Learning to Replicate Expert Judgment in Financial Tasks (Bridgewater x Thinking Machines)

Authors: Sarah Su, Kevin Zhu, Emily Xiao, Rohan Alur, Daniel Kang (Bridgewater AIA Labs), in collaboration with Thinking Machines Lab URL: https://thinkingmachines.ai/news/learning-to-replicate-expert-judgment-in-financial-tasks/ Published: June 30, 2026, on the Thinking Machines Lab news blog Found via: Ravid Shwartz Ziv’s July 1, 2026 tweet, whose only image is a screenshot of this post’s title card. One-line: Bridgewater fine-tuned an open-weight model on expert-labeled data and beat every frontier model it tested at triaging financial documents, at about a fourteenth of the inference cost.

Summary

The problem. Alpha comes from judgment, and judgment is hard to articulate. The team picked a deliberately narrow slice of an investor’s day: triaging the flood of news, research, central-bank documents and emails for what is relevant and worth reading.

Six tasks. Financial-article relevancy, central-bank document relevancy (does it signal the direction of rates?), generic document relevancy against an investor’s question, ad hoc content labeling, document truncation and email truncation (finding where boilerplate begins). The worked example contrasts an FT piece on Greenland, judged not relevant, with one on new China tariffs, judged relevant. Both touch geopolitics and finance, but only one matters to markets.

Frontier models fall short. Gemini, Claude and GPT variants averaged roughly 50% accuracy with a naive prompt. Expert-written instructions, plus a three-way label (relevant and interesting / relevant but uninteresting / irrelevant), lifted them into the mid-70s. Automatic prompt optimization added nothing. The best frontier result was 78.2%, below the 80% bar investors said they would trust. Newer models were barely better per dollar. The authors’ explanation is that a prompt only carries what an expert can put into words.

Data came first. Vendor (non-expert) labels were often wrong. Their fix was a verification loop: train on the noisy set, re-score the same data, and send only the examples where the model disagreed with the label to Bridgewater’s experts.

Training recipe. Built on Tinker, Thinking Machines’ fine-tuning platform, with Qwen3-235B as the base model. Plain GRPO took accuracy from 44.8% to 73.5%. Three changes pushed it further: interleaved round-robin batching across tasks (+12.1% over fully mixed batches), CISPO loss with asymmetric clipping (+10.1%), and on-policy distillation where the teacher is promoted every 20 steps only on a new validation high (+3.1%).

Results. 84.7% average accuracy versus 78.2% for the best frontier model, which works out to 29.8% fewer mistakes, at a 13.8x lower inference cost per task. The authors say the pattern holds on many internal tasks beyond the six published. Their thesis is “differentiated intelligence”: custom models tuned to one organization’s taste outperform general frontier models.

Why it matters

This is the proprietary-data argument made concrete by a firm whose edge is judgment. The moat is the expert labels, not the base model, and a smaller specialist model beats the frontier on both accuracy and cost.

Full text

Archived privately against link rot, with the post’s two example images: ../attachments/learning-to-replicate-expert-judgment-in-financial-tasks/learning-to-replicate-expert-judgment-in-financial-tasks.md

Connections

• Bridgewater and Thinking Machines just… (tweet) — the saved tweet; its image is this post’s title card (verified 2026-09-15).
• Bridgewater Associates — AIA Labs is Bridgewater’s AI research group.
• Thinking Machines Lab — Mira Murati’s lab; co-publisher, and maker of the Tinker training platform used here.
• Ray Dalio — founder of Bridgewater, the firm whose investors’ judgment is being distilled.
• Artificial Intelligence and Open Source — an open-weight base model fine-tuned past closed frontier models.