TANREN: Evolving Code Instead of Training Weights
An evolutionary forge where a cheap frozen LLM writes candidate code and a deterministic verifier scores it — reaching results that rivaled large-scale RL, on one desktop for a few dollars. From Atari Pong's theoretical ceiling (21-0 in all 40 games) to a cache eviction policy that beat twenty years of classics on production traces.
8 articles
Part 1. TANREN — Atari Pong 21–0, and the Kill-Shot the Loop Found on Its Own
How an evolutionary loop — a cheap LLM writing code, scored by a deterministic verifier — reached the theoretical ceiling on Atari Pong (21–0 across 40 games) from nothing but the 128-byte RAM, and discovered a single winning shot without being taught it. Includes the honest protocol and the diagnostic loop that made it work for a few dollars.
Part 2. TANREN — 11 Wins to 1 Against Twenty Years of Cache Classics, on the Future of a Production Trace
Cache eviction — deciding what to drop when a cache is full — has been studied for more than twenty years. On that ground, a 53-line evolved Python function beats ARC, LIRS, and W-TinyLFU on future segments of Twitter's production traces. This post covers how the first attempt lost by overfitting to one time period, the two-window training that fixed it, and the public code anyone can rerun.
Part 3. TANREN — Patching the Evolved Cache Policy into a Real vLLM and Measuring It
The cache policy that won in simulation is patched into a live vLLM and measured on real hardware. The simulated win vanished at first; after fixing where and what to measure, tail latency (p99 TTFT) came out 13-16% lower under memory pressure. The prototype's costs and the conditions where it does not help are reported as measured.
Part 4. TANREN — an Evolved Scheduler Patched into vLLM Halves Tail Latency Under Heavy Load
The scheduler — the heart of LLM serving — evolved by TANREN and A/B tested inside a real vLLM. Across 16 real-trace configurations it beats the default FCFS 15 times (p=2.6e-4), cuts mean/p99 TTFT by 35-49% under heavy load, and does no harm when idle. Includes the honest record of going 0/6 on unseen data and six rounds of diagnosis before the win.
Part 5. TANREN — Discovering Breakout's Tunnel Strategy on Its Own, and Measuring Exactly Why the Rest Is Out of Reach
In Breakout, evolution discovered the classic tunnel strategy without being taught it and reached about 6.1x the human benchmark. Just as important, the reason it could not reach the strongest RL agents was pinned down by measurement, not speculation — drawing the boundary of what a few dozen readable lines of code can do.
Part 6. TANREN — On Atari Skiing, Deep RL's Hardest Slope, Readable Python Beats Every Published RL Agent
On Atari Skiing — where Agent57 needed roughly 78 billion training frames to surpass the human benchmark — about 130 lines of readable Python scored −3310.7 in a held-out evaluation, beating every published RL agent and the human benchmark. Includes the entry gate that estimated the odds before running, the techniques evolution added on its own, and the cost contrast: one day and a few dollars.
Part 7. TANREN — Atari Bowling, the Lost Campaign That Built the Entry Gate: Weighing a Fight Before Running It
On Atari Bowling, TANREN beat R2D2, the human benchmark, and DQN, then hit a ceiling at 74% of the strongest RL. This is the record of establishing — by measurement — that the ceiling was structural to the class of readable reactive code, and of how that loss produced the entry gate that now guards every experiment.
Part 8. TANREN — The Number I Almost Claimed and Then Withdrew: Atari Freeway and Measurement Honesty
A score of 31.9 looked like a DQN-beating result — until re-measuring under the published side's exact protocol dropped it to 28.6. This is the record of withdrawing the number, repairing the verifier, and re-running the evolution under the correct protocol to reach 31.2: past the human benchmark and DQN, at 96% of Agent57. An article about evaluation honesty.