necrashter/symbol-invariant-transformer

Symbol-Invariant Transformer

This repository contains the models for the ICML 2026 paper, Names Don't Matter: Symbol-Invariant Transformer for Open-Vocabulary Learning.

Links:

1. arXiv
2. Project Page
3. Code

Models

In the root directory, you will find a directory for each task:

• ltl: Linear Temporal Logic task
• prop: Propositional Logic task

In each task directory, you will find:

1. ablation: All ablation models presented in the appendix
2. converted: Results of model conversion & fine-tuning experiments in Section 5.8
3. generalization: Main experimental results in Figure 2
4. table: Experiments with reduced & renamed datasets (Table 1)

Note that the baseline models in Table 1 can be found in the previous paper's repository.