Naming the Model's Own Prior: Parametric Foils for Faithful Retrieval-Augmented Generation

Language models that answer questions from retrieved documents still get things wrong in a specific, stubborn way. When a document says one thing and the model "remembers" another from training, the model often goes with its memory. This happens even when the document is correct, and even when the model has plainly read it. Our proposal is to make the model state its own remembered answer out loud before it answers. It then writes a short argument about why the retrieved passage does or does not justify overriding that memory. In the language of contrastive explanation, which justifies a decision relative to a specific rejected alternative (the "foil"), the model's own prior becomes the foil. This is a checkable foil: we can verify that what the model says it remembers is what it actually answers when given no documents.

The opportunity is that two literatures have each found half of this and neither has joined them. Work on retrieval-augmented generation (RAG, where a model answers conditioned on passages fetched from an external corpus) has shown that contrastive reasoning helps small models ignore distractor passages. Work on knowledge conflict has shown that the main source of unfaithful answers is internal: memory-injecting layers override context the model has already attended to. Nobody has asked the model to reason contrastively against its own memory. We also think this is the version of contrastive explanation whose faithfulness can actually be tested.

Core Claim

A RAG model that first surfaces its closed-book answer $Y$ and then produces a rationale contrasting the retrieved claim $X$ against $Y$ will handle knowledge conflicts better than a model that only reasons about passages. The rationale ends in one of three verdicts: the passage supports $X$ over $Y$, the passage does not address the question, or the passage is internally inconsistent. The key requirement is performance across a mixture of conflict types. On counterfactual or updated context it must follow the passage. On irrelevant or distracting context it must fall back on memory or abstain. On unanswerable context it must say so. The approach should beat both "always trust context" and "always trust memory," and it should beat simply detecting the disagreement without explaining it.

We add this requirement because the naive version of the idea is too easy to match. On a purely counterfactual benchmark, a two-line rule ("if the open-book and closed-book answers differ, output the open-book one") would do as well as any rationale. The contrastive explanation earns its place only if weighing $X$ against $Y$ helps the model decide when the disagreement should be resolved in favor of the passage. The experimental design is built around that question.

Relation to Prior Work

The closest work is Eliciting Critical Reasoning in Retrieval-Augmented Generation via Contrastive Explanations. Its C-RAG pipeline has the model argue why one retrieved passage is relevant and another is a distractor. It distills 2,000 GPT-4 demonstrations into Llama-2 students and reports beating Self-RAG by 7.2% and standard RAG by 55.4% on average across open-domain QA. C-RAG supplies our training recipe and our main contrastive competitor. Its foils are always external documents, so it has no way to represent the case where the passage is fine and the model's memory is the problem.

Distinguish Before Answer takes a similar "distinguish first, then answer" approach for multiple-choice questions, with foils given by the answer options. It shows that discriminative text helps, but its foils are handed to it, whereas ours are elicited from the model. CoRE constrains our design: it found that showing a model failed reasoning paths alone hurts, and that negatives help only when paired with positives. We read this as a warning that surfacing the parametric answer $Y$ without an explicit contrast could anchor the model on $Y$. We test for this directly.

Three mechanistic papers motivate the proposal. ReDeEP finds that in 65–77% of RAG hallucinations the attention heads that copy from context did look at the right tokens, but memory-injecting feed-forward layers overrode them. Probing Latent Knowledge Conflict (CLEAR) shows that conflict is linearly detectable in final hidden states and fixes it with attention-guided fine-tuning. ParamMute dampens the offending feed-forward layers and shows that larger models are more stubborn. All three treat the parametric belief as noise to suppress. None of them externalizes it, and ParamMute notes the risk of over-suppressing when the context is itself wrong. We treat the belief as content the model should name and argue against.

FaithEval supplies both the evaluation and the key negative results. Chain-of-thought is "a Band-Aid," GPT-4o falls to 47.5% on counterfactual contexts, and conflict-sensitive prompting causes over-abstention (Phi-3-medium falls to 7.4% on unanswerable contexts) and invented conflicts. That last finding is exactly what our three-way verdict must avoid.

RAG-HAT uses contrast only offline, through preference pairs. Diagnosing AI Explanation Methods with Folk Concepts of Behavior argues that explanations lacking a contrast case leave users with broken mental models of the system. This is our reason to expect the stated foil to matter for human oversight, and not only for accuracy.

Research Program

The method has three steps. First, the model answers the question closed-book with greedy decoding, producing $Y$. It also records a confidence, the average token probability of $Y$. Second, given the retrieved passages, the question, and $Y$, the model writes a rationale. The rationale quotes the passage span that bears on the question, says whether that span asserts something incompatible with $Y$, and weighs the two. When the passage directly and specifically asserts $X$, it adopts $X$. When the passage is off-topic or merely co-mentions entities, it keeps $Y$ or abstains. When the passage contradicts itself, it flags the inconsistency. Third, it answers conditioned on the rationale.

We build a prompting-only version and a distilled version. The distilled version matches C-RAG's budget of 2,000 teacher demonstrations. There is one important twist: each demonstration is built around the student's own closed-book answer, not the teacher's. The teacher therefore learns to argue against the specific prior the student actually holds. Training examples are balanced across four context types: supporting, counterfactual, distracting, and unanswerable. This prevents the student from learning "always side with the passage."

The design has an unusual property: the explanation's key claim can be checked. We can measure how often the foil stated in the rationale matches the model's independent closed-book answer. We can also test causally whether the foil drives behavior, by swapping in a foil from a different question and seeing whether the verdict changes. As an analysis layer on Llama models, we use ReDeEP's parametric knowledge score to ask whether surfacing $Y$ reduces the feed-forward override on conflict items. A reduction would give a mechanistic account of why naming the prior helps.

We run the method across model scales, using the Qwen2.5 family from roughly 1B to 14B as in ParamMute, plus Llama-3.1-8B. This directly tests the tension in the file. Small models are thought to fail on distractors and large models on entrenched priors. We predict gains on counterfactual items grow with scale, because larger models hold more confident, more specific foils to argue against. We predict gains on distractor items are largest for small models.

Baselines

The do-nothing reference is standard RAG: the same retriever, the same passages, and the question answered directly. Alongside it we run the instruction-only reference, which adds "answer only from the passage; say unanswerable if it does not contain the answer." We also run chain-of-thought under the same instruction. FaithEval reports that these do not close the gap, and they calibrate how much of any gain is simply "more tokens of reasoning."

The strongest contrastive competitor is C-RAG. We reproduce it with the same student models, the same 2,000-demonstration budget, the same teacher, and the same balanced mixture of context types, so the only difference is whether the foil is a distractor passage or the model's own prior. The file reports C-RAG's gains on PopQA, NQ, TriviaQA, and FEVER, but gives no reported result on Lacuna for C-RAG on FaithEval, ConFiQA, or CoFaithfulQA, so we measure it ourselves.

The strongest knowledge-conflict method is CLEAR. It reports a FaithEval F1 of 74.4% with LLaMA-3.1-8B, against 71.6% for the best baseline it compared to (Probing Latent Knowledge Conflict). We run it on LLaMA-3.1-8B with its published fine-tuning setup. For reference, FaithEval reports 76.8% overall normalized accuracy for Claude 3.5 Sonnet. ParamMute reports +5% average context recall and −4% memory recall over standard RAG on CoFaithfulQA (ParamMute), and we run it on the same Qwen and Llama checkpoints. Beating CLEAR's reported 74.4 F1 on FaithEval with an 8B model, while also not losing to standard RAG on the distractor-noise setting, would count as a clear win. Matching it with a prompting-only or 2,000-example method that also produces a checkable rationale would count as a useful result.

The disagreement-rule baseline is the one that most directly threatens the core claim. It runs closed-book and open-book generation. If the answers agree, it outputs them. If they disagree, a small classifier trained on the same 2,000 examples chooses among the passage answer, the memory answer, and abstention, using retrieval score and closed-book confidence as features. A second non-verbal baseline is decoding that subtracts closed-book next-token probabilities from open-book ones, in the spirit of the contrastive decoding used by RCD. It is tuned on the same development split.

Three ablations isolate the key ingredient. The stated-only ablation shows $Y$ but omits the contrast step; per CoRE, this may hurt. The shuffled-foil ablation replaces $Y$ with another question's closed-book answer. The cross-model-foil ablation uses the closed-book answer from a different model. If the shuffled or cross-model foils perform as well as the true foil, the benefit comes from the format, not from naming the model's own prior.

Why not just…?

Why not just detect the disagreement and trust the passage? This is the real threat. On purely counterfactual sets it will likely tie with our method, and we say so. It should fail on the mixture, because disagreement alone cannot tell a correct update from a distractor that happens to mention a different entity. CLEAR found that irrelevant context is amplified when it matches the model's beliefs, and C-RAG shows that small models follow distractors. The disagreement-rule baseline with a learned classifier settles this. If it matches us on the balanced mixture, the rationale is decoration, and we would report the method as an explanation interface only, not as a faithfulness gain.

Why not just suppress parametric memory internally, as ParamMute or CLEAR do? These methods improve faithfulness to the passage, but they commit to the passage by design. ParamMute itself flags over-suppression when context is ambiguous or incomplete, and neither method tells a reviewer what belief was overridden. Our distractor and unanswerable splits test the first weakness. The foil-agreement and human-audit measures test the second. If CLEAR beats us on all four context types, the explanation must justify itself on auditability alone.

Why not just use chain-of-thought with a "rely on context" instruction? FaithEval tested exactly this. It found that chain-of-thought helps but leaves large gaps, that GPT-4o drops to 47.5% on counterfactual contexts, and that conflict-seeking prompts cause invented conflicts and over-abstention. We include this baseline to confirm those findings on our models and to show that the gain does not come from reasoning length.

Why not just contrast passages against each other, as C-RAG does? When there is only one relevant passage and it conflicts with memory, passage-versus-passage contrast has nothing to argue against. The matched-budget C-RAG comparison on counterfactual items, where every retrieved passage agrees with the other passages but not with memory, isolates exactly this gap.

Proposed Experiments

The main evaluation uses FaithEval's unanswerable, inconsistent, and counterfactual subsets, ConFiQA for single- and multi-conflict items, and CoFaithfulQA for context recall and memory recall. We add a noise condition built as in C-RAG: PopQA and NQ with injected distractor passages and shuffled passage order. We report results per context type and on a macro-average across types. We also report the abstention rate on answerable items, since that is the failure FaithEval warns about.

To check that gains carry over to long-form generation, we use RAGTruth's question-answering subset. There we measure the span-level hallucination rate with its human labels and check whether hallucination still rises toward the end of long answers.

Explanation quality gets three measures. The first is foil agreement: how often the stated $Y$ equals the independent closed-book answer. The second is causal sensitivity: how often verdicts flip under foil swaps. The third is a small audit study modeled on the clinician study in Trustworthy and Practical AI for Healthcare. Participants see answers either with or without the parametric-foil rationale and must flag cases where the model wrongly trusted or wrongly ignored the passage.

The decisive checks are these. First, on the balanced mixture at 8B, the distilled method must beat the disagreement-rule baseline, C-RAG, and CLEAR on macro-average. Second, the shuffled-foil ablation must lose most of the gain. Third, foil agreement must be high, which we would take as above 90% on greedy closed-book answers. Fourth, gains on counterfactual items should increase with model size across the Qwen2.5 family.

If the first check fails but the others hold, the contribution narrows to a faithful, checkable explanation of knowledge conflicts, with accuracy no better than existing methods. If the second check fails, the idea is wrong and the benefit came from the prompt format.