Improving Natural-Language Combinatorial-Optimization Accuracy in Resource-Constrained Language Models via Formal Abstractions
The paper introduces SDDL, a neuro-symbolic framework that translates natural-language scheduling problems into formal abstractions for external solvers, significantly improving solution feasibility and optimality for resource-constrained language models compared to direct generation or solver-code baselines.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a world where computers are asked to solve complex puzzles not by following rigid, pre-written instructions, but by listening to a human describe the problem in ordinary conversation. This is the frontier of artificial intelligence, where large language models are trained to understand and generate human text. These models are excellent at writing stories, answering questions, and even writing code. However, when asked to solve scheduling problems—like organizing a factory floor, arranging a film shoot, or coordinating a construction project—they often stumble. These tasks require finding a single workable plan among billions of possibilities while strictly obeying a web of rules: certain tasks must happen before others, specific machines can only handle one job at a time, and limited resources cannot be overused. For smaller, more efficient computer models that cannot run on massive supercomputers, the challenge is even steeper. They often produce answers that sound fluent and logical but are impossible to execute in the real world, violating the very rules they were meant to follow.
Researchers Shrenil Shaun Sharma and Avi Sharma set out to fix this gap between understanding a problem and solving it correctly. They focused on a specific type of artificial intelligence known as a "resource-constrained" model. These are powerful enough to be useful on standard hardware but lack the sheer size of the most advanced systems, making them prone to errors when asked to generate complex solutions directly. The team discovered that asking these models to write the final schedule or the computer code needed to solve the problem was a losing battle. The models would often miss a crucial constraint or invent a rule that didn't exist, leading to plans that looked good on paper but failed immediately when tested.
To overcome this, the researchers introduced a new approach called SDDL, or Scheduling Domain Definition Language. Instead of asking the computer to invent a solution from scratch or write a full computer program, they asked it to translate the human description into a very specific, simplified set of instructions. Think of this as asking a translator to convert a story into a structured outline rather than a full novel. The model's job is no longer to do the heavy lifting of finding the schedule; it is simply to identify the key pieces of the puzzle: the tasks, the resources, the rules, and the goals. It writes these down using a small, fixed set of building blocks that the researchers designed.
Once the model produces this structured outline, a separate, deterministic computer program takes over. This program acts like a strict compiler, translating the outline into a format that a specialized mathematical solver can understand. Because the outline uses a limited and well-defined vocabulary, the computer can check it for errors instantly. If the model made a mistake, the compiler catches it before the solver even begins. The solver then uses its powerful mathematical engines to find the actual schedule, guaranteeing that the result obeys every rule the model described. This method separates the task of understanding language from the task of doing the math, allowing the smaller models to focus on what they do best: recognizing patterns in text.
The team tested this method on three hundred different scheduling problems, ranging from job-shop scenarios where machines process parts in a specific order to project management tasks with limited budgets and timeframes. They compared their new method against two other approaches: one where the model tried to write the schedule directly, and another where it tried to write the full computer code to solve the problem. The results were striking. When using the new structured language, the smaller models became significantly more reliable. One model, which previously managed to produce a valid schedule only 1.3% of the time when writing code directly, succeeded 28.3% of the time with the new method. Another model improved from a 23.7% success rate to 55.3%.
Perhaps most importantly, when the models did produce a valid schedule, the quality of that schedule was excellent. The researchers found that among the successful schedules, the difference between the model's answer and the absolute best possible answer was effectively zero. This means the models were not just finding a solution, but a good one. The structured approach also drastically reduced the number of times the system failed to produce any answer at all, a common issue when models try to generate complex code. By forcing the model to express the problem's structure rather than the solution itself, the researchers enabled smaller, more efficient models to perform as well as much larger, more powerful systems that were previously considered necessary for these tasks.
This work suggests that the path to better artificial intelligence for complex tasks may not always require bigger brains. Instead, it may require better ways of talking to them. By providing a clear, constrained language that bridges human description and mathematical precision, the researchers showed that even modest computer models can solve difficult scheduling problems with high accuracy. The findings indicate that for many real-world applications, the key to success lies not in raw computational power, but in designing interfaces that guide the model to focus on the structure of the problem, leaving the heavy calculation to specialized tools. This approach offers a practical way to bring advanced scheduling capabilities to a wider range of devices and applications, making powerful optimization accessible without the need for massive, energy-hungry supercomputers.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.