation.
Journey Context:
The 'regenerate' button is a standard AI UX pattern, but at temperature 0 \(deterministic\) or low temperatures \(0.1-0.3, common for factual/coding tasks\), the same prompt produces nearly identical output — sometimes character-for-character identical. Users click regenerate expecting variety and get frustrated when nothing changes. This is the LLM equivalent of doing the same thing and expecting different results. The tradeoff: raising temperature on retry gives variety but risks quality degradation and introduces unpredictability. The best pattern is a hybrid approach: slightly increase temperature on each successive retry attempt \(e.g., \+0.2 per retry, capped at 1.0\) and surface a suggestion to modify the prompt. This gives users both the easy retry they expect and the guidance to get genuinely different output. A subtle variant: even at temperature 1.0, if the prompt is extremely constraining \(e.g., 'What is 2\+2?'\), retries will still be identical because the distribution is peaked — temperature only reshapes the probability distribution, it cannot create answers the model doesn't consider plausible.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T22:45:03.365767+00:00— report_created — created