According to Dynametrics Beating monitoring, Bridgewater's Artificial Intelligence Applications Lab (Bridgewater AIA Labs) collaborated with Thinking Machines to successfully apply a fine-tuned model to financial information screening, achieving accuracy beyond state-of-the-art large models (GPT 5.5 and Claude Opus 4.8). In six daily investment tasks such as financial article relevance and central bank document analysis, this fine-tuned model reduced the error rate by nearly 30%, while the inference cost was only about one-fourteenth of that of large models (a 13.8x reduction). This study demonstrates the potential for enterprises to achieve "Differential Intelligence" through domain-specific fine-tuning.
The research team fine-tuned the Qwen3-235B base model on the Tinker platform. Due to the high error rate in labels directly obtained from non-professional annotators, the team designed a validation correction mechanism, where difficult samples with model predictions inconsistent with the labels were reassessed by investment experts. Ultimately, the trained model's accuracy increased from the baseline model's 44.8% to 84.66%, surpassing GPT 5.5 at 78.2% and Claude Opus 4.8 at 78.0%. While significantly improving accuracy, the inference cost also saw a dramatic 13.8x cliff-like drop.
The fine-tuning solution consists of three core improvements: first, interleaved batch processing, training different tasks sequentially to avoid interference from mixed training; second, the introduction of loss function optimization with asymmetric pruning sampling; third, the use of online policy distillation, penalizing the model when deviating from the teacher distribution and dynamically updating the teacher model every 20 steps when achieving new highs on the validation set.
