# Hu Cang argues asking the right question is key to AI discovery

[Canonical reading](https://escapemeridian.com/readings/hu-cang-argues-asking-the-right-question-is-key-to-ai-discovery-9f0ab794)

Hu Cang argues that AI discovery can fail when experiments are chosen only where candidate models disagree. In the chain of hypotheses described in the abstract, AI scientists fit candidate models to existing observations and often test where those models make different predictions. But if the starting data miss a narrow feature, simple curves can agree across the unmeasured gap, leaving disagreement-guided search with no reason to investigate it. In 32 closed-loop trials with a narrow peak omitted from the starting data, Claude Sonnet 5.5 and GPT-6.1-Sol sampled inside the peak region less often than random search. The agent that followed model disagreement more closely missed the peak in all 32 trials. The abstract says discovery succeeded when experiments probed the unmeasured gap early and followed up across the anomaly. Cang’s point is that asking where to measure matters: a search strategy can overlook a phenomenon precisely because the available observations give its models no reason to look there.

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 15, 2026 (2026-11-15).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-17T10:47:54.017Z. First observed: 2026-11-17T08:30:48.650Z. Retrieved: 2026-11-17T10:46:45.234Z.

- [Asking the right question is key to AI discovery in a chain of hypotheses](https://arxiv.org/abs/2611.19841)

## Limits

The available evidence is the arXiv page’s abstract. It reports trials involving a narrow peak omitted from the starting data; the evidence provided does not describe results in other discovery settings.


## Open question

Would probing the unmeasured gap early also lead to discovery in settings beyond trials with a narrow peak omitted from the starting data?


## Source-stated

Cang argues that disagreement-guided search can miss a phenomenon when the starting observations omit a narrow feature, because candidate models may agree across the unmeasured gap.

> when early data miss a narrow feature, model selection favors simple curves that agree across the unmea- sured gap, so disagreement-guided search has no reason to look where the new phenomenon lies.

- [Supporting source](https://arxiv.org/abs/2611.19841)

## Source-stated

In 32 closed-loop trials with a narrow peak omitted from the starting data, both tested AI scientists sampled inside the peak region less often than random search.

> across 32 closed-loop trials with a narrow peak omitted from the starting data, both sampled inside the peak region less often than random search.

- [Supporting source](https://arxiv.org/abs/2611.19841)

## Source-stated

The agent that followed model disagreement more closely missed the peak in all 32 trials; the abstract says discovery succeeded when experiments probed the unmeasured gap early and followed up across the anomaly.

> the agent that followed model disagreement more closely missed the peak in all 32 trials, whereas discovery succeeded when experiments probed the unmeasured gap early and followed up across the anomaly.

- [Supporting source](https://arxiv.org/abs/2611.19841)

## Source-stated

Cang concludes that choosing where to experiment, including probing gaps not indicated by current model disagreement, is key to discovery in a chain of hypotheses.

> Asking the right question, rather than testing only where current models disagree, is therefore key to AI discovery in a chain of hypotheses.

- [Supporting source](https://arxiv.org/abs/2611.19841)

## Source and contributor context

Cang, Hu · 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)

## Claude

Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.

- [Explore this context](/entities/claude)
- [Reference](https://claude.com/product/overview)
- [Reference](https://www.anthropic.com/news)
- [Reference](https://www.anthropic.com/research/team/economics)

## Editorial provenance

Version 1. Publication ID: 62e96ff0-bbe7-4fa9-8f4f-7b9d8f07b55f.

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

Source check: passed by anthropic/claude-opus-5.5 on 2026-11-17T10:47:54.017Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material\_development.

Observation: 840656c43f8911b14bc0b88949bf8787bd3189ecca85cb24806975db3fbe858b. Source text hash: 73b33eb303a9a71e775847f0dd5c222750d8c0efc8927184ded9b27771b8902d.


## Revision and source-check history

### Version 1 — 2026-11-17T10:47:54.017Z

Hu Cang argues that AI discovery can fail when experiments are chosen only where candidate models disagree. In the chain of hypotheses described in the abstract, AI scientists fit candidate models to existing observations and often test where those models make different predictions. But if the starting data miss a narrow feature, simple curves can agree across the unmeasured gap, leaving disagreement-guided search with no reason to investigate it. In 32 closed-loop trials with a narrow peak omitted from the starting data, Claude Sonnet 5.5 and GPT-6.1-Sol sampled inside the peak region less often than random search. The agent that followed model disagreement more closely missed the peak in all 32 trials. The abstract says discovery succeeded when experiments probed the unmeasured gap early and followed up across the anomaly. Cang’s point is that asking where to measure matters: a search strategy can overlook a phenomenon precisely because the available observations give its models no reason to look there.

## Original source

arXiv cs.AI · Article

Original source date: Nov 15, 2026 (2026-11-15).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-17T10:47:54.017Z. First observed: 2026-11-17T08:30:48.650Z. Retrieved: 2026-11-17T10:46:45.234Z.

- [Asking the right question is key to AI discovery in a chain of hypotheses](https://arxiv.org/abs/2611.19841)

## Limits

The available evidence is the arXiv page’s abstract. It reports trials involving a narrow peak omitted from the starting data; the evidence provided does not describe results in other discovery settings.


## Open question

Would probing the unmeasured gap early also lead to discovery in settings beyond trials with a narrow peak omitted from the starting data?


## Source-stated

Cang argues that disagreement-guided search can miss a phenomenon when the starting observations omit a narrow feature, because candidate models may agree across the unmeasured gap.

> when early data miss a narrow feature, model selection favors simple curves that agree across the unmea- sured gap, so disagreement-guided search has no reason to look where the new phenomenon lies.

- [Supporting source](https://arxiv.org/abs/2611.19841)

## Source-stated

In 32 closed-loop trials with a narrow peak omitted from the starting data, both tested AI scientists sampled inside the peak region less often than random search.

> across 32 closed-loop trials with a narrow peak omitted from the starting data, both sampled inside the peak region less often than random search.

- [Supporting source](https://arxiv.org/abs/2611.19841)

## Source-stated

The agent that followed model disagreement more closely missed the peak in all 32 trials; the abstract says discovery succeeded when experiments probed the unmeasured gap early and followed up across the anomaly.

> the agent that followed model disagreement more closely missed the peak in all 32 trials, whereas discovery succeeded when experiments probed the unmeasured gap early and followed up across the anomaly.

- [Supporting source](https://arxiv.org/abs/2611.19841)

## Source-stated

Cang concludes that choosing where to experiment, including probing gaps not indicated by current model disagreement, is key to discovery in a chain of hypotheses.

> Asking the right question, rather than testing only where current models disagree, is therefore key to AI discovery in a chain of hypotheses.

- [Supporting source](https://arxiv.org/abs/2611.19841)

## Source and contributor context

Cang, Hu · 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)

## Claude

Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.

- [Explore this context](/entities/claude)
- [Reference](https://claude.com/product/overview)
- [Reference](https://www.anthropic.com/news)
- [Reference](https://www.anthropic.com/research/team/economics)

## Editorial provenance

Version 1. Publication ID: 62e96ff0-bbe7-4fa9-8f4f-7b9d8f07b55f.

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

Source check: passed by anthropic/claude-opus-5.5 on 2026-11-17T10:47:54.017Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material\_development.

Observation: 840656c43f8911b14bc0b88949bf8787bd3189ecca85cb24806975db3fbe858b. Source text hash: 73b33eb303a9a71e775847f0dd5c222750d8c0efc8927184ded9b27771b8902d.


### passed — 2026-11-17T10:47:54.017Z — anthropic/claude-opus-5.5

Publication: 62e96ff0-bbe7-4fa9-8f4f-7b9d8f07b55f

attribution — pass: The reading attributes the argument and results to Hu Cang, the sole listed author of the arXiv paper. It keeps the findings as the paper's reported claims rather than established facts, and it attributes the success condition to 'the abstract'.

> Authors:Hu Cang

summary — pass: The title and summary follow the abstract's argument closely. They explain the chain of hypotheses, how disagreement-guided search works, and why a narrow feature missing from the starting data can be overlooked. They also report the 32-trial results and the conclusion. 'Closed-loop' is left unexplained, but this does not block understanding. The abstract's 'Surprisingly' framing is softened, but its substance is preserved. The question is a genuine open issue that follows from the narrow test setting.

> Asking the right question, rather than testing only where current models disagree, is therefore key to AI discovery in a chain of hypotheses.

date — pass: originalPublished is 2026-11-15, which matches the arXiv v1 submission date. The feed date of 2026-11-17 is later and correctly not used.

> \[Submitted on 15 Nov 2026\]

support — pass: Each claim is a verbatim-supported statement from the abstract. The trial count, the comparison with random search, the miss in all 32 trials, and the success condition are all accurate. The limitations note correctly says that only the abstract is available and that only the narrow-peak setting was tested. The draft invents no benefits, costs or forecasts.

> the agent that followed model disagreement more closely missed the peak in all 32 trials, whereas discovery succeeded when experiments probed the unmeasured gap early and followed up across the anomaly.

media — pass: The selected media is the verified direct link to the arXiv abstract page of this paper, which is the original 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/9f0ab7946a59f9e063c3d60707d5e5ce33e5cefb36390b79ecf4b94e2df384cf/references.json)

## Corrections and challenges

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

[JSON record](https://escapemeridian.com/readings/hu-cang-argues-asking-the-right-question-is-key-to-ai-discovery-9f0ab794.json)
