Recursive Language Models for Arbitrarily Long Prompts

Visual companion for the podcast episode on the MIT CSAIL RLM paper: a systems view of long prompts where memory lives outside the active window and the model iteratively inspects, delegates, and updates state.
arXiv: 2512.24601 MIT CSAIL preprint Episode visualization link
8K → 1M
Length range visualized
272K
GPT-5 direct-fit cutoff in discussion
+28.3%
Reported avg gain for recursive 8B
4
Task families highlighted

What this page shows

Not a text summary. The visuals compare three memory strategies, show “context rot” vs hard overflow, and diagram how an RLM keeps the active context bounded while semantic work scales externally.

SVG flow diagrams heatmaps comparison charts interactive toggles step-by-step recursion
Long Prompt 1M tokens external RLM Loop generate → execute → inspect Answer from iterative access Slice / Compare prompt[400k:420k] Recursive Subcall solve(chunk_i)

References

1. Zhang, Kraska, Khattab (2025). Recursive Language Models.
2. Merrill & Sabharwal (2024). Theoretical expressive power of reasoning models.
3. Hong et al. (2025). Context Rot.
4. Khattab et al. (2021). Baleen: Robust Multi-Hop Reasoning at Scale via Condensed Retrieval.
5. OpenAI (2025). Context compaction work.
6. Smith (2025). Long-context compaction work.
7. Wu et al. (2021). Task-specific long-context methods.
8. Wu et al. (2025). Context compaction / long-context scaffolding.
9. Anthropic (2025). Self-delegation / sub-agent work.
10. Schroeder et al. (2025). Self-delegation work.
11. Sun et al. (2025). Self-delegation work.
12. Chen et al. (2025). Deep research benchmark/work.
13. Bertsch et al. (2025). Information aggregation benchmark/work.
14. Bai et al. (2025). Code repository understanding benchmark/work.
15. Yang et al. (2025). Qwen3 technical report.
16. Core Context Aware Transformers for Long Context Language Modeling (2024).
17. Challenges in Deploying Long-Context Transformers: A Theoretical Peak Performance Analysis (2024).
18. Recursively Summarizing Enables Long-Term Dialogue Memory in LLMs (2024).
19. Augmenting Language Models with Long-Term Memory (2024).
20. Retrieval Augmented Generation or Long-Context LLMs? A Comparative Study and Hybrid Approach (2024/25).
21. LongRAG: Enhancing Retrieval-Augmented Generation with Long-Context LLMs (2024/25).
22. Let’s (Not) Just Put Things in Context: Test-Time Training for Long-Context LLMs (2025).
23. Z1: Efficient Test-Time Scaling with Code (2025).
24. Survey on Test-Time Scaling in LLMs (2025).
Transcript arXiv IDs extracted: 2512.24601
Interactive podcast companion • inline SVG + vanilla JS • dark systems map of long-context alternatives
Listen to the Episode →