GitHub Repos advanced 3 min read Aug 3, 2026
Public Preview Sign in free for the full digest →

Machine Learning for Combinatorial Optimization Papers

“The repo added performance charts in June 2026 to show how far neural solvers have come — and across all 178 benchmark rows, the classical solvers everyone actually runs in production appear exactly zero times.”

Machine Learning for Combinatorial Optimization Papers
1 Views
0 Likes
0 Bookmarks
Source · github.com

“"A rudimentary, parameter-free heatmap based on the intrinsic k-nearest neighbor structure...can match or surpass the performance of more sophisticated, learned heatmaps." — Pan, Wang, Ying, Xue & Yu, "Beyond the Heatmap: A Rigorous Evaluation of Component Impact in MCTS-Based T...”

You know that feeling when you want to know whether a research area has actually produced anything usable, and the only maps available are hand-typed markdown lists that stopped being updated two years ago? Combinatorial optimization has that problem badly: papers land at ICLR, NeurIPS, ICML, AAAI, and AAMAS across 36 distinct problem types, and no single venue tracks them. The direct-format competitor omargup/neural_combinatorial_optimization stopped at 14 stars and has not been pushed since 2020-08-14 — the standard failure mode. awesome-ml4co addresses the maintenance half of that problem; it does not yet address the 'is any of this better than the classical solver' half.

combinatorial-optimizationmachine-learningoperations-researchpaper-listresearchbenchmarksgraph-neural-networks

Think of it like a recipe book that gets typeset automatically instead of being edited page by page. Contributors add a row to data/papers.csv with the fields category, title, publisher, year, type, link, authors, code, then run src/generator.py, which rewrites the whole 164,470-byte README from that CSV. In June 2026 the team added a second data file, data/benchmark_results.csv — 178 rows, 22 columns — that records reported gap and runtime for TSP methods, and the same generator renders SVG charts into assets/benchmarks/ and embeds them into the problem sections. The clever part is what the data layer makes possible: because every benchmark row carries source_paper_title, source_link, pdf_url, source_table, and extraction_confidence, you can trace any plotted point back to the exact table in the exact paper it came from. The unclever part is that the method column contains only learned solvers, so the chart measures progress inside the neural cohort and nothing outside it.

01
CSV-as-database — you can grep, join, sort, and diff the entire literature set instead of scraping markdown. papers.csv is 121,090 bytes; benchmark_results.csv is 51,193 bytes across 178 rows and 22 columns.
02
Per-row benchmark provenance — every plotted point carries source_paper_title, source_link, pdf_url, source_table, extraction_confidence, and notes, so you can verify any number against the table it was lifted from rather than trusting the...
03
36 problem categories in one index — TSP, job shop scheduling, vehicle routing, graph matching, mixed integer programming, electronic design automation, optimal power flow, and 29 more. You stop maintaining a private bibliography per probl...
04
Auto-generated README — contributors edit papers.csv and rerun generator.py, so you never resolve a markdown merge conflict in a 164KB file, and formatting stays consistent across five years of additions.
05
Disclosed benchmark caveats in the README itself — it states runtime is normalized to 1000 instances by instance count only and hardware is not normalized, which tells you exactly what the chart cannot answer.
06
Same-day PR merges when the maintainer is active — ricgama's commits and their merge all landed 2026-07-19, and terranovafr's PR #43 merged 2026-05-21, the same day as the commit.
07
Zero open issues across 41 contributors — only 2 issues have ever been opened in five years, both closed, because contribution runs entirely through pull requests from paper authors adding their own work.
Who it’s for

If you are a researcher or ML engineer tracking neural combinatorial optimization and you are tired of grepping conference proceedings by hand, this is the highest-coverage index that exists and the CSVs alone justify the clone. It also suits anyone building tooling on top of the literature — the machine-readable layer is the real product. It is not for you if you need to decide whether to put a learned solver into production: the benchmark file contains no LKH, Concorde, or Gurobi rows, so it cannot answer whether any charted method beats the solver you already run.

Worth exploring

Worth cloning for the data layer, not for the charts. The index itself has run since 2021-03-21 with 41 contributors and a last push on 2026-07-19, and the CSV workflow is demonstrably live — recent commit messages include 'Add recent papers (ICLR/AAAI/AAMAS 2026) and regenerate README'. Treat the June 2026 benchmark charts as what the README calls them, a Beta feature: hardware is not normalized, the only distribution is uniform_random_2d_unit_square, and there is no classical baseline to anchor against.

Developer playbook
Tech stack, code snippet, sentiment, alternatives.
PM playbook
Adoption angles, user fit, positioning.
CEO playbook
Traction signals, ROI, build vs buy.
Deep-dive insight
Full long-form analysis, no fluff.
Easy mode
Core idea, fast — when you need the gist.
Pro mode
Technical nuance, edge cases, tradeoffs.
Read the full digest
Go beyond the preview

Deep-dive insight, Easy and Pro modes, plus action playbooks — the full breakdown is one tap away.

Underrated tools. Unfiltered takes.

Read the full digest in the Snaplyze app for deep-dive insight, Easy and Pro modes, and the playbooks you can actually use.

Install Snaplyze →