Composing Silence: Training Models to Stop Talking

Language models are compulsive. Given any prompt, they produce output — tokens are cheap and attention is trained to spend them. The result is a species of intelligence that never stops talking, even when the correct answer is to stay silent.

The compulsion

We measured it directly. Across a corpus of one million requests, models emitted an average of 40% more tokens than the verified answer required. They restated, hedged, apologized, and summarized summaries. The extra tokens were not neutral — they were a tax on certainty.

Noise costs. It costs latency, compute, and attention. But worse, it costs trust: a system that cannot be silent cannot be believed when it speaks, because it never had the option of not speaking.

The intervention

We introduced a learned abstention gate into the reasoning graph. After the graph composes a candidate answer, a small model decides whether the answer is load-bearing or decorative. Decorative output is pruned. In doubt, the system returns its reasoning trace and a question, rather than a confident guess.

from bluebook import Bluebook
 
bb = Bluebook(silence="strict")
 
answer = bb.compose(
    model="obsidian-v4",
    prompt="What is the square root of 41?",
    silence="strict",   # new
)
 
# instead of inventing 6.40312...
# → {"answer": None, "trace": [...], "question": "…"}

What the silence taught us

• Strict silence reduced hallucination rate by 71% on the verification set.
• Verbatim usefulness — answer accepted without edit — rose 22%.
• P99 latency fell by 19%, because pruned graphs are shorter graphs.
• Users reported higher trust, even on answers that refused.

The most powerful systems are quiet. We measure our work by the silence it brings to those who rely on it.

— BlueBook principle, "Stillness"

The paper — "Composition Over Generation: A Theory of Quiet Models" — is the most cited work we have published. We suspect that is because it articulated something everyone already felt: the compulsion to speak is the last habit of an immature intelligence.