[ Alignment ]

Models that remain helpful.

Instruction tuning, preference optimization, over-refusal control, and abstention.

Papers: 7
Latest: 2026-01-20
Area: Alignment

Focus areas

Preference optimization

KL-controlled updates with anti-sycophancy axes.

Honesty

Abstention and calibration probes beside helpfulness.

Publications

7 papers in this area

2026-01-20Alignment

Calibration and Abstention — Saying “I Don’t Know” Reliably

Probes and training signals for abstention when evidence is missing—progress and gaps on Origin and NextGen.

Read →

2025-10-22Safety

Safety Research — Red-Teaming, Alignment, and Deployment Guardrails

IdonAI’s safety framework for Origin, NextGen, and Quark — threat modeling, red-teaming, evaluations, product guardrails, and limitations.

Read →

2025-08-07Alignment

Over-Refusal Reduction without Safety Regressions

How we lowered benign over-refusal toward single digits while holding jailbreak metrics steady.

Read →

2025-04-28Safety

Prompt Injection Defenses in Retrieved and Pasted Content

Indirect prompt injection threat model, evaluations, and mitigations for Quark’s document-aware workflows.

Read →

2024-02-20Alignment

Preference Optimization v1 — Reward Models without Reward Hacking

First production preference-optimization stack: rater guidelines, KL control, and anti-sycophancy checks.

Read →

2023-07-09Alignment

Instruction Tuning Pilot — From Base LM to Helpful Assistant

First SFT mixtures and evaluation rubrics that turned early checkpoints into usable assistants—precursor to Quark behavior.

Read →

2022-04-12Publications

IdonAI Research Charter — Capability, Safety, and Longevity

Founding research principles governing what IdonAI optimizes for, what we refuse to optimize for, and how experiments compound into production systems that serve users without harming them.

Read →

[ Research ]

Explore the full archive.

Browse every area, or email the research team about collaborations, datasets, and eval access.

All research Email research