How Can We Make LLMs Exceptional?

What makes a person exceptional? How can we create exceptional things? What separates exceptional work from mediocre work?

Author: Koutian Wu; GitHub: ktwu01

I have been searching for excellence from the beginning. I care about taste: the ability to distinguish exceptional work from mediocre work, in research and in other areas.

For AI, that leads to a question: how can we make LLMs exceptional? I want a model that can use limited prior knowledge to reason through things it has never been taught. Faced with an unfamiliar problem, it should be able to form a judgment, check an answer, and recognize its own mistakes.

I know this is difficult. I want to understand who is working on it. These projects each investigate part of that question.

• François Chollet and ARC Prize. Chollet’s On the Measure of Intelligence frames intelligence around the efficiency of acquiring new skills, accounting for prior knowledge and experience. ARC tests generalization from a few examples. The ARC Prize 2025 technical report discusses iterative refinement, program synthesis, and small networks trained without broad pretraining. ARC performance measures a particular set of capabilities; broader intelligence needs further evidence.

• Alexia Jolicoeur-Martineau: Tiny Recursive Model (TRM). Less is More: Recursive Reasoning with Tiny Networks studies a small network that repeatedly updates an internal state and a proposed answer. The paper reports puzzle-solving results with 7 million parameters. TRM is a specialized reasoning model. Its results leave open how well the approach transfers to general language tasks.

• DeepSeek-AI: R1-Zero. DeepSeek-R1 explores reinforcement learning for reasoning. R1-Zero skips a preliminary supervised fine-tuning stage and learns from rewards on tasks with checkable answers. It starts from a pretrained base model, so its results depend on substantial prior training.

• Google DeepMind: AlphaGeometry. AlphaGeometry combines a learned proposal system with symbolic deduction and uses synthetic geometry data for training. It demonstrates advanced reasoning within a structured mathematical domain. Transfer to other domains remains a separate research question.

The question I want to keep asking: how much can a model infer on an unfamiliar task, with how little prior knowledge and computation?