Is Direct-Answer Bias a Model Fact or an Attribution Artifact?

Retrieval-augmented generation (RAG) systems first fetch documents and then have a language model answer from them. They are trusted partly because they can say which document an answer came from. Methods that assign credit to documents, from the model's own inline citations to game-theoretic scores computed by deleting documents and re-running the model, are now judged by whether they credit the "right" document. In multi-hop questions, one document supplies a bridge fact and a second contains the answer. The most careful study of these methods found that they credit the answer-bearing document and give almost nothing to the bridge. That result is read as a failure of attribution.

The reading may be wrong. An attribution score describes what this particular model depended on, not what a human reader would need. If the model already knows the bridge fact from training, the bridge document really is unnecessary, and giving it zero credit is correct. There is a second possibility: the bias may come from how credit is computed, because scoring the likelihood of a fixed answer string favors whichever document contains that string. This note proposes testbeds in which the model's true dependence on each document is fixed by construction. With those in hand, direct-answer bias can be split into three parts: model shortcuts through memorized knowledge, artifacts of the scoring function, and approximation error. Each part can then be measured on real multi-hop data.

Core Claim

Explanation quality for RAG attribution cannot be measured against human relevance labels. The ground truth has to be the model's behavior under controlled intervention, and that requires inputs where memorized knowledge cannot be used.

Our specific, falsifiable claim has two parts. First, the direct-answer bias reported by Source Attribution in Retrieval-Augmented Generation will largely disappear on fictional-entity questions where the bridge fact cannot be memorized, provided credit is scored contrastively. Second, the bias that remains on real HotpotQA bridge questions will concentrate on questions whose bridge fact the model can answer without any documents.

If both hold, the field has been penalizing attribution methods for being right. If the bias persists even when the bridge is provably necessary, the additivity critique in that paper is confirmed, and we will have a clean benchmark for interaction-aware methods.

Relation to Prior Work

Two papers motivate the proposal by showing that "correct" and "causal" diverge.

Correctness is not Faithfulness in Retrieval Augmented Generation Attributions injects answer phrases into other documents and shows Command-R+ then cites those documents. False citation rates were 12% for random documents and 57% for relevant but uncited ones. That study is black-box and uses a single model. We reuse its injection logic but build the entire input distribution so that the true dependence is known in advance.

Source Attribution in Retrieval-Augmented Generation compares leave-one-out, Shapley sampling, Kernel SHAP and ContextCite on redundancy, complementarity and synergy cases. It reports Precision@1 generally above 0.8, but finds bridge documents go uncredited. Its credit signal is the teacher-forced log-likelihood of the target answer. It never asks whether the models already knew the bridge facts. Those two features are exactly the confounds we isolate.

Several papers supply methods we reuse.

CEBaB shapes the design in two ways. It shows that explainers should be scored against measured effects of human-built counterfactuals. It also shows, as a caution, that sophisticated concept explainers lost to a naive sampling baseline. We therefore expect simple ablations to be strong, and we include them.

Research Program

Controlled testbeds. We first build fictional-world question sets. Entities, relations and values are generated and renamed, so a model cannot answer without documents and cannot route around a document through memory. Each item instantiates one known dependency structure:

  • a single necessary document;
  • two redundant documents;
  • complementary documents that each supply part of the answer;
  • synergy, where document A names the bridge entity and document B contains the answer.

In the synergy case, B always contains $m$ candidate answers, one per possible bridge entity, so B alone leaves the model guessing. A parametric-route control uses the same templates with famous real entities.

Ground truth is checked behaviorally rather than assumed. An item enters the testbed only if closed-book accuracy is at chance and accuracy with B alone is near $1/m$. This makes "A and B are each necessary and jointly sufficient" a measured fact about the model, which is what the explanation-quality literature means by a ground truth that does not rely on intuition.

Separating the scoring function from the estimator. We compare the standard log-likelihood credit signal with a contrastive one:

$$U(S) = \log p(a \mid q, S) - \max_{a' \in F} \log p(a' \mid q, S)$$

Here $S$ is a subset of documents, $q$ the question, $a$ the gold answer, and $F$ the set of foil candidates present in the documents. Under plain likelihood, removing A may barely move $\log p(a)$, because $a$ still appears in B. Under the contrastive signal, removing A drops the margin by roughly the cost of an $m$-way guess.

Exact Shapley values over a handful of documents are cheap to compute. They remove approximation error entirely, so any residual bias can be attributed to the scoring function or to the model.

Transfer to real data. On HotpotQA bridge questions (HotpotQA) we measure, per question and per model, whether the bridge fact is memorized. We do this by posing the bridge sub-question closed-book. We then stratify direct-answer bias by that variable.

Baselines

Naive reference: self-citation. The model is prompted to cite documents inline. Wallat et al. show this mechanism produces false citations 57% of the time on relevant uncited documents. MIRAGE reports a citation F1 of 30.6 for it on ELI5 with Zephyr-7B. It is run with the same prompts and documents as every other method. Any method should beat it on recovering the constructed ground truth.

Simplest causal method: leave-one-out ablation. It costs only $n$ forward passes. It is the method most likely to be "good enough" in the single-document and synergy cases, and known to fail under redundancy.

Strongest existing methods: Kernel SHAP and ContextCite. We run them as in Source Attribution in RAG, with 100 subset evaluations. That paper reports correlation above 0.95 with exact Shapley and Precision@1 generally above 0.8. Beating these means matching that Precision@1 on non-synergy items while crediting the bridge document in the top two on at least most synergy items.

White-box comparator: MIRAGE. It is run on the same open-weights models with its calibrated threshold. Its reported reference points are 86.7% average human agreement on XOR-AttriQA and citation F1 45.6 on ELI5.

Interaction-aware alternative. We compute pairwise Shapley interaction scores from the same subset evaluations. This is the strongest threat to the claim, so it gets the same evaluation budget as the other methods.

Ablations. Every method is evaluated under both scoring functions and on both fictional and real-entity variants. This isolates the two proposed confounds.

Fairness controls. All methods see identical documents, document orders and models: Llama-3.2, Mistral-7B and Qwen as in the source-attribution study, plus Zephyr-7B for MIRAGE. Methods are matched on number of forward passes. Document order is permuted to control the first-is-best position bias that study found.

Why not just…?

Why not just use HotpotQA's supporting-fact labels as ground truth? Those labels record what a human needs to answer the question, not what a specific model used. The parametric-route control is designed to show the gap directly. If attribution methods "fail" against human labels exactly on items where the model can answer the bridge sub-question closed-book, then human labels are the wrong yardstick.

Why not just compute exact Shapley values and be done? Exact Shapley removes estimator error. It does not remove the effect of scoring answer-string likelihood or of memorized shortcuts, and our design runs it as a baseline for exactly that reason. If exact Shapley under log-likelihood already credits fictional bridges, the scoring-function hypothesis fails, and we will report that.

Why not just add pairwise interaction terms to Kernel SHAP? This is a real competitor. It may be all that is needed once memorized shortcuts are removed. Our contribution does not depend on it losing. The testbed is what makes that question answerable at all, and the experiments compare interaction terms against plain Shapley under both scoring functions.

Why not just verify citations with an entailment model or a GPT-4 judge? Entailment checks whether text supports a claim, not whether the model used it, so it cannot detect memorized shortcuts. RAGTruth reports GPT-4 at only 63.4% response-level and 28.3% span-level hallucination-detection F1. We include an entailment check as a sanity baseline on the testbed, where its blindness to memorized routes should show up on the real-entity control.

Proposed Experiments and Decisive Checks

First check: fictional synergy items. We report how often the bridge document is ranked in the top two by each method under each scoring function. The prediction is that contrastive exact Shapley credits the bridge on most items, while likelihood-based credit underweights it. If likelihood-based credit already succeeds, the scoring-function explanation is refuted.

Second check: the gap between controls. We compare matched fictional and real-entity templates. A larger bias on real entities, tracking closed-book bridge accuracy, attributes the bias to memorized shortcuts.

Third check: HotpotQA. On bridge questions, we regress bridge credit on closed-book bridge knowledge. If bias is flat across that variable, direct-answer bias is a genuine method failure, and interaction-aware methods become the priority.

Secondary check: redundancy items. Redundant items replicate leave-one-out's known collapse and the first-is-best position bias under controlled order permutation. This confirms the testbed reproduces prior findings before we trust its new ones.