Research at Hricor
Better intelligence.
By asking better questions.
Our research agenda connects model behavior, real-world evaluation, and practical deployment. These are the questions we’re building around.
Post-training
Our focus is the work that follows pre-training: shaping model behavior with instruction tuning, carefully designed data, and feedback from real tasks.
- How can training data better reflect the work people actually do?
- Which behaviors improve with supervised fine-tuning, and which need a different approach?
- How do we improve instruction following without losing general capability?
Reasoning & reinforcement
We investigate reasoning-oriented fine-tuning and reinforcement learning, with an emphasis on planning, checking work, and learning from outcomes.
- How can models learn to verify intermediate steps and recover from mistakes?
- Which feedback signals improve task completion rather than just answer style?
- How do we evaluate reasoning through observable, reproducible outcomes?
Evaluation & reliability
We are developing an evaluation approach around complete workflows, repeatable tasks, and explicit success criteria, so model improvements can be judged in context.
- Does an improvement transfer from a benchmark to everyday work?
- What fails when a task spans tools, documents, and multiple steps?
- Can the same outcome be reproduced across runs and environments?
Low-cost inference
Our future agenda includes quantizing and post-training open-weight language models and serving them ourselves on rented GPUs from partner clouds, while measuring the effects on quality, latency, and resource use.
- How much capability survives lower-precision inference?
- What is the right quality, cost, and latency tradeoff for a real workload?
- How can deployment evaluations make the models we serve easier to trust?
This page describes our research priorities and future directions. It is not a catalog of released models, benchmark results, or published papers. We will share results with the methods and evidence needed to assess them.
