Language models score noticeably higher when they pick from a list of answer options than when they must write an answer themselves. In medicine this gap is often read as evidence that multiple-choice exams overstate clinical ability. Once the options are removed, though, someone has to grade the free-text answers, and increasingly that grader is another language model, an "LLM judge." Medical concepts have many names. The same drug can appear as hydroxychloroquine, Plaquenil or HCQ, and the same gene as TP53 or p53. A judge comparing a student answer against a reference string may mark a correct synonym as wrong. If it does, some of the measured gap reflects the judge's vocabulary rather than the student model's knowledge.
This note proposes a controlled audit that separates the two. We take medical exam questions whose answers are named entities, strip the options, and grade the free-text answers with LLM judges. Then we rewrite correct answers using other names for the same concept, drawn from the Unified Medical Language System (UMLS), a curated thesaurus that assigns one identifier to every synonym of a biomedical concept. Because the synonyms are verified to share an identifier, any verdict that flips under substitution is a judge error by construction. Totaling these errors gives the share of the "generative gap" that belongs to the grader. It also shows whether that share differs enough across models to reorder a leaderboard.
Core Claim
We claim that, on entity-answer medical questions, LLM judges wrongly reject correct answers phrased with less common valid synonyms. We expect this error to grow with how far the synonym's spelling is from the reference string. We further expect it to account for a measurable fraction of the gap between multiple-choice and open-ended accuracy, and to fall unevenly on student models with different naming habits. The last point matters most. A bias that shifts every model's score equally is a nuisance. A bias that depends on each model's vocabulary changes conclusions, and it would even undermine pairwise ranking, which otherwise holds up well.
The claim is falsifiable in both directions. If verified synonym swaps rarely flip verdicts, or flip them equally for every student model, then the generative gap reported by Can Multiple-choice Questions Really Be Useful in Detecting the Abilities of LLMs? stands as a measure of the student models. That paper reports GPT-4 at 69% on multiple choice versus 58% on long-form generation. Either outcome is useful.
Where This Sits
Three strands of work motivate the question, and none of them tests it.
Li et al. established the multiple-choice versus long-form gap by permuting option order. They did not vary wording and did not examine who grades the long-form answers. DOVE showed that option-label formatting alone swings accuracy, but its variations stay within closed-set syntax. Correct Looks Better supplies our conversion method. It strips options from MMLU Pro and GPQA Diamond, grades with pairwise LLM comparisons aggregated by a Bradley-Terry model, and finds that the rankings track true accuracy closely. It never asks whether the judge handles lexical variants of correct answers.
On the judge side, The Silent Judge measures how often verdicts flip when metadata tags such as author type or year are injected. We borrow its verdict-shift metric but move the perturbation into the answer text itself. RobustJudge studies deliberate adversarial attacks. We study the naturally occurring variation that a benign, correct model produces.
Two papers constrain the design. Are Synonym Substitution Attacks Really Synonym Substitution Attacks? found that most "synonym" swaps in attack studies were not true synonyms. In PWWS, 75.4% were sense-mismatched, and cosine-similarity filtering reached only 0.627 AUPR. Any brittleness finding built on embedding-generated synonyms is therefore suspect. This is why we take synonyms only from UMLS concept identifiers and have annotators check them in context. Biomedical Concept Normalization by Leveraging Hypernyms shows that normalizers confuse specific diseases with their parent categories. We use such parent-category substitutions as negative controls: a judge should reject "lymphoma" when the answer is a specific subtype.
The normalization method comes from Self-Alignment Pretraining for Biomedical Entity Representations (SapBERT). SapBERT trains on UMLS so that all synonyms of a concept sit close together in embedding space. Finally, Large language models encode clinical knowledge shows why automated grading is needed. Its long-form medical evaluation required clinician panels, which the automated judges studied here are meant to replace.
Research Program
Question selection. We build the test set from the English MedQA (USMLE) questions of What Disease does this Patient Have?. We keep only items whose gold answer is a single named entity: a diagnosis, drug, organism or gene. We link each gold answer to a UMLS identifier using SapBERT nearest-neighbor retrieval, and annotators confirm each link. As a secondary set that ties back to Correct Looks Better, we use the health-related MMLU Pro items that paper already converted.
Student models and judges. Roughly eight open and API student models answer the stripped questions. We include general models and medical fine-tunes such as Llama-3-Meditron, since medical tuning may change naming habits. Two to three judges grade the answers in two modes. In pointwise mode the judge sees the reference answer and returns correct or incorrect. In pairwise mode we follow the Bradley-Terry protocol of Remeli and Hardt.
Counterfactual variants. For every answer the judge marks correct, we create variants that swap the answer entity for another UMLS synonym of the same concept, leaving the rest of the text untouched. We sort these synonyms by character-level distance from the reference string: near variants, abbreviations, brand or trade names, and lay terms. We add two sets of controls. Parent-concept substitutions and sense-mismatched strings are cases where rejection is correct. Paraphrases of the judge's instructions, in the style of ProSA, let us compare lexical brittleness against ordinary instruction brittleness.
Human labels and decomposition. A clinically trained annotator panel labels a stratified sample of about 600 original student answers as correct or incorrect. This produces the key decomposition. The multiple-choice minus judged-open gap splits into two parts: multiple-choice minus human-graded open accuracy, which is the student-side gap, and human-graded open accuracy minus judged open accuracy, which is the judge-side gap. We also report the judge-side gap per student model.
Baselines
Naive reference. The do-nothing reference is the gap as usually reported: multiple-choice accuracy minus open-ended accuracy under a judge that sees only the reference string. For context, Flan-PaLM reaches 67.6% multiple-choice accuracy on MedQA (Large language models encode clinical knowledge). We found no reported judge-graded open-ended MedQA figure on Lacuna, so our own measurement of this gap is the starting point.
Pairwise ranking. The strongest existing method is the pairwise Bradley-Terry ranking of Correct Looks Better. It reports Spearman's ρ above 0.9 against ground-truth accuracy, and a Kendall distance of 20% versus 38.2% for direct judging even with a weak judge. Its robustness is the main reason one might think lexical bias does not matter. We run it on identical answers, with identical judges and the same number of comparisons per model pair. It "fails" in our sense if substituting synonyms into one model's answers changes that model's Bradley-Terry rank relative to human-graded accuracy, while the unsubstituted ranking stays within its reported agreement.
Cheap fixes. Two fixes are the strongest competitors. The first is reference expansion: give the judge the full UMLS synonym list for the gold concept. The second is normalize-then-match with no judge at all: link the student's answer entity to a UMLS identifier with SapBERT and compare identifiers. Normalizer accuracy on benchmarks is a known quantity; BIOSYN reports 91.1% top-1 on NCBI Disease, and we check our linker on NCBI Disease and BC5CDR before use. Both fixes use the same answers and are scored against the same human labels.
Ablation. The ablation that isolates the proposal's key ingredient is to replace the UMLS-verified synonyms with embedding-generated candidates of the kind Chiang and Lee criticize. If verdict-shift rates differ sharply between the two, then the verification step is doing real work. Otherwise the brittleness measurements in prior attack studies can be trusted as they stand.
Why not just…?
Why not just use a stronger judge or tell it to accept synonyms? This may shrink the effect, and we test it directly: the best available judge is run with an explicit "accept equivalent names" instruction. The Silent Judge suggests the effect may persist, since frontier judges shifted verdicts while never acknowledging the cue (0% acknowledgment rate). The rarity-stratified variants show whether an instruction helps with abbreviations and lay terms or only with near-spellings. If it closes the gap everywhere, we report that and the problem becomes a one-line fix.
Why not just give the judge the synonym list? This is a real threat to any claim that a new method is needed, which is why this proposal is framed as a measurement rather than a method. Reference expansion is likely to fix answers that name the gold concept. It cannot fix answers that correctly describe the concept without naming it, or that use slang missing from UMLS, a limitation SapBERT itself notes. The human-labeled sample tells us how many residual judge errors fall into these categories. Whatever that number is, the decomposition of the published gap remains the contribution.
Why not just normalize and exact-match, skipping the judge? Linking works for single-entity answers, but it inherits linker errors and confuses parents with children, as the hypernym findings of Yan et al. show. It also cannot grade answers that hedge or list several differentials. We compare it against the judges on the human sample and report where each fails. If it wins, that is itself a finding: for entity-answer medicine, a thesaurus beats a judge.
Proposed Experiments
Decisive check 1: synonym verdict-shift rate. The first check is the verdict-shift rate on verified synonym swaps, stratified by lexical distance and paired with false-acceptance rates on parent-concept controls. A monotone rise with distance, alongside low false acceptance, confirms that the effect is a vocabulary failure and not general leniency. We found no reported result on Lacuna for LLM-judge robustness to ontology-verified biomedical synonyms.
Decisive check 2: judge-side share of the gap. The second check is the judge-side share of the multiple-choice to open-ended gap against human labels, with a bootstrap confidence interval. We repeat three runs at temperature zero, following the reproducibility guidance of Towards Reproducible LLM Evaluation.
Decisive check 3: differential bias. The third check is whether the judge-side error differs across student models, especially general versus medically tuned ones, and whether that difference reorders pointwise and Bradley-Terry rankings relative to human-graded accuracy. If the rankings hold, then the Remeli and Hardt result extends to lexical variation in medicine. If they break, open-ended medical leaderboards need synonym-aware grading before their gaps can be interpreted.
References
- Can Multiple-choice Questions Really Be Useful in Detecting the Abilities of LLMs?
- DOVE: A Large-Scale Multi-Dimensional Predictions Dataset Towards Meaningful LLM Evaluation
- Correct Looks Better: Pairwise Comparisons Reveal Accuracy Rankings
- The Silent Judge: Unacknowledged Shortcut Bias in LLM-as-a-Judge
- LLMs Cannot Reliably Judge (Yet?): A Comprehensive Assessment on the Robustness of LLM-as-a-Judge
- Are Synonym Substitution Attacks Really Synonym Substitution Attacks?
- Biomedical Concept Normalization by Leveraging Hypernyms
- Self-Alignment Pretraining for Biomedical Entity Representations
- Large language models encode clinical knowledge
- What Disease does this Patient Have? A Large-scale Open Domain Question Answering Dataset from Medical Exams
- Llama-3-Meditron: An Open-Weight Suite of Medical LLMs Based on Llama-3.1
- ProSA: Assessing and Understanding the Prompt Sensitivity of LLMs
- Biomedical Entity Representations with Synonym Marginalization
- Towards Reproducible LLM Evaluation: Quantifying Uncertainty in LLM Benchmark Scores