Researchers have introduced SGNE-CC v2, a weight-editing method for the DeepSeek-R1-Distill-Qwen-1.5B model. The technique modifies 205 columns in the mlp.down_proj layer, which accounts for 0.0177% of the model's parameters. This approach aims to reduce repetitive reasoning loops and incomplete responses without requiring retraining or code changes.

The method was tested across five benchmarks, including MATH-500, MBPP, HumanEval, ARC-Challenge, and BBH Logic, using a budget of up to 32,768 tokens. At a temperature of T=0, the number of incomplete responses on MATH-500 dropped from 139 to 16, while the average response length decreased from 11,000 to 4,700 tokens.

On the MBPP benchmark, the edited model achieved a performance of 81/120 at T=0, matching the original model's performance at T=0.6, while producing responses 3.7 times shorter. In the ARC-Challenge, the modification eliminated incomplete reasoning across all tested temperature settings.

The SGNE-CC v2 method is a refinement of the original SGNE-CC approach, which previously required editing 1,984 columns. The new version demonstrates that spherical rotation of output columns can be applied to pure transformer architectures, such as the 1.78 billion parameter DeepSeek-R1-Distill-Qwen-1.5B, with significantly fewer modifications.