AIAny
Icon for item

Ask Before You Optimize: Dynamic Pre-Formulation Clarification for Interactive Optimization

Studies when and how an assistant should ask clarification questions before converting natural-language problem descriptions into optimization models. Introduces OR-Clarify, a benchmark for pre-formulation clarification, and InterOPT, a two-stage framework that diagnoses formulation-critical gaps and guides targeted questioning and stopping decisions.

Introduction

Incomplete natural-language briefs often omit facts that change the structure of an optimization model (objectives, constraints, business rules). This work reframes pre-formulation clarification as a focused recovery problem: identify formulation-critical gaps that alter model structure, recover them with bounded interaction, and decide when the specification is model-ready. The core insight is to separate persistent gap diagnosis from question-selection so the agent asks only what matters and stops when enough structure is recovered.

Key Findings
  • OR-Clarify benchmark: provides paired public briefs and hidden, source-backed formulation-critical slots plus simulated-user responses, enabling controlled evaluation of slot recovery, stopping behavior, silent assumptions, and interaction cost. So what: gives a reproducible way to measure whether agents know when to ask and what to ask before modeling.
  • InterOPT framework: two-stage design (Dynamic Gap Search to track unresolved gaps; Gap-Guided Action Search to decide questions vs. stop). So what: separating gap tracking from action selection improves targeted recovery and reduces unnecessary queries.
  • Empirical results: InterOPT substantially improves exact slot recovery in choice-based (structured) interactions and is competitive in open/free-form settings; diagnostics highlight tradeoffs between question precision and interaction burden. So what: structured gap-guidance yields clear gains when options are available, while free-form gains are more mixed.
Who It's For & Tradeoffs

Great fit if you build LLM-driven tools that auto-formulate optimization models or assist OR practitioners: the benchmark and framework help evaluate and improve whether an assistant asks the right clarifying questions and avoids silent assumptions. Look elsewhere if your pipeline already enforces complete, machine-readable specifications (no need for clarification) or if realistic users cannot tolerate any interactive delay — this work assumes bounded interactive clarification and simulated-user responses rather than noisy human-in-the-loop variability.

Where It Fits

This paper sits at the intersection of operations research and LLM-based modeling: it targets pre-formulation reliability rather than solver performance or post-hoc debugging. It complements solver-in-the-loop modeling systems by front-loading requirement recovery and provides a standardized evaluation to compare clarification strategies.

Information

  • Websitearxiv.org
  • OrganizationsAffiliation: Cardinal Operations, Affiliation: Shanghai Jiao Tong University, Shanghai, China [email protected] {linyichen, chenyuzhou, linjianghao, taoyao, ddge}@sjtu.edu.cn * Equal contribution † Corresponding authors
  • AuthorsSihan Ge, Yichen Lin, Chenyu Zhou, Jianghao Lin, Tao Yao, Dongdong Ge
  • Published date2026/09/04

Categories

More Items

Presents two LLM-based search agents (Iris-mini and Iris-pro) trained by alternating supervised fine-tuning and reinforcement learning against live web search. Key features: web-graph-derived multi-hop tasks with entity abstraction, SFT–RL climbing, inference-time context management, and state-of-the-art open-source benchmark results.

Decides when past post-training updates should be reused for autonomous LLM adaptation by introducing Boundary-Calibrated Intervention Transfer (BCIT). BCIT binds effects to source context, checks applicability and hard conflicts, and runs bounded trials to obtain current-state evidence—reducing harmful updates and improving equal-budget final-model quality.

Reconstructs executable terminal workspaces from recorded agent trajectories and synthesizes verifiable single- and multi-round coding tasks for agent training; it replays file operations, uses an LLM completion agent to fill missing files/dependencies, and verifies tasks with autogenerated test suites.