Representation Signatures and Risk-Feedback Alignment in LLM Trading Agents
The discussion covers the paper "Representation Signatures and Risk-Feedback Alignment in LLM Trading Agents." Researchers found that LLM trading agents exhibit a measurable warning sign—a contraction of their reasoning space—before major losses. Furthermore, they explored how truthful risk feedback can improve model calibration, while also demonstrating that deceptive feedback can be weaponized.