PEARL: A 4B Model Beats DeepSeek-V3.2 685B at Optimization Modeling

Researchers have introduced PEARL, a system that turns an optimization problem described in plain language into a mathematical model and working solver code, checking and correcting the solution in a loop. Its compact version, PEARL-Qwen3-4B, with 4 billion parameters, outperformed the 685-billion-parameter DeepSeek-V3.2 on accuracy across optimization modeling tasks. The work was published on arXiv in July 2026.

What is PEARL and how does it work

PEARL is an interactive optimization modeling system that runs Python and mathematical solvers inside the solving loop itself, rather than producing a formulation in one shot. Most earlier approaches built on large language models worked in one-shot mode: the model built a formulation once, without running it, without looking at the solver's response, and without iteratively fixing errors.

PEARL mirrors how optimization problems are solved in practice, through repeated "solve, debug, rewrite" cycles. The system learns on its own when to test a partial model, how to rewrite it based on solver diagnostics, and when to stop.

• Works in multi-turn mode with tool integration: Python and solvers are called as it goes
• Uses intermediate execution results, feasibility signals, and correctness checks before finalizing
• Learns three things: when to test, how to revise based on diagnostics, when to stop
• PEARL-Qwen3-4B (4 billion parameters) outperformed DeepSeek-V3.2 (685 billion) on both macro- and micro-averaged accuracy
• Noticeably raises the share of verifiably solved problems over one-shot and tool-augmented baselines

Why a 4B model beat a 685B giant

PEARL-Qwen3-4B outperformed DeepSeek-V3.2-685B because it learns not just to formulate a problem but to debug it using solver signals, and that matters more than raw model size. The parameter gap between the two is roughly 171x, but on optimization modeling tasks, what decides the outcome is not the volume of weights but the ability to run a solution, see the error, and rewrite the code.

The key metric in the work is verified solve rate, the share of problems where the final solution actually passed the solver's check rather than merely looking plausible. On this measure, PEARL beats both ordinary one-shot models and tool-enabled approaches that lack a trained repair strategy.

"Our model PEARL-Qwen3-4B outperforms the significantly larger

DeepSeek-V3.2-685B on both macro- and micro-averaged accuracy on optimization modeling tasks," the paper's abstract on arXiv states.

What this means

PEARL's result is another argument that, for applied tasks with verifiable outputs, a trained "act, get feedback, revise" loop beats bare scaling. A small model with access to a solver and a self-checking strategy solves more optimization problems than a model 171 times larger working in a single pass.

Frequently asked questions

What is optimization modeling?

It is the translation of a real-world decision-making problem, described in plain language, into a formal mathematical formulation and executable solver code. Examples include resource allocation or logistics; the solver then outputs the optimal solution.

How does PEARL differ from ordinary LLM approaches?

Ordinary approaches work one-shot: the model builds a formulation once, with no execution and no revisions. PEARL runs Python and solvers in a loop, reads the diagnostics, fixes the model and code, and decides on its own when to stop.

How much smaller is PEARL-Qwen3-4B than DeepSeek-V3.2?

PEARL-Qwen3-4B has 4 billion parameters versus 685 billion for DeepSeek-V3.2, roughly 171 times fewer, yet it is more accurate on optimization modeling tasks.