Mira Murati’s Thinking Machines Turns Bridgewater’s Secret Expert Judgment into a Trainable Skill—and Beats Frontier Models by 29.8% Fewer Errors

In the high-stakes world of macro investing, the difference between alpha and mediocrity often lies not in access to information, but in the subtle, experience-honed judgment of what information actually matters. Every day, investors at firms like Bridgewater Associates sift through floods of financial articles, central bank reports, research documents, emails, and internal memos. The real skill isn’t reading — it’s triage: deciding what deserves attention first.

This seemingly simple task has proven remarkably difficult for large language models. Frontier models like GPT, Claude, and Gemini, when given straightforward prompts, hovered around 46–50% accuracy — essentially coin-flip performance.

Even sophisticated expert-crafted prompts could only push them into the mid-to-high 70s. But a collaboration between Bridgewater’s AIA Labs and Thinking Machines Lab (founded by former OpenAI CTO Mira Murati) has changed the game.

By transforming proprietary expert judgment into high-quality training data, they created a custom fine-tuned model that achieves 84.7% average accuracy — delivering 29.8% fewer errors than the best frontier models while slashing inference costs by 13.8x.

The Core Challenge: Teaching Intuition That’s Hard to Articulate

Bridgewater and Thinking Machines focused on six practical financial tasks drawn from investors’ daily workflows:

Classifying whether a financial article is relevant and interesting to a C-suite macro investor.
Determining if a central bank document signals future interest rate moves.
Assessing if a research document answers a specific investor question.
Labeling ad-hoc vs. boilerplate content and locating issue-specific sections.
Truncating documents and emails at the point where boilerplate begins.

Mira Murati’s Thinking Machines Turns Bridgewater’s Secret Expert Judgment into a Trainable Skill—and Beats Frontier Models by 29.8% Fewer Errors For human experts, these judgments feel instinctive. A headline about tariffs might scream “must-read” in one market context, while a flashy geopolitical story registers as noise in another. LLMs struggled because these decisions rely on deep market context, priorities, and professional “taste” that prompts can only partially capture.

Non-expert labeled datasets performed poorly. Vendor labels introduced too much noise, as the nuances of macro relevance require seasoned investor insight.

Building a Dataset Worthy of Expert Judgment

Mira Murati’s Thinking Machines Turns Bridgewater’s Secret Expert Judgment into a Trainable Skill—and Beats Frontier Models by 29.8% Fewer Errors Bridgewater’s solution was an iterative cleaning process. They started with vendor labels, trained an initial model, and then routed disputed examples—where the model disagreed with the labels — back to expert investors for re-evaluation.

This “contentious rounds” approach filtered out errors and injected genuine expert patterns into the data. The result: a dataset encoding subtle decision-making heuristics that experts intuitively apply but rarely spell out in full instructions.

Such proprietary data is a true moat. Competitors cannot simply download equivalent quality from Hugging Face.

A Smart Training Recipe for Generalizable Judgment

Using Thinking Machines’ Tinker platform for rapid iteration, the team fine-tuned Qwen3-235B with several key innovations:

Interleaved Batching: Instead of mixing tasks randomly or training sequentially, they interleaved batches in round-robin fashion. This improved accuracy by 12.1% over fully mixed batches by helping the model generalize expert judgment across similar scenarios rather than overfitting to narrow patterns.
CISPO Loss with Asymmetric Clipping: This refined loss function carefully limited parameter updates, reducing the risk of latching onto spurious correlations in the data.
On-Policy Distillation with Strong Teachers: The model was iteratively distilled from its own stronger checkpoints. Every 20 steps, the best-performing checkpoint became the teacher for the next iteration—regularizing the student while allowing progressive improvement.

These techniques, combined with the high-quality labels, propelled the model well beyond prompt-engineered frontiers.

Impressive Results with Real-World Impact

The custom model not only hit 84.7% accuracy (up from the frontier’s 78.2%) but proved far more efficient due to its smaller, specialized size. For an organization running thousands of such triage operations daily, the cost savings are transformative. Bridgewater reports this accuracy level is now sufficient for reliable integration into daily workflows, freeing analysts for higher-value synthesis and decision-making.

The Broader Vision: Differentiated Intelligence

This project illustrates a powerful path forward: differentiated intelligence. Rather than chasing ever-larger generalist frontier models, organizations can create custom models tuned to their unique expertise and needs. High-quality proprietary datasets — rooted in hard-to-articulate human judgment—become the new competitive edge.

As Mira Murati and the teams at Thinking Machines and Bridgewater have shown, the future of AI in specialized domains like finance won’t just come from scaling parameters. It will come from teaching machines to replicate the hard-won intuition of the best human experts.

The full technical post is available on the Thinking Machines site: Learning to Replicate Expert Judgment in Financial Tasks.