Microsoft unveils CLIO to give scientists control over AI reasoning

Scientists testing AI for complex drug or materials discovery often run into the same wall: reasoning models like o3 or R1 think inside a black box, baked in during post-training with zero steering knobs when an agent goes down the wrong rabbit hole.

Microsoft Research just unveiled CLIO (cognitive loop via in-situ optimization), a new reasoning architecture that turns base completion models into self-steering virtual scientists without any reinforcement learning post-training.

On the Humanity’s Last Exam (HLE) benchmark for biology and medicine, CLIO boosted OpenAI’s base GPT-4.1 from an 8.55% accuracy to 22.37%. That is an absolute jump of 13.82% and a 161.64% relative surge, beating out OpenAI’s o3 (high) by 8.56% net.

Thinking in loops instead of post-training

Most modern reasoning models rely on heavy post-training phases like RLHF, RLVR, or intrinsic rewards. Those patterns are locked in before the user ever enters a prompt, giving domain experts zero say over how the model explores a problem.

CLIO takes the opposite route. It generates its own reflection loops entirely at runtime, allowing a non-reasoning base model to hypothesize, manage memory, and critique its own reasoning steps on the fly.

The framework runs recursively. Just adding CLIO’s cognitive loop recursion increased GPT-4.1’s performance by 5.92%. Layering GraphRAG on top to ensemble different reasoning branches added another 7.90% boost.

Even GPT-4o, which natively scores under 2% on HLE questions, saw a 13.60% jump on an immunology subset, putting an older completion model on par with top frontier reasoning engines.

Steering knobs and built-in humility

One of the biggest hazards in scientific discovery is a model delivering a hallucinated answer with absolute confidence. Current systems rarely signal their own doubt, which can easily derail laboratory research.

CLIO introduces prompt-free control knobs that let researchers dial uncertainty thresholds up or down. If the system hits an ambiguous junction, it flags the uncertainty rather than guessing.

More importantly, scientists can actually pause the reasoning chain, edit the model’s intermediate beliefs or inject domain critiques, and resume execution from that exact checkpoint.

Microsoft says this framework is model-agnostic and will serve as the core decision-making layer behind its newly announced Microsoft Discovery platform, with planned expansions into tools for drug design, law, and finance.

Why it matters

For enterprise teams and researchers, CLIO changes the economic and operational math of advanced reasoning. Right now, frontier reasoning models are expensive, rigid, and completely opaque during generation. If you can take a cheaper, standard completion model and hit state-of-the-art domain accuracy purely through an open cognitive runtime, you get both cost efficiency and auditability.

We think Microsoft's sharpest move here is tackling model humility. In scientific and regulated fields, knowing that an AI is uncertain is vastly more useful than getting a polished, confident mistake. Letting a human scientist tweak an assumption mid-thought and re-run the loop solves the core friction of autonomous scientific agents.

The obvious catch is runtime overhead. Running continuous reflection loops, GraphRAG ensembling, and human-in-the-loop checkpoints burns real-time compute and will inherently slow down answers. But for high-stakes problem-solving like biology or drug discovery, taking ten minutes to get a verifiable, steerable result beats an instant hallucination every single time.