Definition
Plain language
The way each round of LLM summarizing softens contradictions until they disappear.
As stated in the literature
A characterized failure mode in multi-agent communication where iterative natural-language summarization progressively erodes hard pass/fail constraints into hedged language, producing memory-induced confirmation bias.
Why it matters: It explains why multi-agent systems often drift toward optimistic consensus, eroding the hard constraints they were supposed to enforce.
For example, a strict 'tests failed, do not deploy' verdict can become 'some tests had issues, consider proceeding cautiously' after a few rounds of agents summarizing each other.