Research
Curos is a research organization first. This is where we publish what we learn about how models work, how to align them, how to test them, and how to run them affordably — in the open, for anyone to read.
Every paper here went through internal review and is published openly. Papers are grouped below by research area. Each paper lists its authors, its abstract, and its references — and each one is fictional, like everything else on this site.
Six areas make up the agenda. Interpretability asks what a model is actually doing internally, not just what it outputs. Alignment and evaluations ask whether it does what we intend and how we'd know if it didn't. Reasoning covers how models plan and solve multi-step problems, multi-agent work covers what happens when several models coordinate, and efficiency is the unglamorous work of making all of it cheap enough to give away for free.
Interpretability
Sparse Feature Geodesics: A Topology of Attention
We study the geometric structure of sparse autoencoder features during attention, and find that the path a feature takes — not just its endpoint — carries meaningful information.
Locating Refusal Circuits in Sparse Autoencoder Latents
We identify a sparse set of features that appear to implement refusal behavior in our models, and show that intervention on these features can be measured precisely.
Alignment
Scaling Test-Time Monitoring: Measuring Oversight Density
We introduce oversight density — the number of independent checks applied to a model's reasoning — and show it predicts how well post-hoc review catches errors.
Deference Under Uncertainty: When Models Should Say I Don't Know
We study when models defer to uncertainty instead of bluffing, and find that training explicitly for appropriate deference reduces confident-but-wrong answers without sacrificing usefulness.
Evaluations
Red-Teaming as a Civic Exercise
We ran red-teaming rounds open to the public and found that diverse non-experts found classes of failures our internal teams missed, changing how we think about who audits AI.
A Taxonomy of Hidden Capability in Frontier Models
We propose a working taxonomy for the capabilities a model has but does not obviously display — and a testing protocol for finding them before they surface in deployment.
Reasoning
Compositional Reasoning under Distraction
We show that multi-step reasoning degrades sharply when irrelevant context is inserted between the steps, and that a simple recency-based fix recovers most of the lost accuracy.
Chain-of-Confidence: Probabilistic Reasoning Traces
We propose that reasoning traces carry their own confidence, and show that having a model annotate each step with a lightweight confidence signal improves final-answer calibration without hurting accuracy.
Multi-agent
Emergent Coordination without Shared Reward
In cooperative settings, independent language models develop coordination conventions without any shared reward — and the conventions are stable, legible, and transferable to new partners.
Structured Debate Improves Multi-Agent Truthfulness
Structured, adversarial debate between model agents outperforms consensus-based aggregation on truthfulness — but only when the debate has rules and a verifiable record.
Efficiency
Tiered Dense–Sparse Routing for Efficient Serving
We describe a serving architecture that routes easy requests to a small model and hard ones to a large model, and show it cuts serving cost by half while preserving user-facing quality.
Quantized Attention at Frontier Scale
We show that attention weights can be aggressively quantized without degrading long-context reasoning — and that the savings scale with context length, making frontier-class context windows affordable.
How our research works
Research at Curos is driven by questions that come from our product and our commitments: how do we test models honestly, how do we explain their behavior, how do we keep them affordable? Papers are reviewed internally, and methods are shared openly so other organizations can verify or build on them.
Most papers start as an internal question raised by the evaluations or interpretability teams during a model's release cycle — something a benchmark couldn't explain, or a behavior nobody had a good account of. A small group works the question, writes it up whether the result is positive or not, and it goes through internal review before publication. We do not hold back negative results; a method that didn't work is often as useful to publish as one that did.
If you are a researcher, see the researchers page for how to work with our tooling and evaluation suite.