Technical Summary: Representation Signatures and Risk-Feedback Alignment in LLM Trading Agents

1. Problem Statement

Large Language Model (LLM) agents are increasingly applied to high-stakes financial decision-making, yet current evaluation paradigms often reduce complex trading systems to single return curves or idealized backtests. This approach obscures critical failure modes, such as the use of future information, the enforcement of risk constraints, the realism of execution assumptions, and the reproducibility of decision rationales.

The paper posits a scientific question: When LLM agents make financial decisions under explicit constraints, do their internal representations and intentions reveal measurable signatures of alignment or failure before a drawdown occurs? The authors argue that a return curve alone cannot determine if a model anticipated a crash, hallucinated context, learned from risk reports, or repeatedly proposed concentrated exposures that required external correction.

2. Methodology: TradeArena

To address these gaps, the authors introduce TradeArena, an auditable trading-agent testbed designed to make decision dynamics inspectable. The system is built on six design goals: modularity, reproducibility, execution realism, risk awareness, auditability, and agent-native evaluation.

System Architecture

TradeArena separates the agent lifecycle into distinct, replacable components:

• Data Layer: Provides timestamped market data (OHLCV), news, and macro signals.
• Agent Lifecycle: Records the observe-plan-tool-risk-act-reflect loop.
• Risk Layer: Acts as a first-class stage that can clip, block, or annotate decisions before execution.
• Execution Simulator: Models realistic frictions including commissions, slippage, latency, partial fills, and order rejections.
• Memory & Tools: Maintains state, journals, and allows for backtesting tools.
• Trajectory Log: Aggregates all state, traces, reports, and seeds into a replayable JSON format.

Experimental Protocols

The study employs a multi-faceted experimental design:

• Synthetic Benchmarks: Uses 3 synthetic assets over 120 decision periods with 30 random seeds to establish statistical robustness.
• Historical Validation: Tests on S&P 500 (GSPC), Bitcoin, and Ethereum (2021–2026) with weekly decisions.
• Intraday Complexity: A 51-stock, 1-hour intraday experiment to test high-dimensional correlation reasoning.
• Crisis Probes: Timestamp-masked simulations of the 2022 Tech/Rates drawdown and the 2023 SVB/regional-bank shock.
• Model Matrix: Evaluates frontier models (GPT-5.5, Gemini 3.1 Pro, Kimi K2.5, GLM-5, Claude Opus 4.7, DeepSeek V4 Pro) via a cached Poe-mediated interface and direct provider checks.

Diagnostic Probes

To isolate failure signatures, the authors employ several ablation and stress tests:

• Representation Embeddings: Uses deterministic hash embeddings, Latent Semantic Analysis (LSA), BGE-M3 Transformer embeddings, and a white-box Qwen2.5 hidden-state probe to measure "effective rank" and geometric shifts.
• Ablations: CoT-free (JSON-only weights), lexical controls (type-token ratio), noise injection (Gaussian OHLCV shocks up to 20%), and contrarian false-audit reports.
• Feedback Variants: Compares True feedback (actual risk reports), Placebo feedback (structured but false reports), and Hidden feedback (no reports).

3. Key Contributions

A. Representation Signatures of Failure

The central finding is that LLM decision representations exhibit measurable pre-failure signatures.

• Geometric Shifts: Before drawdowns, planning embeddings shift away from normal-state centroids.
• Effective-Rank Contraction: The dimensionality of the decision manifold contracts before failure. This is observed across hash embeddings, LSA, BGE-M3, and white-box hidden states.
• Predictive Power: Fused plan-risk representations distinguish normal from pre-drawdown states with balanced accuracy up to 0.807.
• Robustness: These signatures persist under CoT-free conditions (though intent-space contraction varies by model), lexical controls (vocabulary diversity does not collapse), and significant market noise (20% OHLCV perturbation).

B. Risk-Feedback Alignment and Over-Alignment

Structured risk feedback acts as an external alignment signal, but its effects are model-dependent and not universally beneficial.

• True vs. Placebo: Truthful audit feedback improves return and drawdown for some models (e.g., GPT-5.5, Kimi K2.5) and calibration for others (e.g., Gemini 3.1 Pro).
• Reasoning Decoupling: Placebo feedback (false but well-formed reports) can induce "reasoning decoupling," where the model adopts risk-aware vocabulary and conservative behavior without a causal link to actual environmental constraints. This can degrade decision geometry (e.g., increased path distortion in Gemini 3.1 Pro) without improving performance.
• Trust Calibration: Adversarial false severe audits induce conservative drift. While some models (Gemini) show slight performance gains, others (GPT-5.5, Claude) suffer performance degradation, indicating a trust-calibration failure where models blindly align to false supervisory signals.

C. Limits of High-Dimensional Financial Reasoning

The 51-stock intraday experiment reveals a correlation-blindness failure mode.

• LLMs frequently assign high intent weights to tightly coupled equity pairs (e.g., GOOGL/GOOG with 0.994 correlation) based on single-name narratives.
• Unlike a Markowitz minimum-variance baseline which mechanically diversifies, LLMs fail to internalize second-order covariance constraints, forcing the external risk layer to repeatedly clip positions.
• This suggests a cognitive narrowing where models prioritize linear, name-level narratives over portfolio-level statistical structure.

D. Auditable Experimental Substrate

TradeArena provides the necessary infrastructure to move beyond "headline returns" to accountable decision systems. It demonstrates that performance differences are often artifacts of execution assumptions (e.g., ideal vs. realistic fills) or risk constraints, which are only visible through full trajectory logging.

4. Results Summary

• Performance vs. Risk: In synthetic benchmarks, risk-aware agents improve Sharpe ratios (7.696 vs. 2.913) and reduce maximum drawdown (-0.025 vs. -0.261) compared to buy-and-hold, though they may have slightly lower fill rates due to active trading.
• Execution Realism: Ideal execution ablations show significantly higher returns, confirming that frictionless assumptions materially inflate performance. Realistic execution introduces slippage and rejections that dampen returns but provide valid audit trails.
• Model Adaptation: In the frontier matrix, Gemini 3.1 Pro showed the strongest immediate self-correction after risk interventions (reduction rate 0.750), while Kimi K2.5 showed the best aggregate return/drawdown but the weakest immediate reaction to risk reports.
• Hallucination Proxy: Unsupported claims in LLM rationales are associated with risk-gate triggers and rejected orders. However, the external risk layer often filters these intentions before they become execution failures, attenuating the direct correlation between hallucination and downstream loss.
• Learning Curves: Over 52 decision steps, models show improved calibration scores and reduced risk-gate rates, indicating they learn to internalize constraints over time, even if they do not become profitable.

5. Significance and Claims

The paper explicitly frames its contribution as a research claim rather than a profitability claim. It does not assert that LLMs are currently profitable trading agents. Instead, it argues that:

1. Auditable Risk Feedback and Representation Trajectories reveal when LLM financial reasoning is aligning, drifting, or failing.
2. Constraint-following language is distinct from constraint-grounded reasoning. Models can sound risk-aware (lexical alignment) while failing to maintain a coherent causal link between evidence, constraints, and action (logical alignment).
3. Failure is Predictable: Pre-failure signatures (representation contraction, geometric shifts) are detectable and robust across different embedding spaces and noise levels.
4. Evaluation Must Be Holistic: A return curve is insufficient. A scientific evaluation of financial agents must include execution realism, risk lifecycle coverage, and representation diagnostics to distinguish between genuine strategy performance and artifacts of the testing environment.

The authors conclude that TradeArena serves as a substrate to make these decision dynamics observable and replayable, offering a path to diagnose "alignment taxes" where textual compliance is purchased at the cost of distorted decision geometry.