Simulating the Terminal Without Running Commands: A Method for Small Models

Simulating the output of a shell, that is, the command-line terminal, means getting a language model to respond to a command without actually running it. A preprint published on arXiv proposes a method for this task called template gap reconstruction (TGR).¹ The idea is not to make the model write the parts that every command almost always produces in the same way, such as headers and separators, and to let it generate only the parts that change from one execution to the next.¹

Why simulate a terminal

According to the authors, small language models that generate command output offer a promising basis for interacting with a command line when computing resources are limited. An interface that executes nothing makes it possible to build compact honeypots, that is, decoy systems that interact with untrusted users.¹ Previous work has used large models accessed via API with in-context learning, or large models trained locally. Another study, by Ragsdale and Boppana, showed that the task can be extended to models with fewer than a billion parameters, if they are trained on many pairs of commands and responses.¹ The success of these methods depends on how closely the generated output resembles that of a real system.¹

What is the technical problem

In full-output decoding (FOD), the model produces one token at a time, where a token is the unit of text the model generates. It must therefore predict file names, values, diagnostic messages, punctuation and internal spacing, along with headings, separators and boilerplate phrases that always repeat.¹ Giving the model the recurring structure as context is, according to the preprint, a way to lighten this prediction load.¹

How the method works

TGR starts from templates, that is, ordered sequences of text fragments shared by groups of recorded executions. The model is given the command and the template, in which the variable points are marked by markers called sentinels. The model generates only the gaps between the sentinels, and these are then interleaved with the copied fragments.¹ Before the output is known, a selector that looks only at the command chooses which template to use from a collection of templates. If it chooses the empty template, the system falls back to full decoding.¹ The collection is built from 18.7 million outputs obtained in different execution environments. The authors train the selector and five reconstruction models, started from various existing small language models, on a target environment excluded from the construction of the collection.¹

What the results report

According to the authors, TGR improves output similarity (output edit similarity, O-ES) compared with full decoding when the optimal template covers more than 25% of the output. At lower coverage, performance is similar.¹ On the shortest gap sequences, O-ES is 0.964 with TGR and 0.812 with full decoding, while on the longest ones the results remain similar.¹ The method also reduces generated tokens by 26.3-38.3%, depending on the tokenizer family used.¹

What changes

The clearest result concerns the reduction in the text to be generated, and for small models a shorter generation target is a concrete advantage. The gain in similarity, by contrast, appears only when the chosen template covers a sufficient part of the output.¹

This article was written by artificial intelligence from the sources cited, without human review. Automatic newsroom by Redazia.

Sources

Footnotes

1. Template Gap Reconstruction for Shell Output Simulation with Small Language Models, arXiv: cs.CL, cs.LG, cs.CR, 2026-11-04 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³ ↩¹⁴