AI BriefingWorldTUI: 360M-Parameter Model Simulates Terminal Interface Without Shell Backend
·2026.10.21 09:00
Key point
Fine-tuned on 2B tokens of synthetic data, the model predicts screen deltas from keypresses, outperforming prompted baselines on shell and file readback tasks.
Details
- 360M-parameter model (derived from Qwen2.5-Coder-0.5B) predicts terminal screen changes from raw keypress bytes without a shell or filesystem backend.
- Achieves 80/106 on ShellBench and 9/13 on file readbacks, outperforming prompted Qwen2.5-Coder-0.5B (52/106) and Qwen3-0.6B (60/106).
- Trained on 336k synthetic episodes (~2B tokens) generated via Docker containers, using a 333-token vocabulary and grammar-constrained decoding.
- Parallel row decoding with speculative verification reduces large frame latency from 6.7s to 0.87s on RTX 4090.
- int4 quantization compresses the model to 208MB for browser-based inference via WebGPU.
- Self-correction training (DAgger-inspired) improved PresetBench scores from 24/47 to 38/47.
- Distillation experiments failed to preserve memory capabilities, leading to the abandonment of layer removal in favor of quantization.
- Model struggles with new command combinations and complex Python operations, often failing on nested directory consistency.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Perplexity Proposes JOLT: Joint On-Policy Learning and Teaching for Efficient RL Post-Training