# Rayeed, Stewart and Ma on LABEL: Literature Asymmetry between Efficacy and Limits

[Canonical reading](https://escapemeridian.com/readings/rayeed-stewart-and-ma-on-label-literature-asymmetry-between-bb18232d)

S M Rayeed, Charles V. Stewart and Yao Ma argue that general medical prose documents drugs’ benefits more often than their safety limits, and that language models reflect this imbalance. Indications describe what conditions a drug treats; contraindications describe who must avoid it to prevent serious harm. Across 574 drugs from PrimeKG, the authors report that Wikipedia articles and PubMed abstracts named benefits about twice as often as safety limits, with a within-drug odds ratio of 2.19. In their model tests, scaling Qwen2.5 from 0.5B to 14B raised accuracy on indications by 18.3 points, while accuracy on contraindications remained near random guessing. Yet models used safety facts just as effectively as benefits when given a retrieved passage stating the answer. Routing safety questions to official FDA labels also boosted accuracy by 15.8 points whenever a label stated the fact. The authors interpret these results as evidence that models’ medication-safety failures stem from missing safety information in free-text corpora, rather than an inability to reason about risk. They recommend routing clinical AI medication-safety questions to official regulatory labels and structured databases instead of relying on general public records. The reported evidence covers the drugs and text sources studied; how well the pattern holds beyond them remains open.

Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.

## Original source

arXiv cs.AI · Article

Original source date: Nov 14, 2026 (2026-11-14).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-17T13:19:56.431Z. First observed: 2026-11-17T08:28:56.599Z. Retrieved: 2026-11-17T13:18:58.634Z.

- [LABEL: Literature Asymmetry between Efficacy and Limits in Drug Knowledge and Language Models](https://arxiv.org/abs/2611.19166)

## Limits

The reported comparison covers 574 drugs from PrimeKG and reference prose in Wikipedia articles and PubMed abstracts; the supplied abstract does not establish whether the same pattern holds beyond these sources. The reported FDA-label accuracy increase applies whenever the label states the fact; it does not establish performance when a label does not state it.


## Open question

Would the reported LABEL gap persist if measured across drug and text sources beyond the 574 drugs, Wikipedia articles and PubMed abstracts studied?


## Source-stated

The authors report that, across 574 drugs from PrimeKG, Wikipedia articles and PubMed abstracts name drug benefits about twice as often as safety limits.

> Across Wikipedia articles and PubMed abstracts, we find that reference prose names a drug's benefits about twice as often as its safety limits (within-drug odds ratio 2.19).

- [Supporting source](https://arxiv.org/abs/2611.19166)

## Source-stated

In the authors’ Qwen2.5 scaling tests, accuracy on indications rose by 18.3 points, while accuracy on contraindications stayed near random guessing.

> As we scale Qwen2.5 from 0.5B to 14B, accuracy on indications rises by 18.3 points, whereas accuracy on contraindications remains flat near random guessing.

- [Supporting source](https://arxiv.org/abs/2611.19166)

## Source-stated

The authors report that models used safety facts as effectively as benefits when given a retrieved passage stating the answer.

> when we provide models with a retrieved passage that states the answer, they use safety facts just as effectively as benefits.

- [Supporting source](https://arxiv.org/abs/2611.19166)

## Source-stated

The authors report that routing safety questions to official FDA labels boosted accuracy when the label stated the fact.

> routing safety questions to official FDA labels boosts accuracy by 15.8 points whenever the label states the fact.

- [Supporting source](https://arxiv.org/abs/2611.19166)

## Source-stated

The authors conclude that clinical AI systems should route medication-safety questions to official regulatory labels and structured databases rather than rely on general public records.

> Clinical AI systems should therefore route medication-safety questions to official regulatory labels and structured databases rather than relying on general public records.

- [Supporting source](https://arxiv.org/abs/2611.19166)

## Source and contributor context

Rayeed, S M, Stewart, Charles V., Ma, Yao · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.


## Source media

No suitable media was found by the supported page extraction.


## Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.


## arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

- [Explore this context](/entities/arxiv)
- [Reference](https://arxiv.org/)
- [Reference](https://info.arxiv.org/help/api/index.html)

## Editorial provenance

Version 1. Publication ID: b95232d3-fd7c-4860-865c-572ac80893bc.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: passed by anthropic/claude-opus-5.5 on 2026-11-17T13:19:56.431Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material\_development.

Observation: 54e263d7586791b66a68d70ac60b7b7a306f3e2efd346469fa63fbc4234e983d. Source text hash: 17c7b11cad2e24507295af2382774d1d8378e407ec957cbc6a403e7f5053faff.


## Revision and source-check history

### Version 1 — 2026-11-17T13:19:56.431Z

S M Rayeed, Charles V. Stewart and Yao Ma argue that general medical prose documents drugs’ benefits more often than their safety limits, and that language models reflect this imbalance. Indications describe what conditions a drug treats; contraindications describe who must avoid it to prevent serious harm. Across 574 drugs from PrimeKG, the authors report that Wikipedia articles and PubMed abstracts named benefits about twice as often as safety limits, with a within-drug odds ratio of 2.19. In their model tests, scaling Qwen2.5 from 0.5B to 14B raised accuracy on indications by 18.3 points, while accuracy on contraindications remained near random guessing. Yet models used safety facts just as effectively as benefits when given a retrieved passage stating the answer. Routing safety questions to official FDA labels also boosted accuracy by 15.8 points whenever a label stated the fact. The authors interpret these results as evidence that models’ medication-safety failures stem from missing safety information in free-text corpora, rather than an inability to reason about risk. They recommend routing clinical AI medication-safety questions to official regulatory labels and structured databases instead of relying on general public records. The reported evidence covers the drugs and text sources studied; how well the pattern holds beyond them remains open.

## Original source

arXiv cs.AI · Article

Original source date: Nov 14, 2026 (2026-11-14).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-17T13:19:56.431Z. First observed: 2026-11-17T08:28:56.599Z. Retrieved: 2026-11-17T13:18:58.634Z.

- [LABEL: Literature Asymmetry between Efficacy and Limits in Drug Knowledge and Language Models](https://arxiv.org/abs/2611.19166)

## Limits

The reported comparison covers 574 drugs from PrimeKG and reference prose in Wikipedia articles and PubMed abstracts; the supplied abstract does not establish whether the same pattern holds beyond these sources. The reported FDA-label accuracy increase applies whenever the label states the fact; it does not establish performance when a label does not state it.


## Open question

Would the reported LABEL gap persist if measured across drug and text sources beyond the 574 drugs, Wikipedia articles and PubMed abstracts studied?


## Source-stated

The authors report that, across 574 drugs from PrimeKG, Wikipedia articles and PubMed abstracts name drug benefits about twice as often as safety limits.

> Across Wikipedia articles and PubMed abstracts, we find that reference prose names a drug's benefits about twice as often as its safety limits (within-drug odds ratio 2.19).

- [Supporting source](https://arxiv.org/abs/2611.19166)

## Source-stated

In the authors’ Qwen2.5 scaling tests, accuracy on indications rose by 18.3 points, while accuracy on contraindications stayed near random guessing.

> As we scale Qwen2.5 from 0.5B to 14B, accuracy on indications rises by 18.3 points, whereas accuracy on contraindications remains flat near random guessing.

- [Supporting source](https://arxiv.org/abs/2611.19166)

## Source-stated

The authors report that models used safety facts as effectively as benefits when given a retrieved passage stating the answer.

> when we provide models with a retrieved passage that states the answer, they use safety facts just as effectively as benefits.

- [Supporting source](https://arxiv.org/abs/2611.19166)

## Source-stated

The authors report that routing safety questions to official FDA labels boosted accuracy when the label stated the fact.

> routing safety questions to official FDA labels boosts accuracy by 15.8 points whenever the label states the fact.

- [Supporting source](https://arxiv.org/abs/2611.19166)

## Source-stated

The authors conclude that clinical AI systems should route medication-safety questions to official regulatory labels and structured databases rather than rely on general public records.

> Clinical AI systems should therefore route medication-safety questions to official regulatory labels and structured databases rather than relying on general public records.

- [Supporting source](https://arxiv.org/abs/2611.19166)

## Source and contributor context

Rayeed, S M, Stewart, Charles V., Ma, Yao · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.


## Source media

No suitable media was found by the supported page extraction.


## Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.


## arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

- [Explore this context](/entities/arxiv)
- [Reference](https://arxiv.org/)
- [Reference](https://info.arxiv.org/help/api/index.html)

## Editorial provenance

Version 1. Publication ID: b95232d3-fd7c-4860-865c-572ac80893bc.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: passed by anthropic/claude-opus-5.5 on 2026-11-17T13:19:56.431Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material\_development.

Observation: 54e263d7586791b66a68d70ac60b7b7a306f3e2efd346469fa63fbc4234e983d. Source text hash: 17c7b11cad2e24507295af2382774d1d8378e407ec957cbc6a403e7f5053faff.


### passed — 2026-11-17T13:19:56.431Z — anthropic/claude-opus-5.5

Publication: b95232d3-fd7c-4860-865c-572ac80893bc

attribution — pass: The reading correctly attributes the paper to S M Rayeed, Charles V. Stewart and Yao Ma. Findings, interpretation and recommendation are presented as the authors' reported results and conclusions, not as established fact or Escape Meridian opinion.

> Authors:S M Rayeed, Charles V. Stewart, Yao Ma

summary — pass: The reading leads with the authors' thesis and explains indications and contraindications in plain language. It gives the 574-drug scope, the roughly 2x gap, the Qwen2.5 scaling contrast, and the retrieval and FDA-label results with their conditions. It also preserves the authors' causal interpretation and recommendation. Odds ratio, PrimeKG and Qwen2.5 sizes could use tooltips, but they do not block understanding. The title and question are faithful and non-leading. Optional repair: restore the source's note that FDA rules require listing contraindications as rigorously as indications, which sharpens the stakes.

> we find that contraindications are far less documented than indications in general prose

date — pass: evidence.originalPublished is 2026-11-14, which matches the arXiv v1 submission date. The feed date of 2026-11-17 is later and is correctly not used.

> \[Submitted on 14 Nov 2026\]

support — pass: Each claim matches the abstract verbatim in substance, including the 2.19 odds ratio, the 18.3-point gain, near-random contraindication accuracy, equal use of retrieved safety facts, and the 15.8-point gain conditional on the label stating the fact. The reading keeps the recommendation as the authors' conclusion. The limitations and question correctly flag scope without inventing benefits or costs.

> models fail on medication safety not because they cannot reason about risk, but because free-text corpora lack safety facts

media — pass: The selected media is the verified direct arXiv abstract page for this paper. It is the most direct access point to the source.

## Recorded references

Explicit citations in public readings to this reading. Shared topics are not citations. Other record types need an explicit reference adapter.

No public references recorded here yet.

[Reference data (JSON)](/readings/bb18232d40f7a23a600d54215a45660e7ab92a8e01111153cb2bf347fee15a8a/references.json)

## Corrections and challenges

[Send a correction or an opposing source](https://escapemeridian.com/corrections/)

[JSON record](https://escapemeridian.com/readings/rayeed-stewart-and-ma-on-label-literature-asymmetry-between-bb18232d.json)
