What is this?
AutoLOGIC is an agentic framework for automated tabular machine learning across classification, regression, and clustering. It turns a task specification into coordinated feature, model, optimization, and output-control plans and then executes them with task-specific tools.
AutoLOGIC is an agentic framework for automated tabular machine learning across classification, regression, and clustering. It turns a task specification into coordinated feature, model, optimization, and output-control plans and then executes them with task-specific tools.Is this a ready-to-use tool or a research framework I script against?It is a research framework: you describe a job as a TaskSpec config and launch the matching Python runner with --task-spec, not a point-and-click app.
It is a research framework: you describe a job as a TaskSpec config and launch the matching Python runner with --task-spec, not a point-and-click app.How do I get a task running?Set your dataset paths and the LLM endpoint through environment variables, then run the task-specific runner with --task-spec pointing at your config file.
Set your dataset paths and the LLM endpoint through environment variables, then run the task-specific runner with --task-spec pointing at your config file.Do I need an LLM API to use this?Yes, the AgenticReasoningAgent relies on a configured LLM endpoint, set alongside your data paths via environment variables.
Yes, the AgenticReasoningAgent relies on a configured LLM endpoint, set alongside your data paths via environment variables.Is this production-ready or more of a research artifact?It reads as research code built to accompany a paper: it ships separate reproduce/fig4, fig5, fig6 scripts to regenerate figures and depends on legacysage/compat benchmark modules, rather than being packaged for production use.
It reads as research code built to accompany a paper: it ships separate reproduce/fig4, fig5, fig6 scripts to regenerate figures and depends on legacysage/compat benchmark modules, rather than being packaged for production use.What are the four plan channels the agent coordinates?It coordinates feature engineering, model construction, optimization, and output-control plans, each revised independently based on execution feedback.
It coordinates feature engineering, model construction, optimization, and output-control plans, each revised independently based on execution feedback.What kinds of outputs can a task produce?Depending on the task, AutoLOGIC can produce stacking, distillation, calibration, prediction-interval, or clustering outputs.
Depending on the task, AutoLOGIC can produce stacking, distillation, calibration, prediction-interval, or clustering outputs.How does the agent decide whether to change its plan?After execution, validation scores, failures, retries, and fallbacks come back as observations, and the agent records a retain, revise, or fallback transition for the affected plan channel.
After execution, validation scores, failures, retries, and fallbacks come back as observations, and the agent records a retain, revise, or fallback transition for the affected plan channel.Can I reproduce the paper's figures with this repo?Yes, the reproduce/fig4, fig5, and fig6 folders contain the benchmark execution, calibration, distillation, and clustering scripts used to generate those figures.
Yes, the reproduce/fig4, fig5, and fig6 folders contain the benchmark execution, calibration, distillation, and clustering scripts used to generate those figures.What are similar projects?sktime, DeepAnalyze, rrsi, strands-decider, autogluon, RD-Agent, homemade-machine-learning, lime
sktime, DeepAnalyze, rrsi, strands-decider, autogluon, RD-Agent, homemade-machine-learning, limeWhich topics is it in?automl, clustering, tabular-data, classification, agentic, data-science, benchmark, framework, research, agent, machine-learning, automation
automl, clustering, tabular-data, classification, agentic, data-science, benchmark, framework, research, agent, machine-learning, automation