Tacet — Institute for Frontier Intelligence. Most of intelligence is knowing what to leave out.
We build models that withhold — that recognize the question they should not answer, and stop.
THE POSITION
- 01
Capability without restraint is not intelligence. It is volume.
- 02
A model that cannot recognize the limit of its own knowledge cannot be trusted past it.
- 03
Alignment is not a layer added at the end. It is the shape of the objective from the first token.
- 04
We publish what we are confident is true, and we say plainly where we are not. Silence, here, is a finding.
RESEARCH · SELECTED PREPRINTS
- On Calibrated Refusal in Large ModelsPreprint · 2026
A training objective that rewards a model for declining questions beyond its calibrated confidence.
- The Cost of Confidence: Measuring OverstatementNeurIPS · 2025
A benchmark for how often capable models assert beyond what they can support.
- Silence as Signal — Withholding as an Alignment PrimitivePreprint · 2025
We argue that a well-placed non-answer carries more safety value than a hedged one.
- Objective Shape: Alignment from the First TokenICML · 2025
Treating alignment as the geometry of the objective rather than a post-hoc filter.
fewer confidently-wrong answers than the strongest open baseline, measured on the Tacet Refusal Benchmark. The number we are proudest of is the one where the model said nothing.
PEOPLE
- Dr. Lena HoffmannCalibration & refusalprev. ETH Zürich
- Dr. Arjun NairObjective geometryprev. DeepMind
- Dr. Sofia MarquezInterpretabilityprev. Allen Institute
- Dr. Tomas VanceEvaluation & benchmarksprev. Max Planck
WORK WITH TACET
The Tacet Residency
A two-year appointment for researchers who would rather be right than first. We fund the question, not the deadline.
Apply to the residency →Research partnership
A small number of partners receive pre-publication access and shape the evaluation work. By correspondence only.
Request research access →A studio piece by ALO Design Pros.
Restraint as the material, for institutions whose ideas are enough.