References & Further Reading

The research this course draws on. These are the works actually cited in the units — follow the sources; the course’s opinions should belong to the literature, not its author. arXiv links are given where the citing unit provides an ID; where it does not, the venue is listed as written.

Papers

• Lost in the Middle — Liu et al. (TACL 2024). arXiv:2307.03172 . Position matters: facts mid-context are recalled worst — Units 0 , 5 .
• RULER — Hsieh et al. (COLM 2024). arXiv:2404.06654 . Usable context is well below the advertised window — Unit 0 .
• Levy et al. (ACL 2024). arXiv:2402.14848 . Added length alone degrades reasoning, holding the task fixed — Unit 0 .
• ReadAgent — Lee et al. (ICML 2024). arXiv:2402.09727 . Gist memory: keep short summaries, page in detail on demand — Unit 8 .
• MemGPT — Packer et al. (preprint). arXiv:2310.08560 . Virtual context: page content between an in-window tier and external store — Unit 8 .
• LLMLingua — Jiang et al. (EMNLP 2023). arXiv:2310.05736 . Prompt-level (token) compression — Unit 10 .
• LLMLingua-2 — Jiang et al. (Findings of ACL 2024). arXiv:2403.12968 . More realistic 2–5× compression on general tasks — Unit 10 .
• Prompt-compression survey — Li et al. (NAACL 2025). arXiv:2410.12388 . Frames hard- vs soft-prompt compression — Unit 10 .
• 500xCompressor — Li et al. arXiv:2408.03094 . A cautionary extreme-compression example — Unit 10 .