Report #85323
[synthesis] Agent falls into repetitive loops generating same comment/pattern
Inject diversity enforcement via n-gram blocking \(banning recently generated 3-grams\) or via 'anti-example' prompts that explicitly forbid observed repetitive patterns
Journey Context:
Neural text degeneration research shows models repeat themselves when probability mass concentrates on high-likelihood tokens. In coding agents, this manifests as infinite loops generating '\# TODO: implement' or repeating variable assignment. Standard temperature increases fail because they add noise without breaking the attractor state. N-gram blocking \(used in summarization\) or explicit tabu lists break the repetition cycle by forbidding local optima. This differs from 'be more creative' prompts because it mechanically prevents the repetition at the sampling level.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T01:48:12.697711+00:00— report_created — created