{"id":1014458,"date":"2026-06-23T16:53:16","date_gmt":"2026-06-23T19:53:16","guid":{"rendered":"https:\/\/www.psr-inc.com\/?post_type=analytics_post&#038;p=1014458"},"modified":"2026-06-23T18:12:36","modified_gmt":"2026-06-23T21:12:36","slug":"ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver","status":"publish","type":"analytics_post","link":"https:\/\/www.psr-inc.com\/en\/analytics-report\/post\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\/","title":{"rendered":"AI reasoning in math and coding: a stochastic generation-expansion solver"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">Introduction<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">AI reasoning in mathematics<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">For most of the deep-learning era, the relationship between large language models and mathematics was uneasy: systems that excelled at natural-language generation remained unreliable on multi-step reasoning and formal mathematics. That changed rapidly after 2024, as frontier systems reached Olympiad-level performance and began integrating directly with formal verification environments such as Lean.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">More interesting than the benchmark progression itself, however, is what this capability means for working mathematicians. Terence Tao has become one of the clearest voices articulating the emerging role of AI in mathematical research. In a January 2025 article in the <em>Notices of the American Mathematical Society<\/em> on machine-assisted proof, and later through the release of a flexible formalization of his textbook <em>Analysis I<\/em> in Lean, Tao argued not that language models had overcome mathematicians, but that they had become useful collaborators. Their value lies in reducing the overhead surrounding research: searching literature, checking standard arguments, formalizing definitions, writing small computational experiments, and testing whether an approach is worth pursuing. In 2023, Tao had already used ChatGPT alongside Lean in work related to formalizing aspects of a proof of the Polynomial Freiman\u2013Ruzsa conjecture.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The relevant shift is therefore not merely that AI systems can solve harder mathematical problems, but that they are increasingly capable of operating across formal reasoning, symbolic manipulation, code generation, and verification within a unified workflow.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">AI reasoning in code<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">The trajectory in software engineering is structurally similar to the one in mathematics, although more visible because its outputs are merged code rather than written proofs. The shift is from suggestion-style autocomplete toward agentic systems that inspect repositories, plan edits across multiple files, execute tests, and iterate on failures. Benchmarks such as SWE-bench Verified now evaluate these systems on real GitHub issues drawn from production repositories, and by 2026 frontier agents were already resolving a substantial fraction of repository-level software engineering tasks on previously unseen codebases.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Beyond benchmarks, the industrial signal is increasingly difficult to ignore. By 2026, coding agents had become embedded in the daily workflows of major engineering organizations for refactoring, debugging, migration, and test generation tasks. Anthropic executives stated publicly that Claude was already writing most of the company\u2019s internal code, while Google\u2019s 2025 DORA State of AI-Assisted Software Development report found AI adoption broadly associated with higher engineering throughput, with productivity depending as much on testing infrastructure and operational discipline as on the specific model used.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Read together, the trajectories in AI for mathematics and AI for code are converging. When supplied with both a research paper and a codebase, frontier systems can participate in a coupled reasoning-and-implementation loop: interpreting algorithms, identifying correctness conditions, translating them into production code, generating regression tests, and iterating through execution and verification. The remainder of this article reports a deliberate experiment in exactly such a workflow, with the human setting high-level objectives, approving implementation strategies, and intervening mainly when tests exposed algorithmic or convergence failures.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">The application: stochastic generation expansion<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">LightBDMM.jl is a PSR Julia package developed against this environment. It implements the accelerated Benders Decomposition with Multiple Masters (a-BDMM) algorithm that was introduced in the <a href=\"https:\/\/doi.org\/10.1109\/TPWRS.2022.3141993\">article<\/a>: <em>An Integrated Progressive Hedging and Benders Decomposition with Multiple Master Method to Solve the Brazilian Generation Expansion Problem<\/em> (A. Soares, A. Street, T. Andrade, and J. D. Garcia). The package structure, implementation, regression tests, documentation, and continuous-integration setup were generated by Claude Code, then iteratively corrected through execution, testing, debugging, and human review. Subsequent extensions \u2014 mixed-integer (MIP) first-stage decisions with quadratic-penalty linearization, and distributed execution over MPI were added by the same assistant.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The project was deliberately set up as a low-interaction experiment: how far can a frontier agentic system carry an applied mathematics implementation when the human author confines themselves to high-level objectives and end-state review, rather than per-step steering? Correctness during the cycle was anchored not on human review of intermediate steps, but on an independent ground-truth validation harness, which is described later.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">We already have implementations of Benders decomposition and Progressive Hedging in separate repositories with MPI parallelization and mixed-integer representation. These repositories were supplied to the coding agent as reference implementations for architectural patterns, testing procedures, and documentation structure.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The mathematical structure underneath stochastic generation expansion is the two-stage stochastic program: an investment decision today (which units to build, which lines to reinforce, which contracts to sign) is made before next year&#8217;s uncertainty (demand, hydrology, fuel prices, renewable output) is observed, an operational recourse decision is then taken once each scenario has materialized.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">The math-reasoning loop<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">The interesting aspect of this implementation was not symbolic mathematics or automated theorem proving. The challenge was translating a research paper into a working optimization package: understanding the algorithmic structure, planning the implementation, identifying convergence challenges, and iteratively correcting them through testing and validation. This experiment was only meaningful because correctness could be checked independently: every implementation was continuously validated against deterministic-equivalent solutions and benchmark behavior rather than judged only by whether the generated code appeared plausible.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The a-BDMM algorithm augments classical Benders along two axes. Multiple masters replace the single aggregated master with one master per scenario, each anchored on a different recourse problem and sharing a common cut pool, producing a tighter recourse approximation in fewer iterations. Progressive Hedging acceleration adds a quadratic consensus penalty that drives the per-scenario decisions toward a common value.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">What makes the algorithm a useful test of mathematical AI reasoning is that several of its correctness conditions are not obvious from the headline description and would be easy to miss in a naive re-implementation:<\/p>\n\n\n\n<ol start=\"1\" class=\"wp-block-list\">\n<li>A separate LP for the lower bound after PH activates. Once the consensus penalty is on, the algorithm&#8217;s lower bound must be computed from a master that contains only the Lagrangian term  <em>w<sub>s<\/sub><\/em> (<em>x<sub>s<\/sub><\/em> &#8211; <em>x<\/em>) \u00a0and not the quadratic penalty. The implementation maintains a second pool of lower-bound masters for exactly this purpose.<\/li>\n\n\n\n<li>If PH is initialized at a degenerate point, for example, with all components of <em>x<\/em>\u00a0set to zero, the algorithm may converge prematurely to a consensus solution that is nevertheless suboptimal.<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">Working through these issues required an iterative planning and debugging process. Before writing code, the assistant generated implementation plans, surfaced convergence risks, and proposed stabilization strategies. These observations still required human verification but surfacing them early reduced the likelihood of propagating conceptual mistakes into the implementation itself.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">The coding loop<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">The coding loop was the second stage of the process and followed the same low-interaction protocol. Claude Code generated the package structure, implementation, regression tests, documentation, and continuous integration setup, while the human author provided only high-level specifications at the beginning of each cycle and reviewed the final outputs. The important point was not that the first implementation was correct, but that the agent could keep the repository, algorithm, and test feedback in context while iterating toward a validated implementation.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The initial implementation already reproduced most of the algorithmic structure correctly, including the Benders decomposition workflow and the benchmark newsvendor test problem. The first major issue appeared in the cut aggregation logic: the implementation initially produced incorrect objective values because scenario contributions were not properly combined. After correcting the aggregation and weighting of cuts, the BDMM variant converged quickly and produced the expected solution.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The more interesting failure emerged in the accelerated a-BDMM variant. All state variables were initialized at zero, and the Progressive Hedging penalty terms immediately pushed the algorithm toward consensus around this uninformed initial point. As a result, the method became trapped in a suboptimal \u201cdo nothing\u201d solution and failed to improve. The eventual fix was to introduce a warmup phase in which the decomposition explored freely for several iterations before activating the consensus penalties. Once implemented, this stabilization strategy resolved the convergence issue on benchmark problems and matched the intuition that agreement mechanisms should not dominate before informative cuts have been generated.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The lower-bound logic, however, required a much more interactive debugging process. Once the PH quadratic penalties became active, the bound calculation no longer behaved consistently. Several implementation attempts failed in different ways: using the upper bound as a proxy produced artificial convergence in the first iteration; using the raw master objectives caused oscillatory lower bounds after PH activation; and subtracting the quadratic penalties directly sometimes generated invalid states with LB UB.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Unlike the warmup mechanism, which emerged with relatively little human intervention, the lower-bound phase required repeated interaction between the human author and the assistant. The debugging process eventually became less about software engineering and more about computational interpretation of the algorithm itself: determining which quantities remained valid lower bounds after PH activation, how convergence should be measured, and which parts of the augmented objective were optimization artifacts rather than components of the original stochastic program. The final implementation eventually recovered a correct lower-bound procedure described on the paper.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Overall, the debugging process was notable because most failures were not syntactic or software-engineering issues, but algorithmic challenges related to convergence behavior and the interaction between decomposition and consensus mechanisms.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Validation on a generation-expansion model<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The solver was exercised on a stochastic generation expansion problem \u2014 fifteen candidate plants, forty-five industrial customers, and fifty stochastic activation scenarios. Each method was run to a 0.1<del> <\/del>% relative gap at the same tolerance.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img fetchpriority=\"high\" decoding=\"async\" width=\"1003\" height=\"552\" src=\"https:\/\/www.psr-inc.com\/wp-content\/uploads\/2026\/06\/image-17.png\" alt=\"\" class=\"wp-image-1014172\" srcset=\"https:\/\/www.psr-inc.com\/wp-content\/uploads\/2026\/06\/image-17.png 1003w, https:\/\/www.psr-inc.com\/wp-content\/uploads\/2026\/06\/image-17-300x165.png 300w, https:\/\/www.psr-inc.com\/wp-content\/uploads\/2026\/06\/image-17-768x423.png 768w\" sizes=\"(max-width: 1003px) 100vw, 1003px\" \/><\/figure>\n\n\n\n<p class=\"legenda-padrao\">Figure 1 &#8211; Relative optimality gap as a function of iteration count for each algorithm on the generation\u2011expansion testbed. a\u2011BDMM collapses the gap in 5 iterations, against 8 for plain BDMM, 21 for multi\u2011cut Benders, and 177 for single\u2011cut Benders.<\/p><br\/>\n\n\n\n<p class=\"wp-block-paragraph\">The qualitative result of interest is the iteration count. a\u2011BDMM closes the bound gap in 5 iterations, against 177 iterations for classical single\u2011cut Benders on the same instance \u2014 a roughly 35\u2011fold reduction. Plain BDMM (without the Progressive Hedging acceleration) lies in between, at 8 iterations, isolating the contribution of the consensus penalty.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">The validation harness<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">In a low-interaction development protocol, the validation harness is not merely a quality-assurance tool; it is the mechanism that makes the workflow viable. Every regression test in the suite is solved both by LightBDMM.jl and by the monolithic deterministic equivalent \u2014 the exact extensive-form formulation of the stochastic program used as an independent reference solution. The results are required to match every benchmark instance, so the implementation is continuously checked against an external ground truth rather than only against its own internal logic. This serves as the numerical counterpart of the earlier reasoning loop: conditions identified during planning and debugging are revalidated computationally against a benchmark that shares none of LightBDMM.jl\u2019s implementation. Any conceptual or implementation drift introduced during the agentic coding cycle would therefore appear as a deterministic-equivalent mismatch before reaching production.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The full regression suite runs in continuous integration on every proposed code change. At the time of writing, the package maintains approximately 98% automated code coverage.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Outlook<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The experiment does not demonstrate autonomous mathematical engineering. It demonstrates something narrower but operationally important: when algorithms have strong mathematical structure, benchmark behavior, and independent validation procedures, agentic coding systems can substantially reduce the engineering cost of translating research methods into production implementations.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Two-stage stochastic optimization is one of the oldest and best-understood frameworks in operations research, and a-BDMM itself has existed in the literature since 2021. The relevant change is therefore not the novelty of the algorithm, but the reduction in the cost and time required to move from paper to tested software. In this case, the combination of mathematical structure, deterministic-equivalent validation, and AI-assisted development made it possible to build a production-quality implementation including decomposition variants, regression tests, benchmarking infrastructure, continuous integration, and documentation, all within a single development cycle.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">More broadly, mathematically structured domains with independent reference solutions appear especially compatible with agentic development workflows. When large parts of the translation from paper to production can be accelerated and continuously validated, organizations gain the ability to experiment with, operationalize, and deploy a much larger fraction of the optimization literature than was previously practical.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n","protected":false},"featured_media":0,"template":"","meta":{"_acf_changed":false},"report_section":[480],"class_list":["post-1014458","analytics_post","type-analytics_post","status-publish","hentry","report_section-indepth"],"acf":[],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.9 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>AI reasoning in math and coding: a stochastic generation-expansion solver - PSR Energy<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.psr-inc.com\/en\/analytics-report\/post\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\/\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"AI reasoning in math and coding: a stochastic generation-expansion solver - PSR Energy\" \/>\n<meta property=\"og:description\" content=\"Introduction AI reasoning in mathematics For most of the deep-learning era, the relationship between large language models and mathematics was uneasy: systems that excelled at natural-language generation remained unreliable on multi-step reasoning and formal mathematics. That changed rapidly after 2024, as frontier systems reached Olympiad-level performance and began integrating directly with formal verification environments such [&hellip;]\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.psr-inc.com\/en\/analytics-report\/post\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\/\" \/>\n<meta property=\"og:site_name\" content=\"PSR Energy\" \/>\n<meta property=\"article:publisher\" content=\"https:\/\/www.facebook.com\/psrenergy\" \/>\n<meta property=\"article:modified_time\" content=\"2026-06-23T21:12:36+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.psr-inc.com\/wp-content\/uploads\/2026\/06\/image-17.png\" \/>\n\t<meta property=\"og:image:width\" content=\"1003\" \/>\n\t<meta property=\"og:image:height\" content=\"552\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/png\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:site\" content=\"@psrenergy\" \/>\n<meta name=\"twitter:label1\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data1\" content=\"12 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/www.psr-inc.com\\\/en\\\/analytics-report\\\/post\\\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\\\/\",\"url\":\"https:\\\/\\\/www.psr-inc.com\\\/en\\\/analytics-report\\\/post\\\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\\\/\",\"name\":\"AI reasoning in math and coding: a stochastic generation-expansion solver - PSR Energy\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/www.psr-inc.com\\\/en\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/www.psr-inc.com\\\/en\\\/analytics-report\\\/post\\\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\\\/#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/www.psr-inc.com\\\/en\\\/analytics-report\\\/post\\\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/www.psr-inc.com\\\/wp-content\\\/uploads\\\/2026\\\/06\\\/image-17.png\",\"datePublished\":\"2026-06-23T19:53:16+00:00\",\"dateModified\":\"2026-06-23T21:12:36+00:00\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/www.psr-inc.com\\\/en\\\/analytics-report\\\/post\\\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\\\/#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/www.psr-inc.com\\\/en\\\/analytics-report\\\/post\\\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\\\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/www.psr-inc.com\\\/en\\\/analytics-report\\\/post\\\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\\\/#primaryimage\",\"url\":\"https:\\\/\\\/www.psr-inc.com\\\/wp-content\\\/uploads\\\/2026\\\/06\\\/image-17.png\",\"contentUrl\":\"https:\\\/\\\/www.psr-inc.com\\\/wp-content\\\/uploads\\\/2026\\\/06\\\/image-17.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/www.psr-inc.com\\\/en\\\/analytics-report\\\/post\\\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/www.psr-inc.com\\\/en\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Posts do Analytics Report\",\"item\":\"https:\\\/\\\/www.psr-inc.com\\\/en\\\/analytics-report\\\/posts\\\/\"},{\"@type\":\"ListItem\",\"position\":3,\"name\":\"AI reasoning in math and coding: a stochastic generation-expansion solver\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/www.psr-inc.com\\\/en\\\/#website\",\"url\":\"https:\\\/\\\/www.psr-inc.com\\\/en\\\/\",\"name\":\"PSR\",\"description\":\"\",\"publisher\":{\"@id\":\"https:\\\/\\\/www.psr-inc.com\\\/en\\\/#organization\"},\"alternateName\":\"PSR Energy Consulting\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/www.psr-inc.com\\\/en\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"en-US\"},{\"@type\":\"Organization\",\"@id\":\"https:\\\/\\\/www.psr-inc.com\\\/en\\\/#organization\",\"name\":\"PSR\",\"alternateName\":\"PSR Energy Consulting\",\"url\":\"https:\\\/\\\/www.psr-inc.com\\\/en\\\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/www.psr-inc.com\\\/en\\\/#\\\/schema\\\/logo\\\/image\\\/\",\"url\":\"https:\\\/\\\/www.psr-inc.com\\\/wp-content\\\/uploads\\\/2023\\\/03\\\/logo-psr.svg\",\"contentUrl\":\"https:\\\/\\\/www.psr-inc.com\\\/wp-content\\\/uploads\\\/2023\\\/03\\\/logo-psr.svg\",\"width\":1056,\"height\":816,\"caption\":\"PSR\"},\"image\":{\"@id\":\"https:\\\/\\\/www.psr-inc.com\\\/en\\\/#\\\/schema\\\/logo\\\/image\\\/\"},\"sameAs\":[\"https:\\\/\\\/www.facebook.com\\\/psrenergy\",\"https:\\\/\\\/x.com\\\/psrenergy\",\"https:\\\/\\\/www.instagram.com\\\/psrenergy\\\/\",\"https:\\\/\\\/www.linkedin.com\\\/company\\\/psrenergy\\\/\"]}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"AI reasoning in math and coding: a stochastic generation-expansion solver - PSR Energy","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.psr-inc.com\/en\/analytics-report\/post\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\/","og_locale":"en_US","og_type":"article","og_title":"AI reasoning in math and coding: a stochastic generation-expansion solver - PSR Energy","og_description":"Introduction AI reasoning in mathematics For most of the deep-learning era, the relationship between large language models and mathematics was uneasy: systems that excelled at natural-language generation remained unreliable on multi-step reasoning and formal mathematics. That changed rapidly after 2024, as frontier systems reached Olympiad-level performance and began integrating directly with formal verification environments such [&hellip;]","og_url":"https:\/\/www.psr-inc.com\/en\/analytics-report\/post\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\/","og_site_name":"PSR Energy","article_publisher":"https:\/\/www.facebook.com\/psrenergy","article_modified_time":"2026-06-23T21:12:36+00:00","og_image":[{"width":1003,"height":552,"url":"https:\/\/www.psr-inc.com\/wp-content\/uploads\/2026\/06\/image-17.png","type":"image\/png"}],"twitter_card":"summary_large_image","twitter_site":"@psrenergy","twitter_misc":{"Est. reading time":"12 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.psr-inc.com\/en\/analytics-report\/post\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\/","url":"https:\/\/www.psr-inc.com\/en\/analytics-report\/post\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\/","name":"AI reasoning in math and coding: a stochastic generation-expansion solver - PSR Energy","isPartOf":{"@id":"https:\/\/www.psr-inc.com\/en\/#website"},"primaryImageOfPage":{"@id":"https:\/\/www.psr-inc.com\/en\/analytics-report\/post\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\/#primaryimage"},"image":{"@id":"https:\/\/www.psr-inc.com\/en\/analytics-report\/post\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\/#primaryimage"},"thumbnailUrl":"https:\/\/www.psr-inc.com\/wp-content\/uploads\/2026\/06\/image-17.png","datePublished":"2026-06-23T19:53:16+00:00","dateModified":"2026-06-23T21:12:36+00:00","breadcrumb":{"@id":"https:\/\/www.psr-inc.com\/en\/analytics-report\/post\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\/#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.psr-inc.com\/en\/analytics-report\/post\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\/"]}]},{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/www.psr-inc.com\/en\/analytics-report\/post\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\/#primaryimage","url":"https:\/\/www.psr-inc.com\/wp-content\/uploads\/2026\/06\/image-17.png","contentUrl":"https:\/\/www.psr-inc.com\/wp-content\/uploads\/2026\/06\/image-17.png"},{"@type":"BreadcrumbList","@id":"https:\/\/www.psr-inc.com\/en\/analytics-report\/post\/ai-reasoning-in-math-and-coding-a-stochastic-generation-expansion-solver\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/www.psr-inc.com\/en\/"},{"@type":"ListItem","position":2,"name":"Posts do Analytics Report","item":"https:\/\/www.psr-inc.com\/en\/analytics-report\/posts\/"},{"@type":"ListItem","position":3,"name":"AI reasoning in math and coding: a stochastic generation-expansion solver"}]},{"@type":"WebSite","@id":"https:\/\/www.psr-inc.com\/en\/#website","url":"https:\/\/www.psr-inc.com\/en\/","name":"PSR","description":"","publisher":{"@id":"https:\/\/www.psr-inc.com\/en\/#organization"},"alternateName":"PSR Energy Consulting","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.psr-inc.com\/en\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"},{"@type":"Organization","@id":"https:\/\/www.psr-inc.com\/en\/#organization","name":"PSR","alternateName":"PSR Energy Consulting","url":"https:\/\/www.psr-inc.com\/en\/","logo":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/www.psr-inc.com\/en\/#\/schema\/logo\/image\/","url":"https:\/\/www.psr-inc.com\/wp-content\/uploads\/2023\/03\/logo-psr.svg","contentUrl":"https:\/\/www.psr-inc.com\/wp-content\/uploads\/2023\/03\/logo-psr.svg","width":1056,"height":816,"caption":"PSR"},"image":{"@id":"https:\/\/www.psr-inc.com\/en\/#\/schema\/logo\/image\/"},"sameAs":["https:\/\/www.facebook.com\/psrenergy","https:\/\/x.com\/psrenergy","https:\/\/www.instagram.com\/psrenergy\/","https:\/\/www.linkedin.com\/company\/psrenergy\/"]}]}},"_links":{"self":[{"href":"https:\/\/www.psr-inc.com\/en\/wp-json\/wp\/v2\/analytics_post\/1014458","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.psr-inc.com\/en\/wp-json\/wp\/v2\/analytics_post"}],"about":[{"href":"https:\/\/www.psr-inc.com\/en\/wp-json\/wp\/v2\/types\/analytics_post"}],"wp:attachment":[{"href":"https:\/\/www.psr-inc.com\/en\/wp-json\/wp\/v2\/media?parent=1014458"}],"wp:term":[{"taxonomy":"report_section","embeddable":true,"href":"https:\/\/www.psr-inc.com\/en\/wp-json\/wp\/v2\/report_section?post=1014458"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}