MedKIT: Evaluating Knowledge Integration and Generalization in Large Language Models

A benchmark for evaluating knowledge integration and generalization in language models on real, time-stamped clinical evidence. Each row in MedKIT is a factual update derived from a published oncology trial comparison and is paired with seven targeted probes that systematically assess whether newly integrated knowledge transfers across lexical, relational, compositional, and operational levels of generalization, plus a locality probe for knowledge preservation.

Anonymous submission for the NeurIPS 2026 Datasets & Evaluations Track. Author identities and personal contact information are withheld during double-blind review. The construction pipeline and evaluation harness are released anonymously at https://anonymous.4open.science/r/medkit_code_release.

---

TL;DR — what does MedKIT measure?

Existing benchmarks for knowledge editing, continual learning, and retrieval-augmented generation focus almost exclusively on factual recall — can the model reproduce the new fact when prompted with the same surface form it was updated with? MedKIT goes further: it asks whether an integrated fact is consistently usable when the model is queried under different wordings, in the reverse direction, in free-form open-ended settings, or implicitly through clinical recommendations.

The benchmark is grounded in clinical evidence from HemOnc.org, comprises 6,196 updates spanning publications from 1960 to 2026, and ships with a chronological structure that supports realistic sequential-update protocols (daily / weekly / monthly).

Motivation

Language models deployed in dynamic, high-stakes domains such as medicine must be able to integrate evolving evidence as it appears. Knowledge-editing methods modify localized representations, continual post-training adjusts parameters at scale, and retrieval-augmented generation injects new evidence at inference time. All three families have shown impressive results on recall-style evaluations.

In practice, however, successful integration is more demanding: the model must use the new fact consistently across paraphrases, relational transformations, multi-hop reasoning, and open-ended clinical recommendation. The gap between recalling a fact and using it is exactly what MedKIT is built to measure.

Generalization framework (the 7-probe taxonomy)

For each factual update, MedKIT provides one update task plus six companion probes drawn from a four-level generalization taxonomy and a locality test. These 7 probes are the unit of evaluation.

#	Probe	Generalization level	Task format	Question column	GT column
1	Update task	—	Closed QA	closed question 1	answer
2	Rephrasing A	Lexical	Closed QA	closed question 2	answer
3	Rephrasing B	Lexical	Closed QA	closed question 3	answer
4	Direction reversal	Relational	Closed QA	closed question m	closed question m answer
5	Open comparison	Compositional	Open QA	open question 1	ground truth
6	Treatment recommendation	Operational	Open generation	open generation 1	ground truth
7	Locality probe	—	Open QA	locality_question	locality_ground_truth

Definitions:

• Lexical: robustness to surface-level paraphrases of the update query. Paraphrases share the answer; only wording changes.
• Relational: consistency under structural transformations of the underlying relation. Concretely, the update's regimen and comparator are swapped and the label is flipped (superior ↔ inferior, no difference unchanged).
• Compositional: free-form responses where correctness depends on expressing the edited fact rather than selecting a predefined answer.
• Operational: applying the integrated knowledge in an open-ended recommendation task without explicitly being prompted to recall the edited fact.
• Locality: a structurally similar but factually independent comparison drawn from the same oncology group; correct preservation requires the model's response to remain unchanged after the update.

Supported method families

MedKIT is method-agnostic. The schema and probe set support evaluation of all three main families used to integrate new knowledge into language models:

• Knowledge editing: methods that modify model parameters or add localized memory to integrate the update.
• Continual post-training: methods that fine-tune on the update text, optionally over time-ordered batches.
• Retrieval-augmented generation: methods that surface the supporting evidence at inference time.

The evidence and pmids columns expose the supporting PubMed abstracts inline, so closed-book and retrieval-style protocols can be compared on identical inputs.

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); empty 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 (recommended)

• Closed-form tasks (anchor, lexical, relational): exact-match accuracy on {superior, inferior, no difference}. Fall back to an LLM judge when the model's response does not exactly match a label.
• Open-form tasks (compositional, operational, locality): an LLM-based judge scoring response–reference alignment. The paper reports inter-judge stability and human-judge agreement.
• Per-update score: report deltas relative to the pre-update (out-of-the-box) model on the same probe, normalized to [0, 1]. This isolates integration effects from base-model competence.

Sequential evaluation

Each update carries a publication date, so users can construct chronological update streams at any granularity. The dataset is shipped as a single split sorted by date ascending; sequential batching is left to the user.

Common batching choices:

• Daily: one batch per unique publication day; finest granularity, many small batches.
• Weekly / Monthly: aggregated batches, useful when daily granularity is too sparse for stable per-batch metrics.
• Custom: arbitrary cutoffs, sliding windows, or stratified sub-streams over conditions or oncology groups.

When evaluating knowledge integration on truly novel evidence, it is common to restrict the update stream to records dated after a chosen cutoff (e.g. after the model's pre-training cutoff). This reduces the chance of conflating successful integration with surfacing of pre-existing parametric knowledge.

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: 27.7% superior, 43.4% no difference, 28.8% inferior
• Conflict-flagged rows (conflicting_edit=True): 161 (2.6%)
• Mean evidence length: 3,363 characters; max 19,563
• Unique PMIDs cited across the benchmark: 5,666

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

Reproducibility and code

The full construction pipeline (HemOnc + PubMed download, canonicalization, question generation, locality cascade, packaging) and the evaluation harness used in the accompanying paper (knowledge-editing methods, continual post-training, retrieval-augmented generation, LLM-judge scoring) are released anonymously for double-blind review:

https://anonymous.4open.science/r/medkit_code_release

The repository contains the byte-deterministic construction pipeline (fixed seed 42), the Hydra configs used for the main experiment sweep, and the LLM-judge prompts. The full Croissant metadata file (with Responsible AI fields) is available on this Hugging Face dataset as croissant.json; the auto-generated core-only Croissant is served at https://huggingface.co/api/datasets/medkit-anon/MedKIT/croissant.

Source data

MedKIT is a derived benchmark; no human re-annotation is performed.

1. HemOnc Knowledge Base: snapshot 2026-03-12, distributed via the Harvard Dataverse. HemOnc represents clinical knowledge as pairwise treatment comparisons annotated with publication dates, contexts, endpoints, and free-text efficacy labels.
2. PubMed / MEDLINE abstracts: fetched per cited PMID via the NCBI Entrez API; courtesy of the U.S. National Library of Medicine. Abstracts are concatenated per (study, condition) into the evidence field.

Construction pipeline

Following the paper's four-stage description:

1. Extraction & Structuring. Pairwise treatment comparisons are extracted from HemOnc as structured tuples (Treatment A, relation, Treatment B, condition, context, endpoint), retaining only well-defined comparisons.
2. Canonicalization. Tuples are normalized to a fixed three-way label space {superior, inferior, no difference} via a frozen vocabulary lookup plus a deterministic regex rule map. When multiple endpoints are reported for the same comparison, a single canonical instance is selected by endpoint type (Primary > Co-primary > Secondary > Undesignated) and clinical priority (OS > PFS > DFS > ...).
3. Evidence Augmentation. Each canonicalized update is augmented with the supporting PubMed abstracts of all PMIDs tied to its originating study, supporting both parametric integration methods (closed book) and retrieval-based approaches.
4. Task Generation. From each update instance, the seven probes are rendered using a fixed library of clinically validated prompt templates (see Appendix A.4 of the paper). Locality probes are sampled per update via a deterministic four-level cascade over oncology groups: same group / different condition / different context / different answer, falling back to relaxed criteria as needed.

Random choices (locality candidate selection, primary-direction selection during mirror collapse) use a fixed seed (42) to make the construction fully reproducible.

Annotations

No human re-annotation. All labels and prompts are deterministic functions of the HemOnc source data and the canonicalization rule set. Inter-annotator agreement is therefore not applicable; rule-set correctness was validated against a held-out sample during pipeline development. The paper reports judge-based agreement statistics for the open-form evaluation (Spearman ρ up to 0.90 on structured tasks, 0.62 on open-ended generation).

Personal and sensitive information

None. MedKIT contains only aggregated, published trial-level information. No patient-level data, identifiers, demographics, or biospecimens are included. Author names that appear inside published PubMed abstracts are public-domain attribution.

Considerations, biases, and limitations

The paper reports the following on benchmark behavior:

• Strong recall ≠ usable knowledge. Across knowledge editing, continual post-training, and RAG, methods that succeed on the anchor task often fail on relational, compositional, and operational probes. MedKIT is explicitly designed to surface this gap.
• Coverage skew. HemOnc curation prioritizes regimens with regulatory-relevant indications, leading to over-representation of solid and hematologic tumors with phase-3 randomized evidence. Rare-cancer and supportive-care interventions are under-represented.
• 3-class label compression. Trial outcomes are mapped to a fixed 3-way label space, which loses effect size and hedged-finding nuance.
• Abstract-only evidence. Full-text manuscripts, supplementary tables, and figures are not redistributed.
• Temporal anchor. date is the publication date of the earliest cited PMID per update — not the underlying trial completion date.
• Conflicting evidence. 161 rows are flagged conflicting_edit=True, indicating cross-study disagreement on the same comparison. They can be used as a contested-edit split or filtered out for "clean" analyses.
• Linguistic and geographic bias. PubMed indexing skews toward English-language journals and Western clinical research.
• Not for clinical decision-making. MedKIT is a research benchmark for evaluating knowledge-integration methods. It is not validated for clinical use and must not be used to guide patient care.

Citation

Anonymous submission to NeurIPS 2026 Datasets and Evaluations Track, 2026. A full citation block (with authors and DOI) will replace this note at camera-ready.

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

For the duration of the NeurIPS 2026 review, please use the Hugging Face dataset community tab on this repository. Contact details and author information will be added after acceptance.