preprint · 2026
A controlled 2×2 factorial study crossing two model families (GPT-5.4-mini, Llama-3.1-8B) with two RAG pipeline architectures (naive vs. agentic self-correction) on HotpotQA. Finds a significant interaction effect: agentic self-correction improves the larger model (+6.0pp) but hurts the smaller one (−7.0pp), suggesting a capability threshold for productive self-correction.
preprint · 2026
A comparative study of Every-Visit Monte Carlo and SARSA(0) with five reward-shaping strategies on a stochastic FrozenLake grid-world. A novel safety-based potential function (distance to nearest hazard) achieves the best throughput for both algorithms, with SARSA + Safety-Based Shaping reaching 71.46% — a 42% improvement over the unshaped baseline.