ASTOR: Benchmarking and Improving Autoformulation for Soft-Constraint Optimization with a Formal Grammar
We introduce ASTOR, an inference-time framework that maps a natural-language decision problem to a grammar-guided abstract syntax tree (AST) before deterministically compiling the resulting representation into executable optimization code.
ProofPaper ↗
Key points
- Real-world decision problems combine two qualitatively different kinds of constraints: requirements that must be satisfied, called hard constraints, and preferences that can be violated, called soft constraints.
- Because these problems are often posed to decision-makers in natural language, large language models offer a natural tool for assisting with their solution.
- This has given rise to autoformulation, which focuses on using LLMs to translate natural-language descriptions of decision problems into executable mathematical optimization models.
- We also introduce SoftOR, a unified benchmark of natural language problems which focuses on soft constraint formulations.
Sources (1)
- [1]ASTOR: Benchmarking and Improving Autoformulation for Soft-Constraint Optimization with a Formal GrammararXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 05:52 PM
We introduce ASTOR, an inference-time framework that maps a natural-language decision problem to a grammar-guided abstract syntax tree (AST) before deterministically compiling the resulting representation into executable optimization code.
Real-world decision problems combine two qualitatively different kinds of constraints: requirements that must be satisfied, called hard constraints, and preferences that can be violated, called soft constraints.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 19, 2026Introducing Clairic, a new System One class of extraction models
- Oct 19, 2026Push-based KV transfer to improve TTFT and TPOT for disaggregated inference in vLLM
- Oct 19, 2026Beyond Behavioral Accuracy: Evaluating Individual-Level Fidelity in LLM Investor Simulation
- Oct 19, 2026From Black Box to Controller: Inference-Time Steering of LLM via Feature Amplification
- Oct 17, 2026Refuse the Harm, Complete the Task via Graph-of-Rules Planning for LLM-Enabled Robots
- Oct 17, 2026EdgeKernelBench: Unified LLM-Driven Kernel Generation for Edge Backends