← All open problems

THEORIA / CATALOG SOURCES

Expert questions.
Independent sources.

The collection combines original papers, named-author problem lists and structured source imports. References attach to one canonical question wherever the scope matches.

In use · editorial research targets

Foundations of AI/ML — learning theory and post-training literature ↗

119 challenges supported by 316 references across 28 subjects, including efficient architectures, Kaplan–Chinchilla scaling, AB-MCTS, KANs, logistic regression/NTK, time-series/tabular learning, differential privacy, jailbreak resistance, homomorphic computation and symbolic ODE/PDE discovery. Discovery combines Undermind scholarly search with primary arXiv, ACL, PMLR, IACR and lab pages; notes distinguish abstracts, inspected full texts and research agendas.

Sixteen entries reformulate published questions, including seven explicit bound gaps; 103 are editorial extensions. Forty-four records state an efficiency payoff. Existing results and subsequent resolutions are recorded; provisional status is not expert certification. Coding-agent access starts at /llms.txt.

In use · explicit bound questions

Learning bounds — sample, model and collaboration complexity ↗

Seven additions define teaching-set size, agnostic multiclass sample complexity, instance-optimal identity testing, sequential calibration, shallow-network encoding size, persuasive collaboration efficiency and fixed-step cyclic self-play.

Models, quantifiers, known bounds and resolution criteria are explicit. The latest realizable PAC and unrestricted ReLU depth claims are treated as baselines; residual logarithmic, encoding-size and minimax gaps are kept separate. Search also includes related full-text matches.

In use · scoped agent questions

Multi-agent systems — latent messages and information graphs ↗

Weighted strategic collaboration has a published exponent gap. Hidden-state message learning and fault-tolerant information-graph design are concrete editorial extensions linked to existing cache-transfer and game-theoretic results.

Charge payload bits, adapter learning and communication edges. Distinguish task-relevant information from raw vector capacity and pure Nash guarantees from worst coarse correlated equilibria under failures.

In use · explicit research directions

Lokhov and collaborators — interaction screening and Ising learning ↗

Two explicit directions concern computationally sufficient moments and provable neural graph recovery. Foundational interaction screening, GRISE, dynamical learning and strongly metastable sampling define the established baselines for four inference targets.

NeurISE and computational-statistics discussion sections were inspected, together with the strong-metastability assumptions. Compact neural approximation is distinct from finite-training structure recovery; statistical sufficiency is distinct from computational and measurement cost.

In use · scoped safety challenges

AI safety & alignment — oversight and assurance literature ↗

Ten primary challenges cover hidden backdoors, debate, tamper resistance, safe exploration, abstention, unlearning, prompt injection, control, reward hacking and weak-to-strong learning. Six related targets are cross-listed from verification, privacy, synthetic-world transfer, inverse RL and robotic controller certification.

Existing impossibility results and positive guarantees define the scope: cryptographic backdoor undetectability, fallible-judge limitations, conditional-coverage impossibility and Lean-verified LLMbda noninterference are baselines. Every safety claim needs a stated threat, access and observation model.

In use · named-author challenges

COLT 2026 — published learning-theory open problems ↗

Three scoped entries draw on Feldman–Kamath–Srebro (standard SGD versus linear dimension), Arvanitakis–Srinivas–Vijayaraghavan (tensor ALS overparameterization), and Nissim–Stemmer–Tsfadia (approximate-private PAC sample complexity). The problem statements and assumptions were checked in the PDFs.

The standard-SGD target excludes the separate SQ question negatively resolved by Patel in September 2026. The privacy entry incorporates Moran–Yan's October near-linear Littlestone bound. Published open labels are qualified by these later results.

In use · named-author challenges

COLT 2025 — data selection and equilibrium learning ↗

Hanneke and collaborators' fixed-learner regression selection profile and Maiti's instance-dependent approximate-Nash identification problem. Their PDF formulations were inspected and compared with 2026 follow-ups.

The regression target excludes solved recovery thresholds and small cells. Equilibrium sample complexity is distinguished from expected estimator bias and Nash regret. The separate sample-compression entry targets linear size beyond the November 2026 near-quadratic preprint.

In use · editorial research questions

Quantitative finance — primary literature ↗

Nine scoped questions and research targets in volatility, martingale mimicking and transport, market impact, HJB computation, financial networks, Epstein–Zin optimization and endogenous efficiency under long-memory order flow. Each detail page links original papers and records which abstracts or statements were checked.

Status notes incorporate restricted positive results, multidimensional counterexamples, the withdrawn rough-Heston uniqueness claim, and 2026 preprints. The long-memory research program starts with an audit of established scaling results. Broad research directions are marked as needing a framework; preprint claims are attributed rather than independently certified.

In use · expert problem list

IQOQI Vienna — Open Quantum Problems ↗

Named-author questions in entanglement, quantum computation, cryptography and thermodynamics. New entries are editorial formulations linked to the numbered problem. The separate solved list and model-specific partial results are checked during selection.

Problem statements and scoped additions; dates on the site are not a complete current-literature audit.

In use · scoped mission targets

Prove2Me — selected mathematical-physics goals ↗

Five editorial records cover the restricted-three-body retrograde global-section problem, separate Thomson optimality targets for 9, 10 and 32 charges, and ballistic-deposition KPZ variance scaling. Sources identify each mission and the precise target inspected; the KPZ mission is tagged Probability on Prove2Me.

An unfinished Lean formalization is not necessarily open research. We checked the selected natural-language/Lean goals, separated known subcases and cited recent partial results. Completed missions may prove a narrower statement than their title; no site-wide completion or review badge is imported as mathematical solved status.

In use · reviewed discovery catalogue

Shao-Jiang Wang — 100 theoretical-physics problems ↗

A named-author, explicitly AI-assisted list spanning gravity, cosmology, QFT and quantum many-body physics. All 100 items were compared with the existing catalogue: 55 new scoped research targets, 27 mapped occurrences and 18 review holds.

New statements are editorial reformulations with numbered source links. The source's status date is August 31, 2026; review checks scope and overlap, not all current literature or proofs. Source-listed arXiv references are labelled as unverified bibliography leads. No bulk-reuse license is asserted.

In use · primary literature

Research papers and reviews ↗

Original conjectures and reviews from arXiv and journals, including operator growth, ETH, random waves, holography, QFT and integrable systems. For example, Parker and collaborators' operator-growth hypothesis is attached to the existing Lanczos-growth record.

Stable paper versions and section/equation locators where inspected. Reading metadata or an abstract is labelled separately from reading the statement.

In use · official problem statements

Clay Mathematics Institute ↗

Authoritative formulations of Yang–Mills existence and mass gap, plus background for the selected Riemann-zero connections. Yang–Mills is cross-listed under QFT and QCD without duplicating its identity.

Official formulations and scope; a general mass-gap question is distinct from Wilson-loop confinement.

In use · community review

Snowmass Theory Frontier ↗

A research roadmap spanning quantum field theory, particle theory, amplitudes, lattice theory, string theory and quantum gravity.

Bibliography and scoped research directions. A roadmap mention alone does not establish an unresolved mathematical statement.

Discovery source · historical

IAMP / Aizenman mathematical-physics problems ↗

Expert-contributed problems assembled in 1998–99 in statistical mechanics, quantum many-body systems, spectral theory and dynamics. Individual pages can contain useful precise models and assumptions.

Requires a modern follow-up check before publication. The historical open labels are not imported wholesale.

Discovery source · not imported

Physics Open Problems — jgmcm ↗

An 84-item mathematical-physics catalog with model statements and bibliographies; its status assessment is dated July 2026.

Discovery and citation leads. Bulk reuse terms and later status changes must be checked before importing source text.

In use · source import

PubPhys ↗

The public API supplies the broad candidate pool. Source-imported records retain their original statements, revision hashes and CC BY 4.0 attribution; they are labelled separately from editorially scoped entries.

Full candidate snapshot, reproducible theory/computation screening, reversible duplicate mappings and explicit status-review holds.

How to read the labels

“Editorially scoped” means the statement and assumptions were formulated for this catalog using cited sources. “Source imported” preserves a screened source record. Neither label means expert-verified open status. A reference check records what was read, not an audit of all subsequent literature.

Source publications keep their own licenses. PubPhys source text is attributed under CC BY 4.0; independently sourced entries are editorial formulations with links, not bulk copies of third-party text.