Research Overview

The scientific foundation behind Tachikoma's design.

Quick Lookup

Technique	What	Used In	Research
Position-Aware Loading	Optimizes context placement	Context Management	Position Bias
Verification Loops	GVR pattern	Skill Chains	Verification
Reflection Phase	Freedom to revisit, rethink	All skills	Verification
Model-Aware Editing	Optimal edit format	Skill Execution	Model Harness

---

Edit format selection matters as much as model choice. Research in Model Harness. | RLM | Large context handling | Subagents | RLM | | Cost-Aware Routing | Match complexity to strategy | Intent Routing | Cost-Aware | | Skill Composition | Modular architecture | Skill Chains | Modularity |

Research Areas

Position Bias

U-shaped attention bias in transformers.

Papers:

• "Found in the Middle" (Hsieh et al., ACL 2024)
• "On the Emergence of Position Bias" (Wu et al., ICML 2025)

Verification Loops

Why verification beats retries.

Papers:

• "Towards Autonomous Mathematics Research" (arXiv:2602.10177)
• "Accelerating Scientific Research with Gemini" (arXiv:2602.03837)

Model Harness

Edit format selection matters as much as model choice.

Source: Can.ac blog (Feb 2026)

RLM

10M+ token contexts through adaptive chunking.

Paper: "Recursive Language Models" (arXiv:2512.24601)

Cost-Aware Routing

Matching strategy to task complexity.

Paper: "When Do Tools and Planning Help LLMs Think?" (arXiv:2601.02663)

Modularity

Why focused components beat monolithic approaches.

Paper: "Agentic Proposing" (arXiv:2602.03279)

Reading Order

1. Position Bias — Context loading (foundational)
2. Cost-Aware Routing — Speed vs accuracy
3. Modularity — Why skills beat monolithic
4. RLM — Large context (advanced)
5. Verification Loops — Reliability
6. Model Harness — Edit optimization

Caveat

Some claims from papers are reported but not independently verified. Always verify with original sources if you need certainty.

See Also

• Architecture
• Capabilities