Research Explainer · XiaoHu Explains

Bridgewater trained a model just for triaging financial information — 84.7% accuracy, methodology public

Partnered with Thinking Machines, fine-tuned an open-source model on expert-labeled data: 29.8% lower error rate than the best frontier model, at 1/14th the inference cost
TL;DR
  • Thinking Machines Lab, together with AIA Labs at Bridgewater, used their in-house fine-tuning platform Tinker to train a custom model specialized in financial information triage.
  • Top frontier models (Gemini, Claude, GPT families) using simple prompts on 6 financial triage tasks averaged only about 50% accuracy. After heavy prompt optimization, the best reached 78.2% — still short of the 80% credibility threshold investors require.
  • The training data initially came from non-expert labelers and was riddled with errors. The team designed a scheme where only "disagreement" samples — where the model's judgment didn't match the label — went to expert review, while the rest were kept as-is, keeping labeling costs down.
  • Starting from the open-source Qwen3-235B base model, standard GRPO reinforcement learning fine-tuning pushed accuracy to 73.48%. Stacking on interleaved batch training, the CISPO asymmetric clipping loss, and online policy distillation with dynamic teacher promotion brought it to 84.66%.
  • The final model hit 84.7% accuracy — a 29.8% lower error rate than the best frontier model tested (78.2%), with per-task inference cost at just 1/13.8 that of the corresponding frontier model.
⚑ Editorial note: this piece is based on an official blog post jointly published by Thinking Machines Lab and Bridgewater's AIA Labs. All models, data, and comparison results are self-reported by the publishers, based on a public subset of their internal data. Figures below are relayed as stated in the original.
1What an investment manager judges in a glance, AI can only guess

News an investment manager judges in a second — AI can only guess

Thinking Machines Lab and Bridgewater's AIA Labs jointly published a post detailing how they used the in-house fine-tuning platform Tinker to train a financial information triage model, along with the results.

What they set out to automate isn't writing research reports — it's the "information triage" investment managers do over and over, every day: pulling out what's actually worth reading from a flood of news, research reports, company filings, and emails. The reading itself isn't hard; what's hard is that layer of up-close judgment, which eats up huge amounts of time. The team wanted to see: can this work be handed off to a model?

Here's the result upfront: top frontier models tackling it directly averaged about 50% accuracy — basically a coin flip — while this custom small model hit 84.7%, at just 1/13.8 the inference cost of a frontier model.
◆
Why it's worth reading: a small model fine-tuned from an open-source base cut the error rate of the strongest frontier model (78.2%) by 29.8% on a concrete judgment task, while slashing per-task inference cost to 1/13.8. This isn't "a smarter general model" winning — it's "a small model tuned for one specific job" winning.
2Six small things to automate

Six "small things" investment managers do every day

These six things are second nature to investment managers — judged in a second. But the moment you have to explain "how exactly did I decide that," people get stuck — which is precisely why they're hard to teach an AI. The team broke them out one by one into a benchmark.