Bridgewater AIA Labs teamed with Thinking Machines Lab to fine-tune the open-weight Qwen3-235B base on expert-labeled financial documents using Tinker, on-policy distillation, and several training tweaks. The resulting model hits roughly 84.7 percent average accuracy across six investor triage tasks, beating the best frontier LLMs by about six points while cutting inference cost by a factor of 13.8.
So, we are back to 2016 after all
Bridgewater, one of the worlds largest hedge funds, a Tinker customer talks through how they've carefully fine-tuned a model focused on what makes interesting financial news. Their fine-tuned model is more effective and cheaper than any frontier model.
At daily workflow volumes even modest accuracy lifts matter less than the per-task price drop, which the team says crosses the threshold for routine analyst use.
The platform let researchers iterate on loss functions and teacher promotion schedules without spinning up their own clusters, though exact dataset sizes and total compute remain undisclosed.
Many users called Bridgewater's fine-tuned models for financial news and documents brilliant and high-ROI because narrow specialization beats general frontier models on enterprise tasks.
No Digg Deeper questions have been answered for this story yet.
@tinkerapi that is a clever way to handle data noise
People sometimes ask why fine-tune when general-purpose models keep getting better. Bridgewater's work is a good reminder that with the right data -- here, expert judgements -- you can beat prompting-only approaches by a lot. @ddkang and the Bridgewater AIA Labs team are great -- glad to see them sharing this.
Sorting which financial docs are worth an analyst's time is surprisingly hard for frontier LLMs. With an expert-labeled dataset and on-policy distillation, Bridgewater fine-tuned a model to do it reliably and cheaply. https://thinkingmachines.ai/news/learning-to-replicate-expert-judgment-in-financial-tasks/
Bridgewater, one of the worlds largest hedge funds, a Tinker customer talks through how they've carefully fine-tuned a model focused on what makes interesting financial news. Their fine-tuned model is more effective and cheaper than any frontier model.
Sorting which financial docs are worth an analyst's time is surprisingly hard for frontier LLMs. With an expert-labeled dataset and on-policy distillation, Bridgewater fine-tuned a model to do it reliably and cheaply. https://thinkingmachines.ai/news/learning-to-replicate-expert-judgment-in-financial-tasks/
Bridgewater and Thinking Machines just published a blog on training a custom model to replicate expert investor judgment. The task is filtering financial documents and news for relevance. Sounds trivial. Turns out it's not.
Sorting which financial docs are worth an analyst's time is surprisingly hard for frontier LLMs. With an expert-labeled dataset and on-policy distillation, Bridgewater fine-tuned a model to do it reliably and cheaply. https://thinkingmachines.ai/news/learning-to-replicate-expert-judgment-in-financial-tasks/
@soumithchintala You basically created an overfitted filter ?
@soumithchintala @MartinShkreli Did not expect a mega HF to be leading the charge on token cost optimization.
@soumithchintala How does cost per task account for the added upfront cost to fine tune? At what task volume does training actually break even?
@tinkerapi Is it open source?
@johnschulman2 @ddkang very cool, a lot of potential for finetuning!
@iAmHenryMascot @tinkerapi No strong reason. This claim that frontier models fail this task 50% is not something I believe.
Three changes closed the gap to 84.7%. Interleaved batching. Instead of mixing all tasks into one batch, each batch came from a single task, rotating round robin. A fully mixed batch means gradients from unrelated tasks fighting each other inside the same update.
@tinkerapi very tuff @SarahhhSuuu
@soumithchintala cheaper because they don't have to pay 95% GM?
End result: 84.7% accuracy, beating every frontier model tested, at roughly 1/14th the inference cost.
The trained model made 29.8% fewer mistakes than the best frontier model.
@tinkerapi 🧠🧠 @SarahhhSuuu
Data quality was the real bottleneck. They fixed it by training on the cheap labels first, then checking where the model disagreed with the labelers. Only those disputed cases got sent to real experts for review. They used Qwen3-235B. Standard GRPO training got them to 73.5% accuracy.
Interleaving keeps each batch's gradient clean while still cycling through every task often enough to avoid the model overfitting to one before seeing the rest. +12.1% accuracy.
@soumithchintala Feels like a VP trying too hard for scopemaxxing!
CISPO loss with asymmetric clipping, replacing standard importance sampling. In policy gradient RL, you reweight updates by the ratio of the new and old policy's probabilities for an action, then clip that ratio so a single update can't swing too far. Standard clipping is symmetric, with the same bound in both directions.
