Two research papers published on arXiv on July 20, 2026, evaluate large language models across distinct specialized domains, revealing both capabilities and critical limitations.
AI Trading Performance
According to a Master’s research paper from Georgia Institute of Technology, researchers systematically evaluated five prominent LLMs—GPT-4 Turbo, Claude 3 Opus, Gemini 1.5 Pro, Llama 3 70B, and domain-specialized FinGPT—for technical market analysis. The evaluation spanned candlestick pattern recognition, directional signal generation, backtesting through simulated execution, and financial report comprehension.
The study found that “GPT-4 Turbo achieves the highest annualized return and Sharpe ratio among general-purpose models, while FinGPT demonstrates competitive risk-adjusted performance due to domain-specific fine-tuning,” according to the paper. Both models outperformed a passive S&P 500 benchmark under tested conditions. However, the research identified “persistent failure modes across all evaluated models, including numerical hallucination, context-window limitations, and inconsistent performance in sideways market regimes.”
Fingerprint Verification Benchmark
Separately, researchers introduced SLAPBench, “the first benchmark for MLLM-based four-finger SLAP fingerprint verification,” according to the second arXiv paper. The benchmark evaluated four open-source models and Claude Opus 4.8 using NIST SD302b data with 7,832 pairs.
Results showed that “prompting governs verification behavior,” with task-description prompting collapsing all four open-source models to near-100% False Accept Rate. Claude Opus 4.8 achieved the best binary result at 20.2% FAR, while Qwen3-VL-8B attained perfect separation (AUC = 1.000), though researchers treat this as potentially diagnostic rather than demonstrating true capability.