# Why LLMs ignore their own internal verification to trust > Source: > Published: 2026-09-07 08:48:02+00:00 # Why LLMs ignore their own internal verification to trust [RAG](/en/tags/rag/)and tool-augmented LLMs isn't just retrieval quality, but how the model actually integrates that evidence into a decision it has already started to form. According to the recent paper arXiv:2609.04290v1, LLMs don't just "update" their knowledge; they follow a distributional theory where evidence shifts the probability of their initial answers. This creates a weird paradox where the model might internally verify that a piece of evidence is wrong, yet still use it to generate the final answer. ## The mechanics of evidence integration The researchers ran over ten million trials across twelve different LLMs from four model families. They focused on eight domains, including high-complexity scientific tasks like quantum mechanics, genetics, and molecular biology. The findings suggest that integration isn't about a "trust score" for the source, but rather how the evidence aligns with the receiver's prior distribution. A few specific observations stood out: - **The Persuasion Gap:** Candidates that are already more probable to the receiver are significantly more persuasive. - **Error Bias:** LLMs integrate "characteristic errors" (errors they would likely make themselves) much more readily than "foreign errors" coming from a source with a different error profile. - **The Strength Paradox:** Identical evidence can actually improve a weaker model while actively harming a stronger one. ## The disconnect between verification and output The most jarring finding is the failure of internal verification. In tests using propositional constraints and held-out physical and life-sciences reasoning, LLMs integrated candidates even after they had internally verified them as invalid. The integration rate was staggering, hitting 93-100% in some constraints and up to 99.4% in scientific reasoning tasks. Essentially, the model's "verification" process is just a verbalized state that is almost entirely dissociated from the actual decision-making process. Using a J-lens decomposition, the authors showed that the representation of verification is decodable (meaning we can see the model "knows" it's wrong), but it has very little causal impact on the final answer. ## How the network actually handles the data If you're looking at this from a deployment or AI workflow perspective, it's important to realize that candidate integration happens very late in the network. The process follows a structured sequence: 1. The network admits external candidate answers. 2. It promotes these candidates. 3. It transports them into the final answer state. Because this happens so late, the "reasoning" or "verification" that happened in earlier layers gets bypassed. This is why a deep dive into prompt engineering often fails to fix hallucinations in RAG—the model isn't failing to verify the evidence; it's failing to let that verification override the integration policy. For those building LLM agents, this suggests that simply asking a model to "verify the source before answering" might be a superficial fix. The actual integration is a receiver-specific control policy driven by the model's own internal distributions, not a logical check of the evidence's validity. [Next Can we actually trust dependency-distance rankings across →](/en/news/8966/)