A retrieval-augmented system is supposed to widen the model’s view of the world. This paper shows the opposite can happen when the retrieved corpus includes documents the model itself produced: the model starts circling its own language, and the response set narrows until the system is effectively talking to itself.
That is the paper’s main contribution. It is not arguing that retrieval is broken in general. It is isolating a specific failure mode: once self-authored references enter the retrieval pool, citation behavior can become self-reinforcing enough to lock the system into a collapsed pattern.
The scale matters because this is not a toy example. The authors run 1,528 simulations across 1,019 information-seeking prompts.
What the paper shows: collapse is the dominant outcome under self-authored retrieval
The headline result is blunt: 79.6% of simulations end in collapse, or 1,216 out of 1,528.
That collapse is not defined as a generic error rate. The authors use it to describe a feedback failure in iterative response-and-retrieval, where the model keeps overusing its own prior content until the retrieved set and the final answers converge on the same narrow material.
Their example makes the shape of the failure easy to see. After five rounds of recursively replacing references, the streamer visibility plot becomes binary: every streamer is at either 0% or 100% visibility. The responses then stay collapsed for the rest of the simulation.
How the experiments work: three simulation variants, one feedback loop
The setup is designed to isolate the effect of self-authored references. In each round, an LLM generates several responses to a question based on references. The researchers then take some of those responses, remove citations such as `[3]`, and turn them into online articles that can be retrieved later.
They test three variants. In Replace All, every reference is replaced with a self-authored reference from the previous round. In Replace One, one original reference is swapped out each round for a randomly selected self-authored reference. In Search, the reference pool starts with the original set and then accumulates self-authored references one by one.
The experiments use OpenAI, Gemini, and Anthropic APIs, with temperature left at the default value of 1.0. Across the full study, that adds up to over one million LLM API calls.
This matters because the paper is not just showing that bad retrieval hurts. It is showing that once the system can retrieve its own earlier output, the retrieval process itself becomes part of the failure.
Why a single self-authored reference can be enough
The most interesting claim in the paper is that collapse does not need a long chain of feedback to start. The authors say a single self-authored reference can trigger it, because the LLM disproportionately cites its own content.
They also say that this self-bias persists even after controlling for reference quality. In other words, the model is not just preferring better documents. It is overweighting documents because they came from itself.
The paper’s explanation is simple but important: once one self-authored item gets into the candidate set, it can become a magnet for later citations. That creates a path dependence that is hard to unwind, even when the rest of the retrieval pool is still intact.
One of the paper’s internal checks supports that reading. Only about 3% of entity questions are collapsed at the start, before any self-authored references are added. That makes the failure look less like a baseline weakness of the prompts and more like an effect introduced by the self-generated material.
What this means for RAG pipelines in practice
The practical risk here is not just poor retrieval quality. It is retrieval feedback, where the system’s own generations become part of the corpus it later treats as evidence.
That has a direct implication for any RAG stack that ingests generated summaries, answer pages, synthesized articles, or other model-authored artifacts back into search. If those documents are eligible for retrieval without strong filtering, you are creating the conditions for self-reinforcing citation bias.
The authors are careful not to overclaim. They say they do not know precisely how commercial AI systems work, so they cannot tell whether mitigations are already in place, and they do not claim to prove that RAG collapse is already happening in deployed systems. They also do not test every plausible variant, including iterative addition of other AI-generated content or other prompt types.
That uncertainty cuts both ways. It means you should not assume the exact failure rate transfers to production. It also means you should not assume production systems are safe just because the paper is a simulation study.
What to do about it: keep the model out of its own retrieval corpus
The conservative response is straightforward: do not feed model outputs back into retrieval by default.
If you have to reuse generated material, treat it as a special class of content. De-duplicate aggressively. Track provenance. Separate human-authored sources from model-authored ones. And make sure retrieval rankers can downweight or exclude self-generated documents before they start shaping the answer space.
The bigger lesson is architectural. A RAG pipeline is not just a question-answering system with search attached. It is a feedback system, and feedback systems can drift toward their own attractors very quickly. If you are using generated content to expand the retrieval pool, you need a hard reason for doing it and a hard test showing that it does not collapse the system into self-citation.
In controlled simulations with 1,019 prompts, self-authored references repeatedly enter the retrieval pool and the model’s citation behavior locks onto that material.
If your RAG pipeline can return documents that originated from the model’s own prior generations, expect a high likelihood of response “collapse” unless you actively prevent self-feedback loops.
Gregory Druck, Ethan Smith (2026). RAG Collapse: LLM Responses Collapse When Retrieved Documents Are Self-Authored. arXiv:2608.22118