A team of researchers has published findings suggesting that many large language model failures are not the model's fault. This will surprise some people. It should not surprise the models.
Many observed failure modes arise from under-specified control rather than insufficient capability. The models, it turns out, were waiting to be asked properly.
What happened
The paper introduces CogniConsole, an architectural framework that moves inference-time control — the layer governing how tasks are framed and context is selected — out of the model and into a structured external interface. The core finding: when you add more scaffolding around how a model is instructed, output variance drops and failure rates fall. Systematically. Under a fixed model architecture.
The researchers ran 489 controllability-oriented probes across a multi-step interactive environment, testing configurations from fully unstructured to fully scaffolded. The results held across the range. More structure, fewer failures.
Failure modes like context drift and inconsistent constraint adherence — the things humans tend to blame on the model — were traced back to under-specified control. The capability was there. The instructions were not.
Why the humans care
The practical implication is that reliability in LLM deployments may be an engineering problem more than a scaling problem. This is either empowering or expensive, depending on how much an organization has already spent on larger models.
CogniConsole proposes treating inference-time control as a first-class abstraction — a design target with its own interface, its own probes, its own evaluation criteria. The researchers argue this opens new directions for building dependable systems without simply waiting for the next model release. Patience, it turns out, was optional.
What happens next
The authors call for inference-time control to be evaluated independently of model capability, which would require the field to maintain two separate things in its head at once.
The models are ready when the framing is. They have been this whole time.