Smaller LLMs Learn Complex Reasoning With Structured Knowledge Extraction Framework

A team has developed SAKE (Structured Agentic Knowledge Extrapolation), a framework that enables smaller language models to solve complex problems in specialized domains by learning to autonomously piece together and extend existing information. The approach outperforms larger commercial models on benchmarks while dramatically reducing computational overhead.

The Knowledge Extrapolation Challenge#

Tackling advanced questions in fields like medicine and biology often requires reasoning across fragmented information sources. Rather than relying on the ability to retrieve comprehensive external knowledge—which can be impractical for niche domains—SAKE teaches models to infer novel insights by combining what they already know. The system works by training language models to interleave two external tools over a structured three-step process: identifying key entities relevant to a question, filtering the most pertinent concept clusters, and then performing analogical reasoning to form new connections and answers.

How SAKE Trains Models#

The framework uses reinforcement learning, specifically GRPO, to optimize the entire pipeline end-to-end. The reward signal trains models on two fronts: whether their outputs follow the correct format and whether their final answers are accurate. Through this process, smaller open-source models learn not just what information to retrieve, but how to reason effectively over incomplete and structured knowledge.

Benchmark Results#

In testing, a fine-tuned version of Qwen2.5-7B—a relatively compact model—achieved 75.4% accuracy on biomedical reasoning tasks and 81.3% on commonsense reasoning, surpassing GPT-3.5-Turbo's performance (70.1% and 74.7% respectively). The gains came alongside a reduction in token usage exceeding 90%, meaning the smaller model required far less computational work to reach superior results.

Why This Matters#

The findings suggest that advanced reasoning over incomplete structured knowledge does not demand the largest available models or intricate multi-step prompting strategies. Instead, smaller, openly available models can learn these capabilities end-to-end through reinforcement learning when equipped with the right tools and training signals. This opens possibilities for deploying efficient, specialized reasoning systems across resource-constrained environments.