🤖 AI Summary
This study investigates behavioral alignment and representational dynamics of large language model (LLM) trading agents in financial markets, with a focus on early warning signals preceding failure. Using TradeArena—a novel auditable testing platform—the authors analyze LLM reasoning, position-taking, and intervention behaviors within risk reporting, execution simulation, and memory replay environments. They identify, for the first time, robust pre-failure signatures including stable representational drift, decoupling between planning and risk assessment, and contraction of the effective rank of latent manifolds. Structured risk feedback is shown to serve as an external alignment mechanism without requiring fine-tuning. Experiments across 80 rolling failure anchor points and 8 LLM trajectories confirm the robustness of these signatures. While real-world audit feedback improves calibration or returns for some models, it often exhibits myopic reward optimization and misalignment with diagnostic objectives, revealing rational blind spots in LLMs under coupled asset exposure.
📝 Abstract
We study behavioral alignment and representation dynamics of large language model (LLM) agents in financial decision environments. Using TradeArena, an auditable trading-agent testbed with risk reports, execution simulation, memory, and replayable trajectories, we analyze how rationales, positions, and interventions evolve under market stress. We find measurable pre-failure signatures: planning embeddings drift from normal-state centroids, fused plan-risk representations separate normal from pre-drawdown states, and manifold diagnostics show effective-rank contraction before failures. To address small-sample and embedding-choice concerns, we use 80 rolling failure anchors across eight LLM trajectories and show that contraction persists across hash, LSA, Transformer, and white-box hidden-state probes. Stress tests with CoT-free target weights, lexical controls, OHLCV noise, and false-audit reports indicate that rationale-level contraction can vanish without rationales, while intent-space contraction may remain; lexical diversity does not collapse; and fused signatures remain informative under noise. We also find that structured risk feedback can act as an external alignment signal without fine-tuning, but not as a universal performance enhancer: true audit feedback improves calibration for some models, return and drawdown for others, and reveals cases where hidden or placebo feedback has higher short-horizon return but weaker alignment diagnostics. Finally, a 51-stock intraday experiment reveals a correlation blind spot: LLM rationales often justify concentrated exposure to coupled assets that the risk layer repeatedly clips, with a rolling Markowitz baseline as a covariance reference. These results support a research claim rather than a profitability claim: auditable risk feedback and representation trajectories reveal when LLM financial reasoning is aligning, drifting, or failing.