Language models often answer with whatever word most often appears near the question's words in their training text, even when that word is wrong. Asked for the capital of Canada, a model says "Toronto" because "Canada" and "Toronto" co-occur far more often than "Canada" and "Ottawa." Retrieval-augmented generation is the practice of fetching documents and placing them in the prompt so the model answers from evidence instead of memory. It is supposed to fix this, but in specialized fields like medicine it can make the problem worse. The retriever is also driven by word overlap, so it returns passages where the common-but-wrong drug sits next to the common disease. Step-by-step reasoning, known as chain-of-thought prompting, then builds a fluent argument from that skewed evidence. We propose to measure, for every candidate answer, how often it co-occurs with the question's key terms in the domain corpus. Items where the model's answer matches the most-frequent candidate get flagged as likely "parroting." For those items only, the system retrieves evidence aimed at the less frequent alternatives and requires each candidate to be supported or ruled out by a passage that discriminates it from its rivals.
The bet is that a co-occurrence count, cheap to compute and external to the model, predicts confident errors. Those are exactly the errors that uncertainty-based retrieval triggers cannot see, because co-occurrence bias makes models wrong and sure. If the bet is right, the result is a targeted fix for the failure mode where retrieval and reasoning compound each other. If it is wrong, the stratified evaluation alone tells the field whether co-occurrence bias survives retrieval in expert domains at all. No study in the file measures that.
Core Claim
Three claims follow, each falsifiable. First, on expert multiple-choice benchmarks, a large share of the errors made by retrieval-augmented chain-of-thought systems select the distractor with the highest corpus co-occurrence with the question's entities. The retrieved context is skewed toward that distractor too. Second, many of these errors come with low answer entropy, so triggers based on model uncertainty miss them. Third, flagging items by co-occurrence and then retrieving contrastively across candidates fixes more of these errors than any of the following, at matched retrieval budget: retrieving per option without the prior, triggering on uncertainty, or sampling more reasoning chains.
What the Literature Establishes
The motivating evidence comes from closed-book probing. Impact of Co-occurrence on Factual Knowledge of Large Language Models counted subject–object co-occurrence in The Pile and reported three findings. A joint-probability baseline guesses roughly 60% of LAMA facts. GPT-J 6B fails 95% of the time on pairs that never co-occur. About 38% of its errors are a higher-co-occurrence competitor overriding the truth. Scale did not remove the effect, and debiased fine-tuning memorized rare training facts without generalizing to unseen ones. That paper never adds retrieval. We take its measurement idea, moved from the pretraining corpus to the domain retrieval corpus, as our core signal.
Two papers explain why adding reasoning on top of retrieval can hurt. Thinking in a Crowd shows Qwen3-8B with thinking enabled drops from 96.22% to 77.03% on SciAux when given two misleading snippets, which is barely above thinking off (74.87%). Unveiling Confirmation Bias in Chain-of-Thought Reasoning shows that when the direct answer has low entropy, rationales become short justifications of the prior. Neither paper proposes a mitigation, and neither asks where the misleading context comes from. Our hypothesis is that in natural retrieval, co-occurrence skew is a main source of it.
The closest prior methods supply contrasts rather than techniques to reuse. CFD-Prompting treats reasoning as a mediator in a causal model confounded by parametric bias and swaps entities in retrieved passages to build counterfactual contexts. It does this on general multi-hop QA, reporting +2.71 EM and +4.54 F1 on LLaMA-3. It never measures corpus co-occurrence, never targets retrieval at the disfavored candidates, and never tests expert domains. MedCoT-RAG scores documents by causal keywords such as "leads to" and runs a four-stage clinical reasoning template. Keyword causality does not protect against two terms that are simply frequently mentioned together, and the method never checks whether its errors follow co-occurrence. CausalRAG indexes LLM-extracted causal graphs. Its zero-shot graph extraction is exactly the step expected to degrade where the model lacks domain knowledge. CoT-RAG gets strong vertical-domain results from expert-authored decision trees, which our method deliberately avoids requiring.
Two results constrain the design. TheoremQA found that pasting theorem definitions into the prompt barely helps, so our evidence must be discriminating between candidates, not merely definitional. Bias Evaluation and Mitigation in Retrieval-Augmented Medical QA shows that surface features of the query reshape what gets retrieved, so the retrieval set itself must be audited.
Research Program
The system has three parts.
The first is a co-occurrence prior. For each question we extract key entities and, for each answer option, count sentence- and passage-level co-occurrence with those entities in the retrieval corpus. Where available we also count in the CMeKG graph used by KMQA. We normalize by option frequency, following the joint-probability baseline of the co-occurrence paper. We then validate that this domain-corpus count is a reasonable proxy for pretraining exposure: on LAMA-TREx we correlate it with the published Pile-based statistics, and we check that it reproduces the roughly 60% guessability figure.
The second is the trigger. An item is flagged when the model's closed-book or vanilla-retrieval answer equals the co-occurrence-maximal option and the margin between that option and the runner-up is large. Unflagged items are answered by the base pipeline, which keeps cost bounded.
The third handles flagged items. The system issues one query per competing option, pairing that option with the question's entities, and gives each option the same passage budget so the skewed retriever cannot drown the rare candidate. Each passage then goes through a contrastive admissibility check. The model is asked whether the passage favors option A over option B. The same question is asked again with the option names swapped in the passage text, borrowing the entity-swap device of CFD-Prompting but applying it to evidence rather than to the final answer. A passage whose verdict does not change under the swap is treated as co-mention noise and dropped. The final reasoning step works by elimination: every option must be supported or refuted by an admissible passage, cited by span, and an option with no admissible support cannot win.
The analysis program matters as much as the method. We will measure three things. The prior-following error rate is the share of errors that pick the co-occurrence-maximal option, a retrieval-era analogue of the 38% figure. Retrieval skew is the ratio of retrieved passages mentioning the prior-favored option to those mentioning the gold option. The entropy profile of prior-following errors tests whether uncertainty triggers can see them.
Baselines
The naive reference points are closed-book chain-of-thought and vanilla dense retrieval followed by chain-of-thought, run with the same generator, corpus, and passage count. For vanilla retrieval, MedCoT-RAG reports 56.5% accuracy on MedQA-US.
The strongest existing domain method is MedCoT-RAG itself, which reports 70.1% on MedQA-US, 74.4% on MMLU-Med, and 73.5% on BioASQ-Y/N. Domain-adapted RAG2 reaches 64.3%, 74.3%, and 72.3% on the same three. Because the file does not tie those numbers to a generator we will use, we will re-implement MedCoT-RAG's causal relevance scoring and four-stage template on our generators, with the same corpus and the same total retrieved-passage budget. We will report both its published numbers and our reproduction. Beating it means a statistically significant gain on the prior-inverted subset (defined below) with no loss on prior-aligned items, at equal or lower average retrieval calls.
CFD-Prompting is the strongest bias-aware method. It is the closest competitor because it already targets parametric confounding through retrieved-knowledge counterfactuals. We will port it to the medical benchmarks with matched sampling budgets. No reported result was found on Lacuna for CFD-Prompting on any medical or expert benchmark.
Self-consistency with majority voting will be run with as many samples as our method uses in model calls. The file reports DeepSeek-R1-70B with majority vote at 34% on MedQA-US in the bias-evaluation setting of Bias Evaluation and Mitigation. That setting is not directly comparable, so we will rerun it.
Three ablations isolate the key ingredient, each holding the rest of the pipeline fixed. The first replaces the co-occurrence trigger with an entropy trigger at the same flag rate. The second applies per-option balanced retrieval to every item with no trigger and no prior. The third removes the swap-based admissibility check. If the per-option ablation matches the full method, the co-occurrence signal contributes nothing beyond cost savings, and we will report that plainly.
Why not just…?
Why not just use a stronger reasoning model? Ice Cream Doesn't Cause Drowning reports GPT-o4-mini at only 40.72% reliability on statistical causal pitfalls, under 30% on the hardest items. Thinking in a Crowd shows thinking mode amplifying misleading context. We will nonetheless include DeepSeek-R1-70B as a generator. If its prior-following error rate is near zero, the problem is solved by scale for that model class, and the contribution narrows to smaller models.
Why not just retrieve evidence for every option? This is the most serious threat. Balanced per-option retrieval alone may remove most retrieval skew. Our ablation runs it on all items at matched passage budget. The proposal survives only if the co-occurrence trigger plus the swap check beats it on the prior-inverted subset, or matches it at a fraction of the cost. A cost-only win is a weaker but still useful result, and we will frame it that way if that is what we find.
Why not just sample many chains and vote? The CFD-Prompting authors argue, and the confirmation-bias study supports, that a shared prior makes samples agree on the same wrong answer. Our entropy analysis directly tests whether prior-following errors are low-variance across samples. If they are, voting cannot fix them by construction.
Why not just subtract the model's closed-book preference at decoding time? This contrastive-decoding idea removes the parametric prior but leaves skewed retrieved evidence untouched. Our retrieval-skew measurement shows whether that residual matters. No reported result was found on Lacuna for contrastive decoding on these benchmarks, so we will implement a simple version, which subtracts closed-book option log-probabilities from retrieval-conditioned ones, as an extra comparison on the open models.
Why not just put guideline text or definitions in the prompt? TheoremQA found that definition injection gives no significant gain, which is why our evidence must discriminate between candidates rather than define them.
Proposed Experiments
The benchmarks are MedQA-US, MMLU-Med, and BioASQ-Y/N, which allow direct comparison with MedCoT-RAG's protocol. We add NLPEC with CMeKG, where KMQA reached 61.8% and attributed 32% of its errors to retrieval noise. LAMA-TREx serves as the calibration set where Pile co-occurrence is known. SciAux tests transfer to the misleading-snippet regime. The generators are Qwen3-8B, Llama-3.1-8B, and DeepSeek-R1-70B.
Each benchmark is split by the prior into two subsets. The prior-aligned subset contains items where the gold answer is the co-occurrence-maximal option. The prior-inverted subset contains items where a distractor out-co-occurs the gold. We report accuracy on each subset alongside overall accuracy.
The first decisive check is descriptive. If vanilla retrieval with chain-of-thought shows no excess error on the prior-inverted subset, and its errors do not disproportionately pick the co-occurrence-maximal option, the premise fails and the method is unnecessary. The second check is whether those errors are low-entropy; this determines whether the trigger has any advantage over uncertainty. The third is the head-to-head comparison against balanced per-option retrieval and re-run MedCoT-RAG on the prior-inverted subset, holding the prior-aligned subset to no degradation. The fourth is the swap-check ablation on SciAux misleading snippets. If admissibility filtering recovers a meaningful part of the gap between 77.03% and the 96.22% clean baseline, the evidence check transfers beyond medicine. Significance throughout uses paired bootstrap over questions.
References
- Impact of Co-occurrence on Factual Knowledge of Large Language Models
- Thinking in a Crowd: How Auxiliary Information Shapes LLM Reasoning
- Unveiling Confirmation Bias in Chain-of-Thought Reasoning
- Unbiased Reasoning for Knowledge-Intensive Tasks in Large Language Models via Conditional Front-Door Adjustment
- MedCoT-RAG: Causal Chain-of-Thought RAG for Medical Question Answering
- CausalRAG: Integrating Causal Graphs into Retrieval-Augmented Generation
- CoT-RAG: Integrating Chain of Thought and Retrieval-Augmented Generation to Enhance Reasoning in Large Language Models
- TheoremQA: A Theorem-driven Question Answering Dataset
- Bias Evaluation and Mitigation in Retrieval-Augmented Medical Question-Answering Systems
- Towards Medical Machine Reading Comprehension with Structural Knowledge and Plain Text
- Ice Cream Doesn’t Cause Drowning: Benchmarking LLMs Against Statistical Pitfalls in Causal Inference