problem_id,statement,categories,why_unresolved,testable_direction,supporting_keys
G-P1,"Can field-normalised citation ranking identify landmark works in small theory categories, given identifier merges, re-deposited classics and generic-title chapters?",algorithms;computational_complexity;formal_verification;compilers,The 150 candidates contain 29 in-scope research works; NCS36 rewards records whose citation counts belong to other works or to canonical textbooks.,"Re-run select_landmarks.py with type filters (exclude book-chapter, dissertation, reprint), a live identifier re-resolution step and venue priors (STOC/FOCS/SODA/POPL/PLDI/CAV/ICSE); measure precision against a specialist-labelled sample.",czerwinski2019petri;chen2022maxflow;kelley2023threeap
G-P2,Does co-classification with the LLM category in software engineering (21.2% of works in 2025) correspond to method adoption or to vocabulary adoption?,software_engineering;programming_languages,Co-classification is lexical; reference flows to LLM works stay at 2.5% of software-engineering references.,Extract method/model nodes for the high-priority set (Part III knowledge graph) and test whether works co-classified with LLM cite LLM method papers at a higher rate than their category baseline.,wang2024llmtesting;fan2023llm4se;hou2024llm4se
G-P3,What is the productivity effect of AI coding assistants on experienced developers working on mature codebases?,software_engineering,Randomised field experiments report +26% completed tasks for enterprise developers while a 2025 randomised trial preprint reports slower completion for expert open-source maintainers; populations and outcomes differ.,"Preregistered multi-site trials stratified by developer experience and repository maturity, with time-to-merge and defect rate as joint outcomes.",cui2026genaiwork;becker2025metr;peng2023copilot
G-P4,Can LLM-based bug finding and compiler fuzzing be given soundness or coverage guarantees comparable to translation validation and differential testing?,programming_languages;compilers;formal_verification,"LLift and WhiteFox report found bugs but, per their abstracts, no completeness or false-negative bounds; Alive2 gives bounded translation validation.","Combine LLM-generated test programs with Alive2-style validation and measure the share of LLM-reported bugs that are confirmed, by optimisation pass.",li2024llift;yang2024whitefox;lopes2021alive2;xia2024fuzz4all
G-P5,"How far can learned provers scale on formal mathematics and verified software, and are reported pass rates comparable across benchmarks?",formal_verification,MizAR reports 60% of Mizar theorems; LLM provers report Lean/Isabelle figures from preprints on different benchmarks with possible training-set overlap.,A benchmark with time-separated test theorems (post-training-cutoff) and a common compute budget; report proof-checking rather than self-reported success.,jakubuv2023mizar;first2023baldur;yang2023leandojo;xin2024deepseekprover;yang2024formalmath
G-P6,"Do the 2022-2025 algorithmic breakthroughs (almost-linear flow, near-linear negative-weight SSSP, sub-Dijkstra directed SSSP, square-root-space simulation) translate into practice, and does the corpus record their diffusion?",algorithms;computational_complexity,None of these works is in the landmark candidate list; their citation diffusion across categories has not been measured.,Trace the citation neighbourhoods of the verified DOIs in the citation graph and classify citing works by category and type (theory vs engineering).,chen2022maxflow;bernstein2022negative;duan2025sorting;williams2025sqrtspace;farachcolton2024openaddressing
G-P7,"Is the metaphor-based metaheuristic literature growing faster than the rest of algorithms research, and does it absorb a growing share of citations?",algorithms,Six of 25 algorithm candidates are such optimisers; the critique literature argues many are re-descriptions of known methods; no corpus-level measurement exists.,"Keyword-defined subseries ('optimizer', 'nature-inspired', 'bio-inspired') with ITS and citation-share tracking, and a sample re-evaluation against the Camacho-Villalon criteria.",sorensen2013metaphor;aranha2021metaphor;camachovillalon2022exposing
G-P8,"What do expressivity bounds on transformers imply for the computational-complexity category, and are such works classified there?",computational_complexity,The TC0 bounds and formal-language survey are classified by OpenAlex topics that may route them to NLP rather than complexity.,Query the classification of the verified DOIs and measure the inflow of LLM-category citations to complexity works over time.,merrill2023parallelism;strobl2024survey
G-P9,"Are AI-discovered algorithms (matrix multiplication, sorting, cap sets) independently reproduced and extended by the theory community?",algorithms,Reports come from the developers (Nature articles and a preprint); no replication appears in the verified bibliography.,Search the citation graph for works that reproduce or improve the reported constructions; classify by author affiliation (independent vs developer).,fawzi2022alphatensor;mankowitz2023alphadev;romeraparedes2024funsearch;novikov2025alphaevolve
G-P10,Why do the formal verification and programming-languages categories decline in primary share while their multi-label counts grow?,formal_verification;programming_languages,Primary-label assignment is an argmax over category scores; works that also carry AI or security labels may lose the primary label to larger categories.,Compare primary and multi-label trajectories by co-assigned category; estimate the share of works whose primary label switched to llm/cybersecurity by year.,