Known risks and prior negative results

1. Compounding error and exposure bias. La Malfa et al. (Oxford / Turing, 2024) show that LLMs already lose accuracy on straight-line code beyond linear complexity and collapse on nested loops; later 2025/2026 work (“Demystifying Errors in LLM Reasoning Traces”, “Beyond Exponential Decay”) confirms error accumulation in long autoregressive traces. A terminal session is precisely a long autoregressive trace conditioned on the model’s own previous (possibly wrong) outputs. arxiv:2401.09074v2
2. Hallucination in computer-use world models. R-WoM explicitly diagnoses this for OSWorld/WebArena and needed external retrieval to fix it. Pure neural terminals will likely need analogous grounding (e.g. mixing real container rollouts in periodically, à la DynaWeb’s interleaving trick).
3. Code execution is known-hard. CWM, despite 32B parameters and massive trace pretraining, only “early-result” simulates Python step-by-step. The shell is broader (filesystem state, processes, network, randomness) and a faithful simulator may be much harder than a faithful Python tracer.
4. Token-faithfulness vs semantic-faithfulness trade-off. The proposal’s own open question. WebDreamer succeeds despite natural-language-only state predictions; that suggests semantic faithfulness may suffice for downstream policy gains and tokenwise matching of ls -la output is overkill.

Gap this work fills

To our knowledge, no published work has (a) trained a model end-to-end to act as a Linux terminal, nor (b) used such a learned terminal as the on-the-fly environment for SFT or RL of a coding agent. The closest analogues are ToolEmu (LM-as-environment, but prompted and used for safety red-teaming, not training), Kimi K2 (production-scale evidence that simulated-environment tool-use trajectories train good agents, but the simulator stack isn’t open and the shell isn’t a dedicated track), CWM (trains on execution traces but doesn’t deploy as a simulator), DynaWeb (does this for the browser), R-WoM (LLM-as-simulator for computer-use, but at the GUI level and retrieval-augmented), GenEnv (LLM-as-task-generator co-evolved with a solver), and ECHO (in-loop env-token CE — in RL it’s L_GRPO + λ · L_env, in SFT it’s just dropping the assistant-only mask; implicit world-modeling via the shared CE, but no separate simulator). TerminalWorld’s three-way comparison — assistant-masked SFT vs. full-transcript SFT (ECHO-in-SFT) vs. SFT-against-a-learned-terminal — is precisely the controlled ablation the field needs to know whether a fully neural terminal gym is a viable substitute for (or accelerant of) the expensive Docker-per-rollout regime that currently dominates terminal-agent training. The Nemotron-Terminal corpus, Endless Terminals’ procedural tasks, and Terminal-Bench 2.0 together make this the right moment to attempt it.