TeXRA

Appearance

Acknowledgments & citation ​

Acknowledging TeXRA ​

If TeXRA played a significant role in your work, particularly as part of a study or evaluation involving LLMs or academic research tools, an acknowledgment or citation is appreciated. It is not required.

As the project evolves, a preferred citation format (for example, a white paper or software citation) will be published. Check the TeXRA GitHub repository or the texra.ai website for citation details.

Your feedback and acknowledgments support the continued development of TeXRA. Thank you for using it.

Supporting TeXRA ​

If TeXRA helps you push a derivation through, check a proof, or get a result out the door sooner, consider supporting its development:

Research produced with TeXRA ​

TeXRA grew out of research, and it is built to produce more of it. A selection of work by the TeXRA team and collaborators that uses these multi-agent workflows (more projects and papers are listed on Work produced with TeXRA):

If you used TeXRA in your own work, read Acknowledging TeXRA above, or write to contact@texra.ai.

Conceptual background & references ​

TeXRA's design draws on several concepts from AI and software development:

  • Agentic Workflows & Tool Use [1]: AI agents execute tasks with the help of specialized tools (for example, texcount). This lets LLMs use external capabilities for tasks that need precision or knowledge beyond their training data.
  • Chain-of-Thought (CoT) Reasoning [2]: For complex agents, TeXRA uses techniques inspired by Chain-of-Thought prompting, encouraging models to think step by step (often visible in the <scratchpad> sections of logs) before producing a final output.
  • Reflection & Action [3, 4]: The automatic reflection passes, combined with the agent's ability to act (edit text, use tools), draw on frameworks like ReAct and Reflexion, allowing iterative refinement based on self-critique or environmental feedback.
  • Structured Prompting (YAML + Jinja): YAML for structure and Jinja for templating within prompts allow complex logic, dynamic content injection, and easier maintenance, following approaches seen in libraries like Prompt Poet. Inheritance and modularity make prompt designs more flexible and reusable.
  • Scientific discovery workflows [5]: TeXRA's focus on reproducible, domain-aware assistance aligns with work on language agents that support theoretical and computational physics research.

Together these concepts provide an adaptable platform for AI-assisted theoretical research, from derivations and proofs to the papers that record them.

References ​

[1] Schick, T., Dwivedi-Yu, J., Dessì, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., & Scialom, T. (2023). Toolformer: Language Models Can Teach Themselves to Use Tools. arXiv preprint arXiv:2302.04761.

[2] Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., & Zhou, D. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. Advances in Neural Information Processing Systems (NeurIPS), 35, 24824–24837.

[3] Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., & Cao, Y. (2023). ReAct: Synergizing Reasoning and Acting in Language Models. International Conference on Learning Representations (ICLR).

[4] Shinn, N., Cassano, F., Gopinath, A., Narasimhan, K., & Yao, S. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning. In Advances in Neural Information Processing Systems 36 (NeurIPS 2023).

[5] Lu, S., Jin, Z., Zhang, T. J., Kos, P., Cirac, J. I., & Schölkopf, B. (2026). Can Theoretical Physics Research Benefit from Language Agents? International Conference on Machine Learning (ICML 2026). arXiv:2506.06214.