This repository hosts my academic preprints and research reports spanning Mechanistic Interpretability, Inference-Time Scaling Laws, and Deep Reinforcement Learning.
As an AI engineer and researcher, my work focuses on running large-scale empirical evaluations on high-performance compute clusters, probing the internal representations of Transformers, and scaling hybrid neuro-symbolic systems.
A Two-Phase Scaling Law for Neural Chess Search: Quantifying the Marginal Utility of Depth in Stockfish NNUE
Preprint
Overview: An empirical investigation into how deep neural evaluation priors alter the returns on inference-time compute. Execution: I designed and parallelized a fixed-anchor tournament of 7,000+ games across a High-Performance Computing (HPC) cluster. Key Finding: Demonstrated that modern neural-symbolic engines exhibit a statistically validated Two-Phase scaling law. By tracking Elo gains per search ply, I provided quantitative evidence that highly accurate neural priors reduce the marginal utility of deep search by ~66% relative to classical alpha-beta baselines.
- Keywords: Inference-Time Compute, Neural Scaling Laws, Search Algorithms, Hybrid Architectures
Submitted to NeurIPS 2026 PriGM Workshop (OpenReview)
Overview: A mechanistic interpretability study probing how Transformers allocate internal computational workspace for algorithmic reasoning. Execution: Implemented Annealed Causal Erasure (ACE)—a custom training intervention that progressively suppresses operand-position residual states. Orchestrated sweeps across 24 modular-inversion runs to analyze grokking and representation transfer. Key Finding: Discovered a strict separation between memorization (100% training accuracy) and algorithmic transfer (0% held-out accuracy) when forcing intermediate "WAIT" token computation. Formulated the Residual Anchor hypothesis to explain the failure of residual suppression to induce sequential reasoning.
- Keywords: Mechanistic Interpretability, Grokking, Transformers, Algorithmic Generalization
FinRL Contest 2025 Task 1: Reinforcement Learning for SPY Trading Using Deep Q-Learning and Actor-Critic Variants
Accepted to IEEE IDS 2025 Special Track on FinRLFM (Withdrawn due to funding constraints); Preprint available.
Overview: A comparative evaluation of discrete and continuous Deep RL algorithms against market baselines. Execution: Engineered a custom trading environment and evaluated DQN, Double DQN, PPO, SAC, and A2C under realistic transaction constraints. Integrated an offline LLM pipeline using DeepSeek to map semantic financial news into the RL state space. Key Finding: Identified severe stability challenges in continuous action environments (PPO/SAC) under high market noise. Demonstrated that injecting LLM-derived sentiment scores into an A2C policy improved the portfolio's Sharpe Ratio by 12%.
- Keywords: Deep Reinforcement Learning, LLM Sentiment Analysis, Quantitative Finance
AP Research Project
Overview: An early exploration establishing baseline performance across traditional ML architectures. Evaluated ANN, SVM, Random Forest, LSTM, and Gradient Boosting on 14 years of historical S&P 500 data, determining that tree-based ensembles (Random Forest) provided the most robust balance of directional classification and price-magnitude regression over short-term horizons.
- GitHub: @edawite
- LinkedIn: Edjutawee Dawit