A community resource · Information theory

AI-assisted research,
from formulation
to articulation.

How was the question proposed? How did AI help solve it? How was the result verified—and explained for human understanding?

Our starting point

A paper, and an
open conversation.

From Formulation to Articulation:
Considerations for AI-Assisted Research in Information Theory

The paper examines documented cases of AI assistance and introduces a framework for describing the research process. It is the starting collection for this project. We invite new cases, corrections, and updates to the evidence as this area develops.

Our focus is the mathematical contribution, the roles of researchers and AI systems, the checks supporting the claims, and the effort to make the ideas understandable.

Open the starting paper (PDF) ↗

The PDF contains the revised case studies, comparison table, acknowledgments, and full references. LaTeX source and BibTeX references are also available. The overview below summarizes the framework.

The classification framework

Four stages.
Five separate classifications.

Each code answers a different question about the process. The codes do not combine into a single score of research quality, autonomy, or intellectual credit.

F

Problem formulation

Who identified and materially shaped the consequential research question?

F-0 Human-formulated
A researcher supplies a sufficiently specific theorem, conjecture, construction, or counterexample problem.
F-1 AI-refined
AI proposes a consequential subproblem, reduction, conjecture, or reformulation that materially changes a human-supplied direction.
F-2 AI-originated
Under broad domain guidance, AI identifies the question and gives a substantive account of its novelty and significance before it is selected.
S

AI-mediated solution

How did researchers interact with the AI system while developing the result?

S-0 Direct prompting
The reported process consists of a prompt and a candidate idea, construction, or proof, possibly with brief follow-up exchanges.
S-1 Interactive collaboration
Researchers and AI repeatedly exchange ideas, objections, computations, and revised arguments, with human feedback guiding development.
S-2 Harnessed search or agentic workflow
AI operates within an automated iteration loop involving execution, evaluators, search, external tools, multiple agents, or proof-state tracking.

Insufficient disclosure: use a dash (—) when the evidence does not support a specific classification. Naming a model does not establish its interaction mode; incomplete disclosure alone does not justify S-0.

Two kinds of evidence

Verification needs both sides of the argument.

A numerical certificate establishes a finite obligation. The analytical reasoning connecting that obligation to the theorem requires its own verification.

A

Analytical reasoning

A-0 No analytical verification
The indispensable analytical reasoning has not been completely checked by human experts or a proof assistant.
A-1 Author verification
The authors check or reconstruct the complete reasoning chain. Responsibility for a manuscript alone is not evidence of a complete check.
A-2 Independent expert verification
A qualified expert outside the proof-generation process checks the particular argument and its indispensable reductions.
A-3 Formal verification
The analytical argument is formalized and checked in Lean or another proof assistant, relative to a stated trusted base.
N

Numerical certification

N-0 No numerical proof obligation
The proof is analytical, or numerical work is exploratory and carries no logical burden for the result.
N-1 Ordinary numerical evidence
Floating-point experiments, Monte Carlo results, or plots support a claim without a rigorous enclosure.
N-2 Rigorous finite certificate
Exact arithmetic, rigorous interval or ball arithmetic, an exhaustive finite check, or a checkable witness establishes a specified finite obligation.
N-3 Independent replay or formal integration
The certificate is reproduced independently, or its checker and connection to the theorem are formally verified.

Overall verification follows the indispensable steps. An A-0/N-2 claim remains provisional when the analytical reduction is unchecked. An independent proof corroborates a theorem statement without verifying another derivation. Formalization also requires checking that the formal statement faithfully represents the theorem in the paper.

T

Translation and articulation

How were the ideas and results developed into an account for human understanding?

T-0 Machine-led presentation
AI produces the final exposition, with human involvement limited to direction, feedback, or a brief final review.
T-1 Author-revised AI exposition
Authors substantively review, edit, reorganize, and clarify AI-originated material. Human-directed AI rewriting also belongs here when authors guide and review the resulting exposition.
T-2 Human-led articulation
Researchers construct or substantially reorganize the exposition around the ideas, techniques, interpretations, significance, and limitations.

Classify the relevant component

A key lemma, a construction, and the rest of a proof may have different histories. Describe distinct components when possible; use a paper-level account when the collaboration is too intertwined to separate meaningfully.

Explain the character of the work

Distinguish conceptual reasoning from search and optimization. A human/AI percentage is not required. The qualitative remarks distinguish reasoning-dominant, mixed, and computation-dominant work, based on expert judgment of discovery and proof; they are not numerical shares.

Keep origin, checking, and explanation distinct

Human verification does not change who developed an argument. Clear exposition does not establish correctness, and formal correctness does not by itself provide an explanation of the ideas.

The initial collection · Open to new cases

Cases from the
starting paper.

The 24 entries below reproduce Table 1 of the revised paper. They were identified through an AI-assisted search of arXiv preprints listed in cs.IT and posted before October 30, 2026, with the latest versions before that date assessed. Two manuscripts from OpenAI’s public mathematics collection are included as exceptions. The selection is illustrative, not exhaustive; future additions and corrections will follow review of public evidence.

Table 1. Comparison of AI-assisted information-theoretic results
ResultFSANTRemark
Gaussian monotonicity counterexample [3]F-0——N-2T-2*Reasoning-dominant
Fisher-information counterexample [5]F-0——N-0T-2*Reasoning-dominant
Multiple unicast: Reed–Solomon construction [6]F-0—A-1N-2—Reasoning-dominant
Multiple unicast: reversible circuits [8]F-0—A-3N-0—Reasoning-dominant
DSBS channel conjectures [9]F-0—A-3N-0 / N-3—Reasoning-dominant
Deletion-channel upper bound [10]F-0—A-3N-3—Mixed
Recursive zero-error constructions [11]F-0——N-2—Mixed
Odd-cycle capacity constructions [12]F-0——N-2T-2Computation-dominant
Bernoulli affine-subspace entropy [13]F-0S-1A-1N-0T-2Reasoning-dominant
Marton inner-bound counterexample [15]F-0S-1A-1 / A-2N-2T-2*Reasoning-dominant
Gaussian MAC: corner points [16]F-0S-1*A-1N-0T-2*Reasoning-dominant
Gaussian MAC: sum-rate face [17]F-0S-1*A-1N-0T-2*Reasoning-dominant
Exact-repair regenerating codes [18]F-0S-1A-1N-2T-1Mixed
Rogers–Shephard entropy inequalities [19]F-0S-1*A-1N-0T-1*Reasoning-dominant
Deletion-correcting codes [20]F-0S-2—N-0 / N-2T-2*Computation-dominant
Deletion-channel capacity approximation [21]F-0S-2A-0*N-2T-0Mixed
Fourier entropy of threshold functions [24]F-0S-2———Reasoning-dominant
Courtade–Kumar: entropy production and spectral arguments [25]F-0——N-2T-2*Reasoning-dominant
Courtade–Kumar: differential equations and Bellman inequalities [26]F-0S-1 / S-2A-3N-3T-2*Mixed
Courtade–Kumar: entropy-flow proof [27]F-0S-1*A-3N-0T-0*Reasoning-dominant
Multibit constructions and conjecture [27]F-0S-1*—N-0 / N-2 / N-1T-2*Mixed
Weak Hellinger inequality [28]F-0S-1A-3N-3T-2Mixed
OpenAI collection: binary-information contraction [2, 29] Lean scopeF-0—A-3—T-0—
OpenAI collection: full Hellinger argument [2]F-0———T-0—

Reference numbers match the accompanying paper. A dash means insufficient information for an assignment; an asterisk marks an inference based on the authors’ expert judgment, and reasonable assessments may differ. Remarks describe the qualitative roles of reasoning and computation during discovery and proof, not human/AI contributions or verification rigor. “Mixed” indicates substantial contributions from both modes without implying equal shares. Inclusion is not an independent verification of the results.

How to interpret claim-specific classifications

For the multibit extension, N-0 applies to the analytical counterexamples for every k ≥ 8, N-2 to exact enumeration within a specified seven-bit family, and N-1 to broader evidence motivating the small-output conjecture. Weindel and Heckel’s one-deletion theorem is analytical (N-0), while finite checks support other constructions (N-2). Pichler’s results include proofs without numerical obligations (N-0) and one with formally integrated computational certificates (N-3).

The Marton entry’s A-1 / A-2 reflects uncertainty about independence of the expert checking. The deletion-channel capacity approximation’s A-0* is an inference about the proof as released, based on the author’s public account. An unspecified A-level does not establish that analytical verification was absent. The T-2* assignment for Chen et al. concerns the accompanying expository account. The T classification does not itself establish analytical verification.

N-0 does not rule out computation during discovery. Formal verification establishes the formalized claim relative to its trusted base; the correspondence to the theorem in the paper still needs checking. See the paper’s comparison and discussion for the basis and scope of these assessments.

Help grow this collection. Suggest another case or share evidence that updates an existing entry.

Help extend the collection

Bring a case.
Show the evidence.

We welcome published papers, preprints, repositories, and public author accounts documenting substantive AI assistance in information-theoretic research.

  1. Check for an existing discussion.

    Search the repository’s issues for the paper title, authors, or arXiv identifier. Add new evidence to an existing issue when appropriate.

  2. Open “Suggest a research case.”

    Sign in to GitHub and use the submission button below, or choose Issues → New issue in the repository. No code changes are needed.

  3. Describe the result and AI contribution.

    Provide the full reference, public source links, and the specific theorem, lemma, construction, or other component. Point to the AI-use disclosure.

  4. Report the verification evidence.

    Explain who checked the analytical steps and what the numerical checks establish. Suggest classifications if the evidence supports them; “not documented” is a useful answer.

  5. Discuss and refine.

    Maintainers review the sources and may request clarification before incorporating a case into the collection or a future revision. An issue submission is not an endorsement of correctness.

Submissions use the hosting repository’s public GitHub Issues. In a downloaded preview, open that repository and select Issues → New issue. The submission links activate automatically on a standard GitHub Pages address.

Acknowledgments

The authors thank Amin Gohari, Chandra Nair, Hessam Mahdavifar, and Vincent Tan for their comments and suggestions, which helped improve the manuscript.