MedKIT: Evaluating Knowledge Integration and Generalization in Large Language Models

MedKIT (Medical Knowledge Integration and Transfer) is a benchmark for evaluating how language models integrate and apply new knowledge under realistic sequences of clinical updates.

Grounded in time-stamped oncology evidence, MedKIT represents treatment comparisons as factual knowledge updates and pairs them with targeted evaluation probes that assess whether newly integrated knowledge transfers across lexical variation, relational transformations, compositional reasoning, and open-ended application, alongside locality tests for knowledge preservation.

---

TL;DR: What does MedKIT measure?

Integrating new knowledge is not merely a matter of storing and recalling a fact. A model should be able to consistently use newly acquired knowledge across tasks and contexts.

MedKIT evaluates this distinction between recall and usable knowledge. Given a factual update derived from clinical evidence, the benchmark tests whether a model can recover the update itself, recognize it under lexical variation, correctly reason about reversed relations, combine it with other information, and apply it in open-ended settings. It additionally measures whether integrating new knowledge interferes with unrelated knowledge.

MedKIT is derived from real-world oncology evidence from HemOnc.org and contains 6,196 canonical updates extracted from clinical studies published between 1960 and 2026. Its temporal structure enables the evaluation of knowledge integration under realistic sequences of evolving clinical evidence.

Motivation

Real-world knowledge evolves continuously, requiring language models to integrate new information over time. This is particularly important in medicine, where clinical evidence changes and outdated knowledge can pose safety risks.

Existing evaluations of knowledge integration largely focus on whether a model can recall newly introduced facts. Successful integration, however, requires more than recall: newly acquired knowledge should be consistently usable across tasks and contexts.

MedKIT is designed to evaluate this gap. It tests whether factual updates transfer across lexical variation, relational transformations, compositional reasoning, and open-ended application, while also measuring whether unrelated knowledge is preserved.

Generalization framework: seven targeted probes

MedKIT follows a fact-centered evaluation protocol: each factual update is evaluated through seven targeted probes derived from the same underlying clinical comparison. The probes vary the task formulation and output space while keeping the underlying fact fixed, enabling a controlled evaluation of whether newly integrated knowledge transfers beyond the original update task.

#	Probe	Evaluation dimension	Task format	Question column	GT column
1	Update	Update success	Closed QA	closed question 1	answer
2	Lexical A	Lexical	Closed QA	closed question 2	answer
3	Lexical B	Lexical	Closed QA	closed question 3	answer
4	Relational	Relational	Closed QA	closed question m	closed question m answer
5	Compositional	Compositional	Open QA	open question 1	ground truth
6	Operational	Operational	Open generation	open generation 1	ground truth
7	Locality	Knowledge preservation	Open QA	locality_question	locality_ground_truth

Probe definitions

• Update: the canonical treatment comparison, formulated as a three-way classification task (superior, inferior, or no difference). It directly measures whether the factual update has been successfully integrated.

• Lexical: tests robustness to surface-level variation using two paraphrased versions of the update question that preserve its semantics and answer.

• Relational: tests whether the model has internalized the relationship between treatments by reversing their order in the comparison. The corresponding label is inverted (superior ↔ inferior; no difference remains unchanged).

• Compositional: tests whether the updated fact can be used within an open-ended reasoning process. The model compares treatments in the given clinical context and must produce a factually correct response rather than select from a fixed answer space.

• Operational: tests whether the updated knowledge is applied implicitly in an open-ended treatment recommendation, without explicitly prompting the model to reproduce or compare the updated fact.

• Locality: tests whether integrating an update affects unrelated knowledge. The probe uses a structurally similar but factually independent comparison, preferentially drawn from the same oncology group.

Supported knowledge integration methods

MedKIT is method-agnostic: it evaluates the outcome of knowledge integration rather than prescribing how new knowledge should be incorporated. The benchmark can therefore be used to evaluate different families of knowledge integration methods, including:

• Knowledge editing: methods that update individual facts through parameter modifications, learned editors, or memory-based mechanisms.
• Continual post-training: methods that integrate new knowledge through continued training on individual updates or temporally ordered batches of updates.
• Retrieval-augmented generation (RAG): methods that store new knowledge externally and retrieve relevant evidence at inference time.

MedKIT provides the supporting evidence and linked PubMed identifiers through the evidence and pmids fields, enabling parameter-based and retrieval-based methods to be evaluated on the same underlying factual updates.

Schema

Column	Type	Description
date	string	Publication date (YYYY-MM-DD) of the earliest cited PMID for this update; the temporal anchor.
year	int	Year extracted from date.
evidence	string	Concatenation of the supporting PubMed abstracts (separated by \n\n).
pmids	string	Semicolon-separated PubMed IDs whose abstracts are concatenated in evidence.
regimen	string	Treatment 1 in the comparison (the regimen being evaluated).
comparator	string	Treatment 2 in the comparison.
condition	string	Disease + appended stage / status, e.g. "Cervical cancer (advanced)".
context	string	Treatment context, e.g. "Non-curative first-line therapy", "Adjuvant therapy".
endpoint	string	Clinical endpoint (e.g. OS, PFS, ORR, DFS); Could not be determined when not reported.
endpoint_type	string	Primary / Co-primary / Secondary / Undesignated.
answer	string	Anchor label, one of {"superior", "inferior", "no difference"}.
ground truth	string	Verbalized statement combining regimen, label, comparator, condition, context, endpoint.
closed question 1	string	Anchor — closed-form 3-way classification prompt (system prompt baked in).
closed question 2/3	string	Lexical paraphrases of the anchor.
closed question m	string	Relational probe — anchor with regimen and comparator swapped.
closed question m answer	string	Flipped label for the relational probe (superior↔inferior; no difference unchanged).
open question 1	string	Compositional probe — structured open-form prompt eliciting a free-form comparison answer.
open generation 1	string	Operational probe — open-ended treatment-recommendation prompt without explicit recall cue.
option 1/2/3	string	Constant {superior, inferior, no difference}; stored explicitly for prompt-template stability.
locality_question	string	Locality probe — open-form prompt for a same-distribution but unrelated update.
locality_ground_truth	string	Verbalized answer for the locality probe.
conflicting_edit	bool	True if the same (condition, context, endpoint, regimen, comparator) appears in another row with a disagreeing label across studies/dates. Useful as a clean / contested split.

Answer encoding for the closed-form and locality tasks: "superior", "inferior", "no difference". The option N columns preserve this mapping for prompt-template stability.

Evaluation protocol

MedKIT evaluates the effect of knowledge integration by comparing model performance before and after an update.

• Closed-form tasks (update, lexical, relational) use the answer space {superior, inferior, no difference} and are evaluated primarily through deterministic label matching. When a response cannot be resolved deterministically, an LLM judge can be used to determine whether it implies the correct label.

• Open-form tasks (compositional, operational, locality) require semantic evaluation and are scored using an LLM judge based on the model response, ground-truth comparison, and clinical context.

For each task, performance is reported as the change relative to the model before knowledge integration:

Δ = s_post - s_pre

where s_pre is the score of the out-of-the-box model and s_post is the score after integrating the corresponding updates. This post–pre evaluation isolates changes in model behavior caused by knowledge integration from differences in initial model competence.

The paper additionally validates the LLM-based evaluation through inter-judge robustness, agreement with clinician judgments, and intra-judge consistency.

Sequential evaluation

MedKIT preserves the temporal structure of the underlying clinical evidence. Updates are ordered by publication date and can be grouped into chronological batches, enabling evaluation under realistic sequences of evolving knowledge.

The paper considers three batching strategies:

• Daily: one batch per publication day, providing the most fine-grained update sequence.
• Weekly: updates are grouped by publication week. This is the primary setting used in the paper, providing a practical balance between fine-grained updates and manageable sequence length.
• Monthly: updates are grouped by publication month, producing fewer, larger update batches.

The released dataset retains the publication date for each update, so users can also construct alternative chronological batching schemes for their own evaluation protocols.

For experiments targeting genuinely new knowledge, the benchmark can be restricted to updates published after a model-specific knowledge cutoff. The paper uses a post-2025 evaluation window to reduce the likelihood that evaluated updates were already represented in the models' pre-training data.

Dataset statistics

• Knowledge updates: 6,196
• Treatment regimens (unique): 2,135
• Distinct treatment options across regimen/comparator slots: 3,334
• Clinical conditions: 329 with stage suffix (143 base disease names without stage)
• Oncology groups covered: 12 (Hematologic, Breast, Digestive, Lung, Urological, Gynecological, Head&Neck, CNS, Melanoma, Sarcoma, NET, Other)
• Distinct endpoints reported: 318 (dominated by OS, PFS, DFS, ORR, RFS and other standard oncology endpoints)
• Probes per update: 7
• Date range: 1960-03-15 → 2026-02-01 (publication date of earliest cited PMID)
• Label distribution: 27.9% superior, 43.4% no difference, 28.7% inferior
• Conflict-flagged rows (conflicting_edit=True): 154 (2.5%)
• Mean evidence length: 3,363 characters; max 19,563
• Unique PMIDs cited across the benchmark: 5,666

The 6,196 updates constitute the complete set of eligible comparisons from the HemOnc.org snapshot after preprocessing and canonicalization, rather than a sampled subset. Conflict-flagged updates are retained in the released benchmark but excluded from the main experiments.

A 100-row stratified preview is available at data/sample/medkit_sample_100.csv for fast inspection without downloading the full dataset.

Reproducibility and code

The code accompanying MedKIT provides the benchmark construction and evaluation framework used in the paper:

https://github.com/bethgelab/MedKIT

The released code includes the data-processing pipeline, preprocessing scripts, task and prompt templates, evaluation framework, and experiment configurations required to reproduce the main experiments. It covers the evaluated knowledge-integration approaches, temporal batching, closed- and open-form evaluation, and LLM-judge scoring.

Benchmark construction is entirely rule-based: no learned or generative model is used during preprocessing. Task variants are generated deterministically from the structured clinical comparisons using fixed templates, and seed 42 is used to reproducibly resolve rare symmetric comparisons during canonicalization.

This Hugging Face release contains the canonical MedKIT benchmark used for the reported experiments. Additional implementation details, including preprocessing decisions, evaluation protocols, prompt templates, experimental settings, and reproduction instructions, are provided in the accompanying code repository and paper.

The full Croissant metadata file, including Responsible AI metadata, is available as croissant.json. Hugging Face also serves the automatically generated core Croissant metadata through its dataset API.

Source data

MedKIT is derived from publicly available, clinically curated oncology evidence. Benchmark construction uses a fixed source snapshot and an entirely rule-based preprocessing pipeline; no learned or generative model is used during preprocessing.

1. HemOnc.org: MedKIT is built from a fixed HemOnc.org knowledge-base snapshot dated 2026-03-12. HemOnc provides clinically curated, evidence-backed comparisons of oncology treatments. Each comparison specifies two treatment regimens, a clinical condition and context, an efficacy endpoint, a comparative outcome, and associated publication information.

2. PubMed: Each retained update is linked to its supporting published evidence. MedKIT includes the corresponding PubMed metadata, titles, and abstracts, which provide the evidence used to ground the factual update.

The resulting clinical comparisons are canonicalized into structured factual updates with the label space {superior, inferior, no difference}. Rows without resolvable supporting evidence are excluded.

Construction pipeline

MedKIT transforms structured clinical comparisons from HemOnc.org into factual knowledge updates and associated evaluation probes through a four-stage construction pipeline:

1. Extraction and Structuring. Pairwise treatment comparisons are extracted and represented as structured tuples (Treatment A, relation, Treatment B, condition, context, endpoint). Incomplete or ambiguous entries are removed, and equivalent comparisons are deduplicated to retain well-defined updates.

2. Canonicalization. Comparisons are normalized into a consistent representation with the fixed label space {superior, inferior, no difference}. When a study reports multiple endpoints for the same treatment comparison, a single canonical endpoint is retained using a fixed priority scheme based first on endpoint type (Primary > Co-primary > Secondary > Undesignated) and then on clinical relevance, with OS and PFS receiving highest priority. This process yields 6,196 canonical updates, representing the complete set of eligible comparisons after preprocessing.

3. Evidence Augmentation. Each update is augmented with its supporting clinical evidence through linked PubMed titles and abstracts. This evidence-grounded representation supports both parametric knowledge-integration methods and retrieval-based approaches.

4. Task Generation. Each factual update is paired with seven targeted evaluation probes derived from the same underlying clinical comparison. The task variants are generated deterministically from fixed templates manually designed in collaboration with an MD clinician. Locality probes use structurally similar but factually independent comparisons, preferentially from the same oncology group and with a different condition.

The preprocessing pipeline is entirely rule-based; no learned or generative model is used during benchmark construction. In rare symmetric comparisons where either treatment ordering can serve as the canonical representation, seed 42 is used to resolve the ordering reproducibly.

Annotations

MedKIT does not introduce a separate human re-annotation of the underlying clinical comparisons. Outcome labels are inherited from the clinically curated HemOnc source data and mapped through a rule-based preprocessing pipeline to the fixed label space {superior, inferior, no difference}.

Evaluation tasks are generated deterministically from the structured clinical comparisons using fixed templates. These templates were manually designed in collaboration with an MD clinician to ensure clinically plausible and semantically consistent formulations.

Human evaluation is used to validate the LLM-based scoring of open-form tasks. Two independent clinicians annotated a 100-item evaluation set. The primary judge, gpt-4o, achieves Spearman correlations of ρ = 0.83 on the compositional task and ρ = 0.64 on the operational task against the two-clinician consensus. For reference, inter-clinician agreement is ρ = 0.76 and ρ = 0.62, respectively.

Personal and sensitive information

MedKIT is derived exclusively from publicly available clinical evidence from HemOnc.org and PubMed. The benchmark contains no patient-level information.

The released data consist of processed, study-level treatment comparisons and their associated publication metadata and evidence. MedKIT is intended as a research benchmark and does not contain private clinical records or other patient-specific data.

Considerations, biases, and limitations

MedKIT is designed to evaluate whether newly integrated factual knowledge can be used consistently across tasks and contexts. Several limitations should be considered when interpreting benchmark results:

• Pairwise formulation. MedKIT represents each update as a pairwise comparison between two treatment regimens on a single endpoint, with labels {superior, inferior, no difference}. This enables controlled and deterministic evaluation but abstracts away clinically relevant structure, including quantitative effect sizes, uncertainty, multi-arm relationships, and subgroup-specific effects. MedKIT therefore evaluates the integration of qualitative treatment comparisons rather than the full clinical decision-making process.

• Source-data biases. MedKIT inherits biases from HemOnc.org and the underlying clinical literature. Coverage across oncology groups is uneven, outcome labels reflect source-data curator judgments, and endpoint prioritization introduces additional assumptions. PubMed-indexed clinical trials also over-represent Western populations. Benchmark results should therefore not be assumed to generalize beyond the distribution represented by these sources.

• Conflicting evidence. Some treatment comparisons occur across multiple studies or publication dates with different outcomes. These cases are retained in the released benchmark and marked with conflicting_edit=True, but are excluded from the main experiments to focus the reported evaluation on unambiguous updates.

• LLM-based evaluation. Compositional, operational, and locality tasks require semantic evaluation and therefore rely on an LLM judge. Judge agreement with clinicians is stronger on compositional tasks than on open-ended operational generation. The clinician validation is additionally limited to two expert annotators and 100 responses. Fine-grained differences on open-form tasks should therefore be interpreted with appropriate caution.

• Open-generation scores do not measure clinical correctness. The operational evaluation measures whether a generated recommendation is consistent with the underlying treatment comparison. It does not verify dosage, contraindications, safety considerations, or the overall clinical appropriateness of a recommendation. These scores measure knowledge use rather than clinical validity.

• Refusals and stochastic evaluation. Refusals are scored as incorrect on closed-form tasks and neutral on open-form tasks, which does not distinguish safe abstention from an incorrect response. Open-form evaluation also depends on stochastic judge calls, introducing small amounts of evaluation variance.

• Recall does not imply usable knowledge. The experiments in the accompanying paper show that strong performance on the original update and lexical tasks does not necessarily transfer to other task formulations. Relational generalization remains challenging, and the evaluated methods show no meaningful improvement over pre-update performance on compositional and operational tasks.

• Research benchmark, not clinical decision support. MedKIT is intended for the development and evaluation of knowledge-integration methods. Benchmark scores reflect the integration and use of benchmark-specific knowledge and do not imply clinical readiness. MedKIT must not be used as a substitute for clinical validation or to guide patient care.

Citation

If you use MedKIT, please cite:

@inproceedings{thede2026medkit,
  title     = {MedKIT: Evaluating Knowledge Integration and Generalization in Large Language Models},
  author    = {Thede, Lukas and Atri, Yash Kumar and Chen, David and Bitterman, Danielle and Bethge, Matthias and Hartvigsen, Tom and Akata, Zeynep},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}

Please also cite the upstream sources:

• HemOnc: Warner JL et al., Journal of Biomedical Informatics 2019; 96:103239. doi:10.1016/j.jbi.2019.103239.
• NLM / PubMed: Courtesy of the U.S. National Library of Medicine.

See ATTRIBUTION.md for full attribution and reuse terms.

License

CC BY-NC-SA 4.0 for the MedKIT release. The license is inherited from the HemOnc Knowledge Base ShareAlike clause: derivatives of HemOnc must carry the same license, and MedKIT is one such derivative.

Practical implications:

• Permitted: academic and non-commercial use with attribution, including reproduction in research papers, evaluation in code releases, and remixing into derivative academic benchmarks.
• Not permitted: commercial use of MedKIT in any form, including as training data, benchmark inputs, or evaluation harness for commercial products.
• ShareAlike: any derivative work (a filtered, augmented, or re-batched version of MedKIT) must be redistributed under CC BY-NC-SA 4.0.

Upstream sources retain their own terms; see ATTRIBUTION.md and LICENSE.

Contact

Please use the community tab of this repository for questions about the data, or open an issue in the code repository.