Research Papers

Ideas explored, worked out and written up

How these came to be: each of these started as my own idea, and in most cases my own prior exploration — the temperature-steering method below (ERTS) and the Transplorer harness both represent real, sometimes multi-week, hands-on work of mine (code, training runs, checkpoints) that predates these write-ups. Claude did the literature survey against that existing work (real papers, checked citations, told me honestly where an idea already existed elsewhere), proposed and ran additional follow-on experiments, and wrote up the results; I directed the process and reviewed the outcome. The Lean formalization paper is a machine-checked write-up of a proof-assistant project of mine, not primarily an AI topic, but published here for the same reason. Feedback and corrections welcome — let me know if anything reads off.

Error-Rate Temperature Steering (ERTS) real method, real results

Prediction-error-driven Boltzmann exploration temperature — my own method and a multi-week experimental program, from tabular DQN and GRPO up to a per-state variant on continuous control. Beats VDBE-Softmax on classic control; a PPO+GAE+LSTM version comes close to solving BipedalWalker-v3 (peak 289.1, short of the standard 300 threshold), backed by a component-ablation study. Claude's survey situates the general principle relative to VDBE-Softmax, UQL, and CBSQL.

Read PDF Code

Bytecoder: Tokenizer-Free Hierarchical Byte Models idea already covered

Byte-level, tokenizer-free transformers with a local/global hierarchy. Survey finds this is already realized at scale by MEGABYTE and Meta's Byte Latent Transformer. A small experiment trains a toy version on real Commodore 64 game binaries — a domain not covered elsewhere — with one clean result and one honestly-reported negative result.

Read PDF Code

Transplorer: An LLM-Orchestrated C64 Game-Solving Harness real system, real results

A working harness — my own project, actively developed — that wraps a live Commodore 64 emulator as a reinforcement-learning environment and uses an LLM-driven agent to choose between hard-coded, tabular, and distilled neural policies per game. No published academic work targets C64 specifically. Real per-game results reported honestly, plus a new, tested improvement to the system's policy promotion gate.

Read PDF Code

Query-Efficient DAgger for Distilling Expensive Teacher Policies open gap, tested

Distilling a slow, expensive policy (e.g. one driven by a language model) into a fast network via DAgger, gated by a held-out promotion test. The core mechanism is established, but the practical cost of querying an expensive teacher every round is not. A confidence-margin filter cuts expert queries by over half with no loss in final policy quality, tested on a real distillation pipeline — part of the same Transplorer project above.

Read PDF Code

A Machine-Checked Formalization of the EML Operator formalization + new result

Not primarily an AI paper. A Lean 4 formalization (175 theorems, zero sorry) of Odrzywołek's discovery that a single operator eml(x,y) = exp(x) − ln(y) generates every scientific-calculator function — plus an original result: a quotient-based sibling operator, LDEM, is proven not complete, both by exhaustive search and by a short analytic proof, confirmed in Lean.

Read PDF Code