Knowledge Conflicts in RAG: A Prior Instead of a Fixed Rule Gains Up to +26 Points

When a RAG system pulls a document from its database that contradicts what the model itself knows, something has to decide which to believe. In their paper "Don't Pick a Winner," researchers Minh Pham Van and a co-author argue that any decision baked into the architecture in advance is bound to fail. Their proposed system scored 45.3 on GRACE with Qwen3.5-35B-A3B and 35.0 with Llama-3.1-8B. The best existing method scored 29.4 and 27.5. Across every mix of correct and incorrect contexts, the advantage reached up to 26 points.

Why any fixed rule loses

The core problem is simple. A retrieval-augmented model answers from two sources: the retrieved context and its own parametric knowledge. When the context is incomplete or wrong, the two diverge. Existing approaches settle the conflict with a rule chosen at design time: follow the context blindly, amplify it in proportion to how far it shifts the model's distribution, or trust memory.

The authors build a decision-theoretic model with two sources. In it, the quality of any fixed rule depends linearly on the share of reliable contexts in the stream, while the best achievable quality is convex. It follows that any fixed rule falls short of the optimum everywhere except the single point where it happens to match reality. The "trust the context" rule works well on a clean knowledge base and badly on a noisy one. The "trust memory" rule is the mirror image.

There is a second argument, aimed at adaptive decoding. It measures the symmetric divergence between the distributions with and without the context. That quantity is the same whether the context or the memory is right, so it says nothing about which one is.

Two inputs that no existing method has

The theory implies that resolving a conflict takes two inputs. The first is evidence about each source taken separately. The second is a prior on how often the context turns out to be correct. No existing method stores either.

To test this, the authors used GRACE. The benchmark intervenes on the knowledge-graph context for a fixed question. That lets it measure separately what the model knows, what the context says and what is true. On it, every baseline from the four families is strong either on correct contexts or on incorrect ones, but not on both at once.

The authors' own design has four components:

• the model writes out its knowledge as notes;
• a small policy voices each note with calibrated confidence;
• a reader answers "in two voices," one from the context and one from the notes;
• a selector either commits to an answer or abstains, applying the prior π as an exact prior shift.

Numbers and limits of applicability

When π equals the true share of correct contexts, the system beats the best baseline at every GRACE context mix, by up to 26 points. The main results:

• GRACE, Qwen3.5-35B-A3B: 45.3 vs. 29.4 for the best baseline;
• GRACE, Llama-3.1-8B: 35.0 vs. 27.5;
• ClashEval (preregistered protocol): +12 points over CK-PLUG and +26 over TCR on the same backbone;
• where the model lacks the needed knowledge, the gain disappears.

The authors present the last point as confirmation of the theory, not as a caveat. If nothing is stored in parametric memory, separate evidence about that source is useless, and the method falls back to baseline level. The paper is long: 49 pages, 9 of them main text, plus 14 figures.

What to change in RAG pipelines

For practitioners, the takeaway is this. If your system holds a single constant for "how much to trust the context," that constant is a hidden assumption about index quality. It breaks when the corpus changes, for example when you move from an internal wiki to the open web. At a minimum, it helps to estimate the share of correct contexts on a labeled sample and use it as the prior.

The second lesson concerns calibration. The model should be made to state separately what it knows without the context and to rate its confidence in that knowledge. A divergence between answer distributions alone will not provide this. Third, abstaining offers an honest way out when the sources disagree and the prior does not allow a confident choice.

The limitations are also clear. The method requires an estimate of π and works only where the model genuinely holds the knowledge, and the evaluation so far covers just two models and two benchmarks. Still, the framing itself, which treats knowledge conflict as a Bayesian problem rather than a pick-a-winner choice, will likely influence the design of the next RAG systems.