[ Research ]

Open inquiry. Rigorous methods.

We publish our work, share benchmarks, and invest in the science required to build safe, capable systems.

Publications

Architecture, scaling, and efficiency — published as we learn.

Benchmarks

Transparent evals across reasoning, coding, and multimodal tasks.

Infrastructure

Training and inference systems built for frontier scale.

Safety

Red-teaming, adversarial testing, and deployment guardrails.

Alignment

Methods for models that stay helpful, honest, and controllable.

Experiments

Live probes and ablations that feed the next model cycle.

Datasets

Curated corpora and eval sets released for the community.

Publications

28 papers

2026-03-12Publications

Quark — A Conversational Assistant Built on Origin and NextGen

Technical report on Quark — system design, training, routing, evaluation, and responsible deployment of IdonAI’s primary assistant product.

Read →

2026-02-18Publications

NextGen — Technical Report on Architecture, Training, and Evaluation

Full technical report on NextGen and Origin — architecture, training stack, Gemini-competitive benchmarks, ablations, scaling, and deployment considerations.

Read →

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 →

2025-12-03Safety

Safety Evaluation Card — 2025 H2 Snapshot

Compact safety scorecard for Origin, NextGen, and Quark—companion to the full Safety Research report.

Read →

2025-11-14Experiments

Agentic Tool Loops — Planning, Verification, and Stop Conditions

Experimental agent patterns on NextGen: plan–act–verify loops, and why Quark keeps tools conservative in production.

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 →

2025-09-25Infrastructure

Enterprise Isolation Patterns for Model Serving

Tenant isolation, VPC patterns, and audit logging considerations for IdonAI enterprise deployments.

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 →

2025-06-12Publications

Distilling NextGen into Origin — Quality without Latency Tax

Task-selective distillation that lifts Origin’s coding and MMLU scores while preserving Flash-tier serving.

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 →

2025-03-19Experiments

Structured Output Constraints — JSON Reliability for Tools

Constrained decoding, schema adherence metrics, and failure taxonomies for Origin and NextGen tool use.

Read →

2025-01-08Benchmarks

Eval Harness Standardization — Templates, Seeds, and Leaderboard Hygiene

How IdonAI locks prompt templates and harness versions so Origin/NextGen numbers stay comparable across cuts.

Read →

2024-11-21Publications

Quark Routing Prototype — Dual-Model Assistant Control Plane

First Quark router design: features, policies, cascade mode, and cost–quality tradeoffs before public launch.

Read →

2024-09-09Experiments

Multimodal Alignment Pilot — Vision and Audio into the LM Space

Early encoder–LM alignment experiments that unlocked NextGen multimodal features.

Read →

2024-08-29Infrastructure

Continuous Batching Notes — Fairness, Prefill, and Cache Hits

Practical serving research: scheduling fairness between short chat and long prefill jobs on shared Origin pools.

Read →

2024-07-18Experiments

Mixture-of-Experts Routing Study for NextGen-Class Models

Expert count, top-k routing, and load-balancing ablations that shaped NextGen’s sparse FFN design.

Read →

2024-05-02Experiments

Long-Context Mid-Training — From 4K to 128K and Toward 1M

Curriculum and position-extension methods that enabled Origin 128K and paved the way for NextGen 1M.

Read →

2024-04-16Datasets

Dataset Mixture Ablations — Code, Math, and Prose Tradeoffs

Systematic mixture ablations showing how code and math density reshape Origin-class downstream profiles.

Read →

2024-03-11Safety

Red-Team Pack v1 — A Minimal Adversarial Suite for Release Gates

First standardized jailbreak and misuse suite used as a hard gate before Origin promotions.

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 →

2023-12-05Infrastructure

Origin Serving Canaries — Latency, Batching, and Failure Budgets

Operational research on validating Origin under load before Quark promotion: SLOs, canaries, and rollback.

Read →

2023-11-16Publications

Origin — Technical Report on IdonAI's Fast Production Model

Full technical report on Origin: architecture, training pipeline, Flash-class benchmark suite, serving infrastructure, distillation methodology, safety interfaces, ablation study, and known limitations.

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 →

2023-05-18Infrastructure

Grouped-Query Attention for Interactive Serving

Why Origin adopted GQA: KV-cache memory, batchability, and quality tradeoffs measured on internal canaries.

Read →

2023-03-22Experiments

Tokenizer Design Notes for a Shared Origin–NextGen Vocabulary

How we chose a byte-level BPE vocabulary for code-heavy workloads and future multimodal text interfaces—fertility measurements, code round-trip tests, and the case for boring infrastructure done right.

Read →

2023-02-14Datasets

Data Quality Filtering v1 — Provenance, Dedup, and Toxicity Screens

First production-oriented data pipeline for IdonAI pretraining: filter architecture, near-dedup strategy, toxicity screens, PII heuristics, and measured impact on training quality versus unfiltered baselines.

Read →

2022-09-08Experiments

Early Scaling Observations on Small Dense Transformers

2022 internal study of loss-vs-compute on dense decoder-only models—what scaled predictably, what failed to transfer to Origin, and what design decisions these experiments permanently shaped.

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 →