Abstract
Large language models are increasingly deployed as autonomous research agents in quantitative finance: they read briefs, write strategy code, evaluate simulator output, and propose refinements in a loop. This report argues that the dominant bottleneck in such pipelines is not model capability or simulation fidelity, but reward design, that is, the choice of what feedback signal an LLM consumes between iterations. PnL only rewards collapse rich behavioural information into a single noisy scalar, encouraging overfitting and obscuring whether a strategy succeeded for the right reasons. We outline a taxonomy of reward signals (PnL, risk adjusted, log derived, multi dimensional) and argue that the most informative signals, namely structured execution logs, are also the ones least compatible with the context windows of current LLMs. We then propose treating log aware reward computation as a long context problem and sketch how the Recursive Language Model (RLM) architecture of Zhang et al. (2025) can serve as a hierarchical evaluator that interrogates logs without flooding the strategy LLM’s context. We compare this approach against grep style tool use and retrieval augmented generation. The framework is intended as a general design pattern for LLM driven research loops in finance.
0 Comments