pith. machine review for the scientific record. sign in

arxiv: 2604.11535 · v2 · submitted 2026-04-13 · 💻 cs.AI

Recognition: unknown

Problem Reductions at Scale: Agentic Integration of Computationally Hard Problems

Xi-Wei Pan , Shi-Wen An , Jin-Guo Liu
Authors on Pith no claims yet

Pith reviewed 2026-05-10 16:03 UTC · model grok-4.3

classification 💻 cs.AI
keywords problem reductionsNP-hard problemsAI coding agentsharness engineeringpolynomial-time reductionsverification systemsoptimization solverstransitive composition
0
0 comments X

The pith

Harness engineering lets AI coding agents build a verified library of 200+ polynomial-time reductions between hard problems.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to demonstrate that a structured harness of constraints, multilayer verification, and automated pipelines can direct AI agents to produce a large, reliable collection of reductions at a speed and scale not previously achieved. A sympathetic reader would care because such a library would let any supported NP-hard problem be routed to any compatible solver through a single interface, with new solvers instantly becoming available across all connected problems due to transitive composition. The authors report constructing a command-line tool backed by 100+ problem types and 200+ rules in over 170k lines of Rust within three months, using no-code expert input, type checks, and agentic feature tests that simulate end users.

Core claim

Harness engineering, defined as the deliberate design of constraints, verification systems, and feedback loops, enables AI coding agents to build a command-line tool supported by a library of more than 100 problem types and 200 reduction rules totaling over 170,000 lines of Rust code. The resulting reduction graph composes transitively, so that registering any solver for a single problem type makes that solver available to every problem reachable through a chain of reductions.

What carries the argument

Harness engineering: the practice of designing constraints, verification systems from type checks to agentic feature tests, and automated pipelines to channel AI coding agents into producing correct reductions.

If this is right

  • Registering a solver for any one problem type makes it available to every problem connected by a reduction path.
  • Domain experts can add new problem types or rules through a no-code route without writing implementation code.
  • Any supported hard optimization problem can be reformulated and sent to any supported solver via a single command-line interface.
  • New quantum hardware or commercial optimizers integrate across the entire connected problem space without additional per-problem work.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same harness pattern could be tested on other large-scale formal software tasks, such as generating verified compilers or proof assistants.
  • If the approach scales further, it could shorten the time between discovering a new solver and making it usable for arbitrary connected problems.
  • Transitive reduction graphs might uncover previously unnoticed efficient routes between seemingly distant problem classes.

Load-bearing premise

The multilayer verification stack is assumed to detect every error so that all generated reductions are correct polynomial-time transformations without requiring manual review of each rule.

What would settle it

An audit that finds even one generated reduction that either maps solutions incorrectly or fails to run in polynomial time would show the harness does not reliably produce correct reductions at the reported scale.

Figures

Figures reproduced from arXiv: 2604.11535 by Jin-Guo Liu, Shi-Wen An, Xi-Wei Pan.

Figure 1
Figure 1. Figure 1: The reduction graph. Each node is a problem type; each directed edge [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The integration problem. (a) a contributor’s new reduction does not fit [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Library architecture. Two layers stack: interfaces (top) and infrastruc [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: CLI using example: creating and solving a Maximum Independent [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Skills architecture. Blue boxes are advisor skills (human in the loop); white boxes are automation skills (fully autonomous). Both are executed by agents, but humans are more involved when advisor skills are loaded. traditional automation (Makefiles, shell scripts), which chains deterministic commands, skills can include abstract steps: “implement the reduction algorithm”(run-pipeline skill in [PITH_FULL_… view at source ↗
Figure 6
Figure 6. Figure 6: Delegation spectrum. Tasks are sorted from fully human (top) to fully [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Contribution pipeline as a state machine on the GitHub project board. Each box is a board column (state); arrows show the single allowed transition [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Excerpts from the add-rule skill. The checklist encodes the mathematical knowledge every reduction requires. Reference implementations prevent convention drift. The paper entry ensures visual verifiability by human reviewers. report surfaces usability gaps, documentation omissions, and semantic errors that unit tests cannot reach. In our practice, agentic feature tests caught most CLI command bugs, and som… view at source ↗
Figure 10
Figure 10. Figure 10: Canonical examples as a single source of truth. The contributor’s [PITH_FULL_IMAGE:figures/full_fig_p007_10.png] view at source ↗
Figure 13
Figure 13. Figure 13: Growth of problem types and reduction rules over time, mined from [PITH_FULL_IMAGE:figures/full_fig_p007_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Trait hierarchy and compile-time validation. The [PITH_FULL_IMAGE:figures/full_fig_p010_14.png] view at source ↗
read the original abstract

Solving an NP-hard optimization problem often requires reformulating it for a specific solver -- quantum hardware, a commercial optimizer, or a domain heuristic. A tool for polynomial-time reductions between hard problems would let practitioners route any supported problem to any supported solver through a single interface. Building such a library at scale, however, has remained out of reach. We show that harness engineering, the practice of designing constraints, verification systems, and feedback loops that channel AI coding agents, can overcome this barrier. Our harness combines a no-code contribution route for domain experts, a multilayer verification stack ranging from type-level checks to agentic feature tests (AI agents role-playing as end users), and a fully automated implementation-review-integration pipeline. In about three months, we built a command-line tool backed by a library of 100+ problem types and 200+ reduction rules in over 170k lines of Rust. The result suggests that a well-engineered harness lets agents build well-tested software at a scale and pace beyond prior reduction-library efforts. Because the reduction graph composes transitively, a new solver registered for any single problem type instantly becomes available to every problem connected by a reduction path. The source code is available at https://github.com/CodingThrust/problem-reductions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript describes the use of harness engineering to direct AI coding agents in building a comprehensive library of polynomial-time reductions between over 100 types of computationally hard problems. The authors report constructing a command-line tool and supporting Rust library with 200+ reduction rules in approximately three months, employing a multilayer verification system from type checks to agentic feature tests. The system enables transitive composition, allowing solvers registered for one problem type to be used for others via reduction paths. The source code is publicly available on GitHub.

Significance. Should the reductions prove correct and the verification stack reliable, this work would provide a practical tool for routing NP-hard problems to appropriate solvers without custom reduction design for each case. It also offers a case study in scaling AI agent contributions to complex software projects through structured constraints and automated pipelines. The provision of the full source code and the scale achieved (170k lines) are notable strengths supporting reproducibility and further development.

major comments (2)
  1. [Abstract and Introduction] The abstract and introduction provide no concrete example of any reduction rule, such as a specific instance mapping from one problem to another together with verification that yes/no answers are preserved and runtime is polynomial. This omission is load-bearing for the central claim of having successfully constructed 200+ correct reductions.
  2. [Verification stack description] In the section describing the multilayer verification stack, the combination of type-level checks and agentic feature tests is presented as sufficient to ensure correctness, but no mechanism is detailed for detecting non-polynomial operations or semantic mismatches in the reduction mappings. Agentic tests can miss such errors, undermining the transitive composition guarantee.
minor comments (1)
  1. [Overall] A summary table or diagram of the reduction graph, showing connectivity between problem types and the number of rules per type, would improve clarity on the library's coverage.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight areas where the manuscript can better support its central claims. We respond to each major comment below and will incorporate revisions as indicated.

read point-by-point responses

  1. Referee: [Abstract and Introduction] The abstract and introduction provide no concrete example of any reduction rule, such as a specific instance mapping from one problem to another together with verification that yes/no answers are preserved and runtime is polynomial. This omission is load-bearing for the central claim of having successfully constructed 200+ correct reductions.

    Authors: We agree that the absence of a concrete example in the abstract and introduction weakens support for the claim of 200+ correct reductions. The repository contains the full library with tests, but the manuscript text does not present one. We will revise the introduction to include a specific example (e.g., a 3-SAT to Vertex Cover reduction with instance mapping, answer-preservation argument, and polynomial-time verification), along with references to the corresponding code, type checks, and agentic tests. revision: yes

  2. Referee: [Verification stack description] In the section describing the multilayer verification stack, the combination of type-level checks and agentic feature tests is presented as sufficient to ensure correctness, but no mechanism is detailed for detecting non-polynomial operations or semantic mismatches in the reduction mappings. Agentic tests can miss such errors, undermining the transitive composition guarantee.

    Authors: The referee correctly notes that the current description does not detail mechanisms for non-polynomial operations or semantic mismatches beyond the stated layers. Our stack relies on type constraints for signatures, agentic tests on sample instances for functional behavior, and human review in the pipeline, but these do not exhaustively catch all possible issues. We will revise the verification section to explicitly describe the polynomiality checks (type-level plus targeted profiling), acknowledge the limitations of agentic tests for semantic mismatches, and clarify that transitive composition holds only for reductions passing the full pipeline with documented oversight. revision: yes

Circularity Check

0 steps flagged

No circularity: engineering construction with external artifacts

full rationale

The paper presents an engineering project for building a reduction library via harness engineering and AI agents, with no mathematical derivations, equations, fitted parameters, or predictions. Claims rest on the constructed artifact (170k lines of Rust, 100+ types, 200+ rules) and the linked GitHub repository as external verification, satisfying the self-contained criterion. No load-bearing steps reduce to self-definition, self-citation chains, or renamed inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The work rests on standard complexity theory and software verification practices rather than new postulates.

axioms (1)
  • standard math Polynomial-time reductions between NP-hard problems preserve hardness and allow solver reuse
    Invoked implicitly when stating that reductions let any solver become available to connected problems.

pith-pipeline@v0.9.0 · 5528 in / 1169 out tokens · 68452 ms · 2026-05-10T16:03:16.620854+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

64 extracted references · 18 canonical work pages · 3 internal anchors

  1. [1]

    Korte and J

    B. Korte and J. Vygen,Combinatorial optimization: theory and algo- rithms. Springer, 2008

    2008

  2. [2]

    Combinatorial optimization with physics-inspired graph neural networks,

    M. J. Schuetz, J. K. Brubaker, and H. G. Katzgraber, “Combinatorial optimization with physics-inspired graph neural networks,”Nature Ma- chine Intelligence, vol. 4, no. 4, pp. 367–377, 2022

    2022

  3. [3]

    Airline crew scheduling: State- of-the-art,

    B. Gopalakrishnan and E. L. Johnson, “Airline crew scheduling: State- of-the-art,”Annals of Operations Research, vol. 140, no. 1, pp. 305–337, 2005

    2005

  4. [4]

    Register allocation via coloring,

    G. J. Chaitin, M. A. Auslander, A. K. Chandra, J. Cocke, M. E. Hopkins, and P. W. Markstein, “Register allocation via coloring,”Computer languages, vol. 6, no. 1, pp. 47–57, 1981

    1981

  5. [5]

    Toth and D

    P. Toth and D. Vigo,Vehicle routing: problems, methods, and applica- tions. SIAM, 2014

    2014

  6. [6]

    R. M. Karp,Reducibility among Combinatorial Problems. Boston, MA: Springer US, 1972, pp. 85–103. [Online]. Available: https: //doi.org/10.1007/978-1-4684-2001-2 9

    work page doi:10.1007/978-1-4684-2001-2 1972

  7. [7]

    M. R. Garey and D. S. Johnson,Computers and Intractability: A Guide to the Theory of NP-Completeness. W. H. Freeman & Co., 1990

    1990

  8. [8]

    An extensible SAT-solver,

    N. E ´en and N. S ¨orensson, “An extensible SAT-solver,” inInternational conference on theory and applications of satisfiability testing. Springer, 2003, pp. 502–518

    2003

  9. [9]

    CaDiCaL, Gimsatul, IsaSAT and Kissat entering the SAT Competition 2024,

    A. Biere, T. Faller, K. Fazekas, M. Fleury, N. Froleyks, and F. Pollitt, “CaDiCaL, Gimsatul, IsaSAT and Kissat entering the SAT Competition 2024,” inProc. of SAT Competition 2024 – Solver, Benchmark and Proof Checker Descriptions, ser. Department of Computer Science Report Series B, M. Heule, M. Iser, M. J ¨arvisalo, and M. Suda, Eds., vol. B-2024-1. Univ...

    2024

  10. [10]

    IBM ILOG CPLEX optimization studio,

    IBM, “IBM ILOG CPLEX optimization studio,” 2026, accessed: 2026-04-13. [Online]. Available: https://www.ibm.com/products/ ilog-cplex-optimization-studio

    2026

  11. [11]

    [Online]

    Gurobi Optimization, LLC, “Gurobi,” 2026, accessed: 2026-04-13. [Online]. Available: https://www.gurobi.com

    2026

  12. [12]

    Parallelizing the dual revised simplex method,

    Q. Huangfu and J. J. Hall, “Parallelizing the dual revised simplex method,”Mathematical Programming Computation, vol. 10, no. 1, pp. 119–142, 2018

    2018

  13. [13]

    Improved approximation algo- rithms for maximum cut and satisfiability problems using semidefinite programming,

    M. X. Goemans and D. P. Williamson, “Improved approximation algo- rithms for maximum cut and satisfiability problems using semidefinite programming,”Journal of the ACM (JACM), vol. 42, no. 6, pp. 1115– 1145, 1995

    1995

  14. [14]

    Quantum bridge analytics I: a tutorial on formulating and using QUBO models,

    F. Glover, G. Kochenberger, and Y . Du, “Quantum bridge analytics I: a tutorial on formulating and using QUBO models,”4or, vol. 17, no. 4, pp. 335–371, 2019

    2019

  15. [15]

    Ising formulations of many NP problems,

    A. Lucas, “Ising formulations of many NP problems,”Frontiers in physics, vol. 2, p. 74887, 2014

    2014

  16. [16]

    Quantum optimization for maximum independent set using Rydberg atom arrays,

    H. Pichler, S.-T. Wang, L. Zhou, S. Choi, and M. D. Lukin, “Quantum optimization for maximum independent set using Rydberg atom arrays,”

  17. [17]

    Quantum Optimization for Maximum Independent Set Using Rydberg Atom Arrays

    [Online]. Available: https://arxiv.org/abs/1808.10816

    work page Pith review arXiv

  18. [18]

    Quantum optimization of maximum independent set using Rydberg atom arrays,

    S. Ebadi, A. Keesling, M. Cain, T. T. Wang, H. Levine, D. Bluvstein, G. Semeghini, A. Omran, J.-G. Liu, R. Samajdaret al., “Quantum optimization of maximum independent set using Rydberg atom arrays,” Science, vol. 376, no. 6598, pp. 1209–1215, 2022

    2022

  19. [19]

    Quantum annealing and GNN for solving TSP with QUBO,

    H. He, “Quantum annealing and GNN for solving TSP with QUBO,” in Algorithmic Aspects in Information and Management (AAIM). Springer, 2024, pp. 134–145

    2024

  20. [20]

    Harness engineering: Leveraging Codex in an agent- first world,

    OpenAI, “Harness engineering: Leveraging Codex in an agent- first world,” 2026, accessed: 2026-04-13. [Online]. Available: https: //openai.com/index/harness-engineering/

    2026

  21. [21]

    SWE-EVO: Benchmarking Coding Agents in Long-Horizon Software Evolution Scenarios

    M. V . T. Thai, T. Le, D. N. Manh, H. P. Nhat, and N. D. Q. Bui, “SWE- EVO: Benchmarking coding agents in long-horizon software evolution scenarios,” 2026. [Online]. Available: https://arxiv.org/abs/2512.18470

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  22. [22]

    SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks?

    X. Deng, J. Da, E. Pan, Y . Y . He, C. Ide, K. Garg, N. Lauffer, A. Park, N. Pasari, C. Rane, K. Sampath, M. Krishnan, S. Kundurthy, S. Hendryx, Z. Wang, V . Bharadwaj, J. Holm, R. Aluri, C. B. C. Zhang, N. Jacobson, B. Liu, and B. Kenstler, “SWE-Bench Pro: Can AI agents solve long-horizon software engineering tasks?” 2025. [Online]. Available: https://ar...

    work page internal anchor Pith review arXiv 2025

  23. [23]

    Agent skills in the sdk,

    Claude, “Agent skills in the sdk,” 2026, accessed: 2026-04-13. [Online]. Available: https://code.claude.com/docs/en/agent-sdk/skills

    2026

  24. [24]

    Exact algorithms for maximum indepen- dent set,

    M. Xiao and H. Nagamochi, “Exact algorithms for maximum indepen- dent set,”Information and Computation, vol. 255, pp. 126–146, 2017

    2017

  25. [25]

    A framework for exponential-time-hypothesis–tight algorithms and lower bounds in geometric intersection graphs,

    M. de Berg, H. L. Bodlaender, S. Kisfaludi-Bak, D. Marx, and T. C. van der Zanden, “A framework for exponential-time-hypothesis–tight algorithms and lower bounds in geometric intersection graphs,”SIAM Journal on Computing, vol. 49, no. 6, pp. 1291–1331, 2020

    2020

  26. [26]

    Optimization, approximation, and complexity classes,

    C. L. Lucchesi and S. L. Osborn, “Candidate keys for relations,” Journal of Computer and System Sciences, vol. 17, no. 2, pp. 270–279, 1978. [Online]. Available: https://www.sciencedirect.com/ science/article/pii/0022000078900090

    work page arXiv 1978

  27. [27]

    An exact method for the minimum feedback arc set problem,

    A. Baharev, H. Schichl, A. Neumaier, and T. Achterberg, “An exact method for the minimum feedback arc set problem,”ACM J. Exp. Algorithmics, vol. 26, Apr. 2021. [Online]. Available: https://doi.org/10.1145/3446429

    work page doi:10.1145/3446429 2021

  28. [28]

    Progressive disclosure,

    J. Nielsen, “Progressive disclosure,” Nielsen Norman Group,

  29. [29]

    Available: https://www.nngroup.com/articles/ progressive-disclosure/

    [Online]. Available: https://www.nngroup.com/articles/ progressive-disclosure/

  30. [30]

    GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models

    I. Mirzadeh, K. Alizadeh, H. Shahrokhi, O. Tuzel, S. Bengio, and M. Farajtabar, “GSM-Symbolic: Understanding the limitations of mathematical reasoning in large language models,” 2025. [Online]. Available: https://arxiv.org/abs/2410.05229

    work page Pith review arXiv 2025

  31. [31]

    The illusion of thinking: Understanding the strengths and limitations of reasoning models via the lens of problem complexity,

    P. Shojaee, I. Mirzadeh, K. Alizadeh, M. Horton, S. Bengio, and M. Farajtabar, “The illusion of thinking: Understanding the strengths and limitations of reasoning models via the lens of problem complexity,”

  32. [32]

    The illusion of thinking: Understanding the strengths and limitations of reasoning models via the lens of problem complexity.arXiv preprint arXiv:2506.06941,

    [Online]. Available: https://arxiv.org/abs/2506.06941

    work page arXiv

  33. [33]

    Faith and fate: Limits of transformers on compositionality,

    N. Dziri, X. Lu, M. Sclar, X. L. Li, L. Jiang, B. Y . Lin, S. Welleck, P. West, C. Bhagavatula, R. Le Braset al., “Faith and fate: Limits of transformers on compositionality,”Advances in neural information processing systems, vol. 36, pp. 70 293–70 332, 2023

    2023

  34. [34]

    FrontierMath: A benchmark for evaluating advanced mathematical reasoning in AI.arXiv preprint arXiv:2411.04872, 2024

    E. Glazer, E. Erdil, T. Besiroglu, D. Chicharro, E. Chen, A. Gunning, C. F. Olsson, J.-S. Denain, A. Ho, E. de Oliveira Santos, O. J ¨arviniemi, M. Barnett, R. Sandler, M. Vrzala, J. Sevilla, Q. Ren, E. Pratt, L. Levine, G. Barkley, N. Stewart, B. Grechuk, T. Grechuk, S. V . Enugandla, and M. Wildon, “FrontierMath: A benchmark for evaluating advanced math...

    work page arXiv 2025

  35. [35]

    A benchmark of expert-level academic questions to assess AI capabilities,

    Center for AI Safety, Scale AI & HLE Contributors Consortium, “A benchmark of expert-level academic questions to assess AI capabilities,” Nature, vol. 649, no. 8099, pp. 1139–1146, 2026

    2026

  36. [36]

    The expressive power of transformers with chain of thought, 2024

    W. Merrill and A. Sabharwal, “The expressive power of transformers with chain of thought,” 2024. [Online]. Available: https://arxiv.org/abs/ 2310.07923

    work page arXiv 2024

  37. [37]

    Superpowers: A plugin for building better AI coding agents,

    J. Vincentet al., “Superpowers: A plugin for building better AI coding agents,” 2025, accessed: 2026-04-13. [Online]. Available: https://github.com/obra/superpowers

    2025

  38. [38]

    [Online]

    OpenAI, “Codex,” 2026, accessed: 2026-04-13. [Online]. Available: https://openai.com/codex/

    2026

  39. [39]

    Claude Code,

    Anthropic, “Claude Code,” 2025, accessed: 2026-04-13. [Online]. Available: https://github.com/anthropics/claude-code

    2025

  40. [40]

    Programming guide for solving constraint satisfaction problems with tensor networks,

    X. Gao, X. Li, and J. Liu, “Programming guide for solving constraint satisfaction problems with tensor networks,”Chinese Physics B, vol. 34, no. 5, p. 050201, 2025

    2025

  41. [41]

    Nguyen, J.-G

    M.-T. Nguyen, J.-G. Liu, J. Wurtz, M. D. Lukin, S.-T. Wang, and H. Pichler, “Quantum optimization with arbitrary connectivity using Rydberg atom arrays,”PRX Quantum, vol. 4, p. 010316, Feb 2023. [On- line]. Available: https://link.aps.org/doi/10.1103/PRXQuantum.4.010316

    work page doi:10.1103/prxquantum.4.010316 2023

  42. [42]

    Encoding computationally hard problems in triangular Rydberg atom arrays,

    X.-W. Pan, H.-H. Zhou, Y .-M. Lu, and J.-G. Liu, “Encoding computationally hard problems in triangular Rydberg atom arrays,”

  43. [43]

    Available: https://arxiv.org/abs/2510.25249

    [Online]. Available: https://arxiv.org/abs/2510.25249

    work page arXiv

  44. [44]

    qubogen: QUBO matrix generator for major combinatorial optimization problems written in Python,

    Y . Tamura and M. Sakai, “qubogen: QUBO matrix generator for major combinatorial optimization problems written in Python,” 2020, accessed: 2026-04-13. [Online]. Available: https://github.com/tamuhey/qubogen

    2020

  45. [45]

    Karp: a language for NP reductions,

    C. Zhang, J. D. Hartline, and C. Dimoulas, “Karp: a language for NP reductions,” inProceedings of the 43rd ACM SIGPLAN International Conference on Programming Language Design and Implementation, ser. PLDI 2022. New York, NY , USA: Association for Computing Machinery, 2022, p. 762–776. [Online]. Available: https://doi.org/10.1145/3519939.3523732

    work page doi:10.1145/3519939.3523732 2022

  46. [46]

    Automatic evaluation of reductions between NP-complete problems,

    C. Creus, P. Fern ´andez, and G. Godoy, “Automatic evaluation of reductions between NP-complete problems,” inTheory and Applications of Satisfiability Testing – SAT 2014. Cham: Springer International Publishing, 2014, pp. 415–421

    2014

  47. [47]

    PyQUBO: Python library for mapping combinatorial optimization problems to QUBO form,

    M. Zaman, K. Tanahashi, and S. Tanaka, “PyQUBO: Python library for mapping combinatorial optimization problems to QUBO form,”IEEE Transactions on Computers, vol. 71, no. 4, pp. 838–850, 2021

    2021

  48. [48]

    qubovert: a Python package for formulating, simulating, and solving problems in boolean and spin form,

    J. T. Iosue, “qubovert: a Python package for formulating, simulating, and solving problems in boolean and spin form,” 2022, accessed: 2026-04-13. [Online]. Available: https://github.com/jtiosue/qubovert

    2022

  49. [49]

    QUBO.jl: A Julia ecosystem for quadratic unconstrained binary optimization,

    P. M. Xavier, P. Ripper, T. Andrade, J. D. Garcia, N. Maculan, and D. E. B. Neira, “QUBO.jl: A Julia ecosystem for quadratic unconstrained binary optimization,” 2023. [Online]. Available: https: //arxiv.org/abs/2307.02577

    work page arXiv 2023

  50. [50]

    D-Wave documentation,

    D-Wave Systems, “D-Wave documentation,” 2026, accessed: 2026-04-

    2026

  51. [51]

    Available: https://docs.ocean.dwavesys.com

    [Online]. Available: https://docs.ocean.dwavesys.com

  52. [52]

    Comput- ing solution space properties of combinatorial optimization problems via generic tensor networks,

    J.-G. Liu, X. Gao, M. Cain, M. D. Lukin, and S.-T. Wang, “Comput- ing solution space properties of combinatorial optimization problems via generic tensor networks,”SIAM Journal on Scientific Computing, vol. 45, no. 3, pp. A1239–A1270, 2023

    2023

  53. [53]

    Swe-agent: Agent-computer interfaces enable automated soft- ware engineering,

    J. Yang, C. E. Jimenez, A. Wettig, K. Lieret, S. Yao, K. Narasimhan, and O. Press, “Swe-agent: Agent-computer interfaces enable automated soft- ware engineering,”Advances in Neural Information Processing Systems, vol. 37, pp. 50 528–50 652, 2024

    2024

  54. [54]

    Introducing Devin, the first AI software en- gineer,

    S. Wu, “Introducing Devin, the first AI software en- gineer,” Cognition Blog, Mar. 2024. [Online]. Available: https://cognition.ai/blog/introducing-devin

    2024

  55. [55]

    OpenHands: An Open Platform for AI Software Developers as Generalist Agents

    X. Wang, B. Li, Y . Song, F. F. Xu, X. Tang, M. Zhuge, J. Pan, Y . Song, B. Li, J. Singh, H. H. Tran, F. Li, R. Ma, M. Zheng, B. Qian, Y . Shao, N. Muennighoff, Y . Zhang, B. Hui, J. Lin, R. Brennan, H. Peng, H. Ji, and G. Neubig, “OpenHands: An open platform for AI software developers as generalist agents,” 2025. [Online]. Available: https://arxiv.org/ab...

    work page internal anchor Pith review arXiv 2025

  56. [56]

    CoRR , volume =

    C. S. Xia, Z. Wang, Y . Yang, Y . Wei, and L. Zhang, “Live-SWE-agent: Can software engineering agents self-evolve on the fly?” 2025. [Online]. Available: https://arxiv.org/abs/2511.13646

    work page arXiv 2025

  57. [57]

    SWE-bench: Can language models resolve real- world github issues?

    C. E. Jimenez, J. Yang, A. Wettig, S. Yao, K. Pei, O. Press, and K. R. Narasimhan, “SWE-bench: Can language models resolve real- world github issues?” inThe Twelfth International Conference on Learning Representations, 2024. [Online]. Available: https: //openreview.net/forum?id=VTF8yNQM66

    2024

  58. [58]

    LongCLI-Bench: A preliminary benchmark and study for long-horizon agentic programming in command-line interfaces,

    Y . Feng, J. Sun, Z. Yang, J. Ai, C. Li, Z. Li, F. Zhang, K. He, R. Ma, J. Lin, J. Sun, Y . Xiao, S. Zhou, W. Wu, Y . Liu, P. Liu, Y . Qiao, S. Zhang, and K. Zhang, “LongCLI-Bench: A preliminary benchmark and study for long-horizon agentic programming in command-line interfaces,” 2026. [Online]. Available: https://arxiv.org/abs/2602.14337

    work page arXiv 2026

  59. [59]

    State of AI vs. human code generation report,

    CodeRabbit, “State of AI vs. human code generation report,” 2025, accessed: 2026-04-13. [Online]. Available: https://www.coderabbit.ai/ whitepapers/state-of-AI-vs-human-code-generation-report

    2025

  60. [60]

    Measuring the impact of early-2025 ai on experienced open-source developer productivity,

    J. Becker, N. Rush, E. Barnes, and D. Rein, “Measuring the impact of early-2025 ai on experienced open-source developer productivity,”

    2025

  61. [61]

    Measuring the impact of early-2025 AI on experienced open-source developer productivity.CoRR, abs/2507.09089, 2025

    [Online]. Available: https://arxiv.org/abs/2507.09089

    work page arXiv

  62. [62]

    Cursor used a swarm of AI agents powered by OpenAI to build and run a web browser for a week—with no human help. here’s why developers are buzzing,

    S. Goldman, “Cursor used a swarm of AI agents powered by OpenAI to build and run a web browser for a week—with no human help. here’s why developers are buzzing,” Jan. 2026, accessed: 2026-04-13. [Online]. Available: https://fortune.com/2026/01/ 23/cursor-built-web-browser-with-swarm-ai-agents-powered-openai/

    2026

  63. [63]

    We analyzed the code of Cursor’s ai-built browser FastRender,

    W. Heijstek, “We analyzed the code of Cursor’s ai-built browser FastRender,” Jan. 2026, accessed: 2026-04-13. [Online]. Available: https://www.softwareimprovementgroup.com/blog/quality-of-fastrender/

    2026

  64. [64]

    Harness engineering for coding agent users,

    B. B ¨ockeler, “Harness engineering for coding agent users,” 2026, accessed: 2026-04-13. [Online]. Available: https://martinfowler.com/ articles/exploring-gen-ai/harness-engineering.html APPENDIXA SYSTEMARCHITECTURE The library’s type system reduces the space of possible agent errors by making incorrect code fail to compile. This appendix describes the ke...

    2026