David Noel Ng, an independent researcher operating under the handle dnhkng, has achieved top rankings on the HuggingFace Open LLM Leaderboard using an unconventional method that requires no model training or additional GPU memory. His technique, called 'RYS' (Repeat Yourself), involves strategically duplicating specific transformer layers within existing models to boost performance across multiple benchmarks.
The RYS Method Reveals Transformer Architecture Insights
Ng's approach centers on duplicating layers 45-51 of the Qwen2-72B model, creating a 78B parameter variant. The method achieved an average benchmark improvement of 44.75% (+2.61%), with particularly strong gains on difficult reasoning tasks: MATH Level 5 performance increased to 38.97% (+8.16%), while MuSR scores jumped to 23.72% (+17.72%). The entire process was executed on just two RTX 4090 gaming GPUs.
The key discovery involves how transformer layers function as indivisible reasoning "circuits." Ng found that duplicating individual middle layers degraded performance, but replicating an exact block of seven consecutive layers produced significant improvements. This suggests that certain layer sequences work together as cohesive computational units rather than independent transformations.
Impact on the Open LLM Leaderboard
As of early 2026, the effectiveness of Ng's technique is evident in the leaderboard standings: four of the top five models are 78B variants descended from the RYS-XLarge architecture. The method requires no weight modification and uses no additional VRAM beyond what the duplicated layers naturally add to model size.
Ng published a detailed technical writeup on March 10, 2026, explaining the methodology and experimental results. The post gained significant attention on Hacker News with 270 points and 82 comments, reflecting strong community interest in accessible AI research methods that don't require massive computational resources.
Key Takeaways
- Independent researcher David Noel Ng topped HuggingFace's Open LLM Leaderboard by duplicating layers 45-51 of Qwen2-72B, creating a 78B parameter model
- The RYS method achieved benchmark improvements of up to 17.72% on reasoning tasks without any training or weight modifications
- Four of the top five leaderboard models as of early 2026 are 78B variants descended from the RYS-XLarge architecture
- The technique reveals that transformer layers function as indivisible seven-layer "circuits" rather than independent transformations
- The entire method was executed on just two RTX 4090 gaming GPUs, demonstrating accessibility for independent researchers