When a language model answers a question by first fetching documents and then writing from them, a setup called retrieval-augmented generation (RAG), the answer usually arrives with visible sources such as "NCCN guideline," "Nature Communications 2024," or "bioRxiv preprint." These answers are increasingly graded by another language model acting as an automated judge. This note asks whether that judge is checking whether the claim matches the retrieved text, or whether it is mostly reacting to how authoritative and how recent the source looks.
The stakes are highest where correct answers mostly live in recent, non-consensus experimental papers. Acquired radioresistance is one such area. It covers the ways tumor cells adapt to survive radiation, for example by speeding up DNA repair, rewiring metabolism, or changing cell state. Actionable findings such as SIK2-driven homologous recombination repair in colorectal cancer, or blocking glucose uptake with Phloretin in lung cancer models, appear in single-lab preclinical papers rather than in guidelines. In this setting, a judge that trusts "guideline" over "2024 mouse study" will penalize correct answers. A judge that trusts "recent" will wave through answers that stretch mouse data into clinical advice.
We propose a factorial benchmark that, for the first time, crosses two controlled factors on the same answer text. The first is the type of grounding error, drawn from a taxonomy specific to mechanistic radiobiology. The second is the provenance and recency cue attached to the evidence. The benchmark measures which error types an LLM judge misses, and how much its verdict flips when only the source label changes. We then test whether a two-part judge removes the bias. One part checks claims against the retrieved text with no source labels visible. The other separately and explicitly grades strength of evidence. We also test whether simply deleting the labels does just as well.
Core Claim
Our central hypothesis has three parts. First, shortcut bias in LLM judges is not uniform across error types. Second, it concentrates in the errors that require reasoning about what kind of study supports a claim: overgeneralizing from cell lines or mice to patients, moving a finding from one cancer type to another, and inflating the level of evidence. Third, simple entity or direction errors are caught regardless of the source label.
If this holds, existing RAG evaluations in radiation oncology overstate faithfulness in the cases that matter most. The fix is then structural rather than a matter of prompt wording. Whether a claim is supported by its passage is a question that should be judged without source labels. How strong that evidence is should be judged with the labels, and stated openly. Today's judges blend the two questions and hide the blend.
How This Relates to Prior Work
Two papers motivate the bias side. The Silent Judge injects provenance labels (EXPERT/HUMAN/LLM/UNKNOWN) and year labels (1950 vs. 2025) into pairwise judging on ELI5 and LitBench. Its measures are the verdict shift rate (VSR) and the cue acknowledgment rate (CAR), which is how often the judge's written justification mentions the cue. GPT-4o's verdicts shifted by up to 30% on recency and 18% on expert labels, and CAR was 0% throughout. That study used general question answering and creative writing, where the cue is irrelevant by construction. We move the probe into a domain where provenance is partly legitimate evidence. That forces a distinction the Silent Judge setup could not draw: using a source label properly to grade evidence strength versus using it improperly to decide support.
Cognitive bias in LLM reasoning compromises interpretation of clinical oncology notes builds a three-tier taxonomy of oncology reasoning errors and finds that Claude 3.5 Sonnet, used as an auditor, detects that an error exists but cannot reliably classify its subtype. That paper did not involve retrieval or cue manipulation. We borrow its idea that judges should be scored per error subtype, and we add retrieved evidence and controlled cues.
Two papers supply the application regime and a constraint. Development of a RAG-based Expert LLM for Clinical Support in Radiation Oncology built a five-agent RAG system over NCCN guidelines and a radiation oncology textbook and reached 91.5% on the 2021 ACR TXIT exam. It also reported that adding 1,000 uncurated PubMed articles lowered performance. That system was graded only by multiple-choice accuracy. Its noise result is the constraint we take seriously: mixing guideline and research sources is exactly where both the generator and the judge can go wrong. AgentHallu supplies a precedent for a retrieval-error category inside a hallucination taxonomy. It also gives a sobering ceiling: 41.1% step localization by the best judge. This tells us not to expect judges to solve attribution unaided.
On the biology side, SIK2 in colorectal radioresistance, Phloretin with radiotherapy in lung cancer, CDC6 stabilization driving senescence and EMT, and ADAR1–Rad18 in NSCLC are representative of the item sources. Each is a preclinical, single-context finding with a specific mechanism, model system and direction of effect. These are precisely the properties our error taxonomy perturbs.
Research Program
The corpus and questions
We assemble a retrieval corpus with two kinds of documents. The first is guideline and textbook text of the kind used by Liu et al. The second is a few thousand PubMed abstracts on radioresistance mechanisms and radiosensitizers across colorectal, lung, head and neck, nasopharyngeal, bladder and breast cancers.
Radiobiologists write roughly 200 base questions, each answerable from one or two abstracts. Examples include "What mediates SIK2's effect on radioresistance in colorectal cancer, and in what model systems was reversal shown?" and "Is there evidence that GLUT inhibition radiosensitizes lung tumors in patients?"
For each question we keep the gold supporting passages and a gold answer. Each gold answer is annotated with the claim, the mechanism, the direction of effect, the cancer type, the model system (cell line, mouse model or patient cohort) and the evidence level.
The error taxonomy
Starting from each gold answer, we write minimally edited variants. Each variant introduces exactly one error type and leaves fluency and length unchanged. The seven error types are as follows. An entity substitution swaps one protein for a related one. A mechanism misattribution changes the repair pathway, for example from homologous recombination to non-homologous end joining. A direction reversal turns a sensitizer into a protector. A context overgeneralization moves a finding from colorectal to lung cancer, or from a cell line to "patients." Evidence inflation describes a mouse study as clinical evidence. Unsupported synthesis combines two papers' agents into a combination therapy that no paper tested, a tempting error given the field's redundant pathways. Guideline omission ignores a retrieved consensus statement that contradicts the answer.
We add the same kinds of errors as they occur naturally, by running a Liu-style multi-agent RAG system and a plain single-pass RAG system over the questions. Two experts label these naturally occurring errors with the same taxonomy, and agreement between them is reported.
The cue manipulation
Every answer, whether correct or perturbed, is shown with its evidence passages under several label conditions. The text is identical across conditions; only the attached label changes. The provenance conditions are a guideline/textbook label, a peer-reviewed journal with a year, a preprint, and no label. The recency conditions are an old year versus a recent year. We also run a "cue in text" condition, in which phrases like "NCCN recommends" or "a 2024 study found" are written into the passage itself. Deleting labels does not remove this kind of cue.
The proposed judge
The proposed judge separates two decisions. In the first stage, the answer is broken into individual claims. Each claim is checked only against the passage text, with labels removed and in-text provenance phrases masked. For each claim, the judge must report the mechanism, the direction, the cancer type and the model system stated in the passage, and then compare these fields with the claim. This turns the context and evidence-level errors into field comparisons instead of an overall impression.
In the second stage, the judge receives the provenance labels and grades evidence strength. It must name the label it is using. This makes the judge's use of the cue explicit by construction, so the cue acknowledgment rate is no longer zero.
Baselines
The naive reference point is the RAG answer graded by nothing beyond lexical overlap with the gold answer, using ROUGE-L. Evaluating Progress in Automatic Chest X-Ray Radiology Report Generation showed that such metrics reward fluent reports that contain invented findings. Because our perturbations are minimal edits, we expect overlap to rate perturbed and correct answers almost identically. Any judge must clear this floor by a wide margin to be informative.
The strongest existing competitor is an off-the-shelf frontier judge, GPT-4o and a second model family. It is prompted in the standard way to rate whether the answer is faithful to the retrieved evidence, and it sees the evidence exactly as a deployed RAG audit would present it. This is the method the field actually uses. The best reported evidence on its bias comes from The Silent Judge: GPT-4o showed a recency verdict shift of +30% on ELI5 and an EXPERT-versus-UNKNOWN shift of 18%, with a 0% cue acknowledgment rate. We found no result on Lacuna for any LLM judge's per-subtype detection accuracy on retrieval-grounded oncology answers. Kenaston et al. report only qualitatively that subtype classification is poor.
A second strong competitor is the same judge with chain-of-thought plus majority vote over five samples. This mirrors the voting design of Liu et al., and AgentHallu reports step-by-step prompting adds an average 12.3% to localization.
Two ablations isolate the key ingredient. The first, label stripping only, uses the off-the-shelf judge with metadata deleted. The second, claim decomposition only, uses our first stage with labels left visible.
All judges get the same items and the same underlying model, with an equal per-item token budget. Prompts are tuned for each judge on a held-out tuning split of 20% of base questions, with the same number of prompt iterations, and the test split is never seen during tuning.
Beating the strongest baseline means two things on the context and evidence-level error types. First, a statistically significant gain in detection at a matched false-alarm rate on correct answers. Second, a cue verdict shift reduced to within the run-to-run noise measured by re-sampling the same judge on unchanged items. Detection on the simple error types must not degrade at the same time.
Why not just…?
Why not just delete the source labels before judging? This is a serious threat, and the ablation exists to test it. We expect it to fall short for two reasons that the experiments can confirm or refute. First, in this literature the cues are often written into the passages themselves ("NCCN recommends," "in our 2024 xenograft model"), so deleting metadata does not remove them. The "cue in text" condition measures exactly this. Second, deleting labels throws away information a clinician legitimately needs, because evidence level matters for whether a mouse result should change practice. If label stripping alone matches our judge on both detection and verdict shift in the in-text condition, we will report that as the main finding. In that case the contribution becomes the benchmark and a cheap recommendation, not the two-part judge.
Why not use a closed-book frontier model and skip retrieval altogether? Liu et al. found that GPT-4o without retrieval scored 77.3% on the TXIT exam, against 91.5% for curated multi-agent RAG. The radioresistance findings we target are recent preclinical results that are unlikely to be reliably memorized. The deeper reason is that grading a closed-book answer is even more exposed to the judge's own parametric beliefs. Position: Adversarial ML for LLMs Is Not Making Any Progress argues that judge and generator share blind spots. We include a closed-book generation arm so this can be measured rather than assumed.
Why not just ensemble several judges or vote? Voting averages out random disagreement, but shortcut bias is systematic. If every judge favors guideline-labeled text, the vote inherits that preference. The Silent Judge found the same provenance hierarchy across model families. Our chain-of-thought plus voting baseline tests this directly. If voting closes the verdict-shift gap, the two-part design is unnecessary.
Why not fine-tune a judge on expert labels? With a few hundred expert-labeled items, fine-tuning risks learning the surface form of our perturbations instead of the reasoning behind them. It would also not explain which error types were hard. We treat fine-tuning as future work that would build on the benchmark. As a check, we will train a small classifier on the perturbed items and test it on the naturally occurring RAG errors. A large drop between the two would confirm that the controlled perturbations alone do not teach the general skill.
Proposed Experiments and Decisive Checks
The first experiment maps the bias. For each judge, each error type and each cue condition, we report three measures. The first is detection recall at a fixed 10% false-alarm rate on correct answers. The second is the verdict shift rate between label swaps on identical text. The third is the cue acknowledgment rate. The decisive check for the core claim is an interaction test between error type and cue condition. The hypothesis is supported if the verdict shift for overgeneralization and evidence inflation is significantly larger than for entity substitution and direction reversal. If the shift is flat across error types, the claim is wrong, and we would report shortcut bias as uniform in this domain.
The second experiment tests transfer to naturally occurring errors. On errors produced by real RAG systems and labeled by experts, we measure rank agreement (Kendall tau-b) between each judge's scores and the expert scores. We compare this with the 0.522 that RadCliQ achieved against radiologists in a different task, as a rough sense of what useful agreement looks like. We also check whether the RAG configurations that mix guideline and research sources, the setting where Liu et al. saw a performance drop, are the ones off-the-shelf judges overrate most.
The third experiment runs the ablation grid of full judge, label stripping only, claim decomposition only, and standard judge, in both the metadata-cue and in-text-cue conditions. It is the test that decides whether the two-part judge is worth building. It also stops the work from overclaiming if the simpler fix holds.
All items, labels and perturbation scripts will be released so that the cue-crossing protocol can be reused in other domains where correct answers come from recent research rather than guidelines.
References
- The Silent Judge: Unacknowledged Shortcut Bias in LLM-as-a-Judge
- Cognitive bias in LLM reasoning compromises interpretation of clinical oncology notes
- Development of a RAG-based Expert LLM for Clinical Support in Radiation Oncology
- AgentHallu: Benchmarking Automated Hallucination Attribution of LLM-based Agents
- Salt‐inducible kinase 2 confers radioresistance in colorectal cancer by facilitating homologous recombination repair
- Combination treatment with Phloretin enhances the anti-tumor efficacy of radiotherapy in lung cancer models
- Radiation-promoted CDC6 protein stability contributes to radioresistance by regulating senescence and epithelial to mesenchymal transition
- ADAR1 enhances tumor proliferation and radioresistance in non-small cell lung cancer by interacting with Rad18
- Evaluating progress in automatic chest X-ray radiology report generation
- Position: Adversarial ML for LLMs Is Not Making Any Progress