New 'RYS' Method Boosts AI Performance Without Training a Single Weight
- What: Release of "LLM Neuroanatomy II," detailing the RYS (Repeat Your Self) optimization method.
- Key Discovery: Evidence of a "Universal Thinking Space" in middle layers where multiple languages converge into identical mathematical representations.
- Technical Scale: Research involved 3,024 beam search candidates and a surrogate model scoring 2 million configurations.
- Results: Previous RYS implementation pushed Qwen2-72B to the #1 spot on the HuggingFace Open LLM Leaderboard without retraining weights.
Independent researcher David Noel Ng has released new findings on "LLM Neuroanatomy," demonstrating that duplicating specific internal layers—a method called "Repeat Your Self" (RYS)—can significantly boost model performance without any additional training or weight changes. The research, published in March 2026, provides mathematical evidence of a "Universal Language" space within Large Language Models (LLMs) where disparate inputs, from English to Base64 code, converge into a format-agnostic reasoning state.
The Evolution of RYS and LLM Neuroanatomy
The findings build upon Ng’s mid-2024 discovery that duplicating a block of seven middle layers in the Qwen2-72B model could propel it to the top of the HuggingFace Open LLM Leaderboard. This original experiment was conducted using only a pair of RTX 4090 GPUs and "hard math probes." The success of the RYS method suggested that LLMs possess a modular internal structure that can be "hacked" to increase reasoning capacity without the massive computational cost of traditional fine-tuning or pre-training.
In this latest update, Ng addresses whether the RYS effect was a "fluke" of earlier architectures or a general property of modern Transformers. Utilizing a high-end dual Grace-Hopper system, the researcher scanned a new generation of open-source models, including Qwen3.5, MiniMax, and GLM-4.7. The focus shifted to the Qwen3.5-27B model, which Ng identifies as the "sweet spot" for the LocalLLaMA community—large enough to possess complex internal circuits but small enough for individual developers to run RYS variants.
Mapping the 'Universal Thinking Space'
The core of the research involves the "ERD" hypothesis: the idea that LLMs function through three distinct phases: Encoding, Reasoning, and Decoding. While early layers (encoding) handle the specific surface form of a language and late layers (decoding) prepare the output, the middle layers appear to operate in a universal, format-agnostic space.
To prove this, Ng expanded on experiments initially suggested by researcher Evan Maunder. By feeding semantically identical sentences in English, Mandarin, and Base64 through a model, the researchers measured the cosine similarity of hidden states at every layer.
The data revealed a striking pattern:
- Encoding Phase: Rapid convergence in the first few layers.
- Reasoning Phase: Near-perfect similarity through the middle layers, regardless of the input language.
- Decoding Phase: Divergence in final layers as the model translates internal "thoughts" back into a specific surface language.
Ng’s latest dataset pushed this further, testing across eight languages—including Arabic, Russian, Japanese, Korean, Hindi, and French—across diverse topics like science, poetry, history, and medicine. The results confirmed that two sentences about the same topic (e.g., photosynthesis) are mathematically more similar in the middle layers than two sentences in the same language about different topics.
Technical Execution and Optimization
To find the optimal "relayering" configuration for Qwen3.5-27B, Ng employed a massive search strategy. The process involved 3,024 beam search candidates and a surrogate model that evaluated approximately 2 million different layer configurations.
This rigorous approach aimed to answer whether relayering still provides a benefit to "stronger" modern models that are already highly optimized. According to Ng’s report, the "short answer is yes, relayering survives." The researcher found that while smaller models often have more "entangled" functional anatomy, the circuit structure responsible for RYS performance remains robust.
Impact for Developers and the AI Industry
The implications of the RYS method are significant for the efficiency of AI deployment. By identifying which layers "earn their extra layers," developers can effectively expand a model's "brain" and reasoning depth at the cost of inference latency rather than the multi-million dollar costs associated with training.
For the open-source community, this provides a blueprint for "upcycling" existing models. If the middle layers truly function as a universal reasoning engine, the industry may shift away from language-specific optimization and toward expanding these "universal" blocks to achieve higher intelligence.
"The middle layers really operate in a universal 'thinking space' where two sentences about the same topic are more similar than two sentences in the same language about different topics," Ng noted, marking a shift in how we understand the internal logic of Transformers.
What’s Next
Ng has released the scanning code and a new set of RYS models based on the Qwen3.5 architecture. The research is ongoing, with MiniMax M2.5 and other modern models currently being processed. As compute-heavy "Hopper" systems continue to grind through configurations, the goal is to determine if these "motifs" of neuroanatomy can be stacked to create even more powerful, training-free model variants.
The researcher also hinted that as models grow beyond 200B parameters, the separation of encoding, reasoning, and decoding becomes even cleaner, potentially allowing for even more surgical "hacking" of the Transformer architecture.
Sources
All technical specifications, pricing, and benchmark data in this article are sourced directly from official announcements. Competitor comparisons use publicly available data at time of publication. We update our coverage as new information becomes available.