Recursive Language Models (RLM) - Research

Focus: Original MIT paper research and algorithm Tachikoma Implementation: See docs/capabilities/rlm.mdFull Documentation: See docs/rlm.md

---

The Problem

Standard LLMs have fixed context windows (128K-200K tokens). How to handle:

• Entire codebases
• Large documentation sets
• Multi-file refactoring
• 10M+ token contexts

---

MIT Paper: "Recursive Language Models"

Paper: "Recursive Language Models: Efficient Long-Context Processing with Limited Computation" arXiv: https://arxiv.org/abs/2512.24601

Key Findings

• Process inputs up to two orders of magnitude beyond context windows
• RLM-Qwen3-8B outperforms base Qwen3-8B by 28.3% on average
• Approaches GPT-5 quality on long-context tasks

---

Key Innovations

1. Symbolic Handle to Prompt

Concept: Prompt lives in REPL (external to LLM)

How it works:

• Large context stored externally (not in LLM context)
• LLM maintains symbolic reference to context
• Only metadata in LLM context (constant size)

2. Symbolic Recursion

Concept: LLM writes code that calls sub_LLM() in loops

How it works:

# LLM generates this Python code
chunks = chunk_indices(size=50000)
results = []
for start, end in chunks:
    chunk = peek(start, end)
    result = sub_LLM("Analyze", chunk=chunk)  # RECURSION!
    if result["success"]:
        results.append(result["result"])

Why this is recursion: LLM is calling itself through sub_LLM() function.

3. Output via Variables

Concept: Results stored in REPL variables (Final)

How it works:

• Intermediate results stored in REPL
• Final synthesis in variable Final
• LLM only sees variable names, not actual results

4. Metadata-Only History

Concept: Only constant-size metadata in LLM context

How it works:

• No actual context chunks in LLM history
• Only metadata: chunk IDs, processing status
• Context window stays fixed regardless of input size

5. Sub-LLM Calls

Concept: LLM calls itself via subagent

How it works:

• Subagent acts as "sub-LLM"
• Processes individual chunks
• Returns structured results
• Main LLM synthesizes from results

---

Performance Results

Metric	Result
Context scaling	100x beyond context windows
Accuracy improvement	28.3% over base model
Quality	Approaches GPT-5 on long-context tasks
Computation	Limited - scales efficiently

---

Limitations from Paper

1. Sequential processing - Chunks processed one at a time
2. Manual chunking - No semantic boundary detection
3. Fixed chunk size - No adaptive sizing based on content

---

Related Research

• Position Bias: LLMs pay more attention to tokens at start and end of context
• Tool-Augmented LLMs: Tools add latency but improve accuracy
• Modularity: Smaller, focused components work better than large monolithic prompts
• Verification Loops: Reflection after execution improves quality

---

Implementation Notes

Tachikoma's RLM implementation extends the MIT paper with:

1. Adaptive chunking - Semantic boundary detection (JSON objects, Markdown headings, code functions)
2. Parallel processing - Process 5 chunks concurrently in waves
3. Plugin system - Native opencode integration for tool discovery
4. Environment variables - Testing and control

See: docs/capabilities/rlm.md for Tachikoma implementation details

---

Last Updated: 2026-02-20