tue09/awesome-reasoning-generalization

[](https://awesome.re) [](https://github.com/tue09/awesome-reasoning-generalization/commits/main) Progress in LLM reasoning is commonly measured on benchmarks that stay close to the training distribution. This list collects work on the question the survey asks: when and how can large language models generalize their reasoning beyond their training distributions, under changes in notation, problem structure, length, difficulty, domain, language, or tool interface? The companion text is in survey.md, and the paper is in fullpaper/. A reasoning-generalization result can be interpreted only when it states what the model has encountered, what changes at test time, and which components remain fixed. Table 5 of the survey lists the minimum report. Gains enabled by prompting, search, verification, or memory need not hold for the model alone: a tool-augmented system can solve longer problems even when the base model has not learned a length-general algorithm. Greedy accuracy shows what a model produces reliably, large-pass@k accuracy what it can produce occasionally, and…