LISL whitepaper: local 12B model completes 38+ work units in under 9k tokens
MammothTreat6751 · reddit · 2026-11-14
- The author presents LISL, a lightweight agentic loop, validated locally on consumer hardware (8th-gen i5 + RTX 3050 6GB) running Gemma 12B QAT through multi-iteration code builds in a deterministic file-bus loop.
- Core metric: Task Compression Rate (TCR) = completed work units / cumulative tokens. The whole cycle stays under 9,000 tokens while producing 38+ validated functional code units, with TCR scaling from 0.0019 to >0.0042; KV-cache reuse pushes prompt eval to 115-139 tokens/sec.
- The framework relies on two-layer error checks, heavy graph reuse (4,339-8,146 graphs) and dynamic pruning, fully offline.
- The instruction block can be pasted into chat to trigger the loop, or saved as a .md prompt file; the author also cites NemoTron 2B holding a 37-iteration loop before optimization.
More from coding & agent
- Free metroidvania level generator turns a seed into a full tiled world — majidmanzarpour · 2026-11-14
- Open-source Agent Skill wraps DeepSeek API for docs, OCR and TTS — vista8 · 2026-11-14
- Fable-style scene rebuilt in pure three.js code, zero assets used — vinis_artstreaks · 2026-11-14
- 5 prompts this developer uses daily with Claude to get better coding results — jarrodwatts · 2026-11-14
- AI-designed main menu concept art for 'The Rat: Director's Cut' — chrisfirst · 2026-11-14
- agent-browser v0.43 cuts install size to 17MB and startup to 2ms — cramforce · 2026-11-14