Where SLMs beat GPT-5

We’ve been seeing a consistent pattern across agent systems:

GPT-5 works well as a judge on average cases—
but breaks down on edge cases and policy boundaries.

That’s exactly where reliability matters.

In our recent work, we took a different approach:

• Generate adversarial edge cases from the spec

• Resolve ambiguity via multi-agent debate

• Train a task-specific small model (SLM) on that data

Paper: https://huggingface.co/papers/2604.25203

What we’re seeing:

• SLMs outperform GPT-5 on boundary decisions

• More consistent (less flip-flopping on similar inputs)

• Fast enough for real-time, per-interaction evaluation

This leads to a different stack:

• GPT-5 → generation

• SLMs → evaluation + guardrails

Curious if others are seeing similar behavior in production.

(If relevant, we also turned this into a product: https://www.producthunt.com/products/plurai