PEARL: как модель на 4B обошла DeepSeek-V3.2 685B в моделировании оптимизации

PEARL: how a 4B model beat DeepSeek-V3.2's 685B in optimization modeling

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

What is PEARL and how does it work

PEARL is a system for interactive optimization modeling that runs Python and mathematical solvers directly inside the solving loop, rather than producing a formulation in one shot. Most previous approaches based on large language models operated one-shot: the model built a formulation once and did not verify it by running it, did not look at the solver's response, and did not iteratively fix errors.

PEARL replicates how optimization is solved in reality — through repeated "solve — debug — rewrite" cycles. The system learns on its own when to test a partial model, how to rewrite it based on the solver's diagnostics, and when to stop.

• Operates in multi-turn mode with tool integration — Python and solvers are called along the way
• Uses intermediate run 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 tasks 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 the problem, but to debug it using signals from the solver — and that turns out to be more valuable than raw model size. The gap in parameters between them is roughly 171-fold, but on optimization modeling tasks, what matters is not the volume of weights but the ability to run the solution, spot an error, and rewrite the code.

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

"Our PEARL-Qwen3-4B model 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 for the idea that, for applied tasks with a verifiable outcome, a trainable "act — get feedback — revise" loop beats sheer scale. A small model with access to a solver and a self-checking strategy solves more optimization tasks 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 code for a solver. For example, resource allocation or logistics — the solver produces the optimal solution as output.

How does PEARL differ from typical LLM approaches?

Typical approaches work one-shot: the model builds the formulation once, without running it and without 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 smaller, yet more accurate on optimization modeling tasks.