Editing 0.0177% of DeepSeek-R1-1.5B weights cuts unclosed reasoning chains from 305 to 41
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Editing 0.0177% of the weights without fine-tuning cut average reasoning length from 10.8 thousand to 4.0 thousand tokens and the number of unclosed responses from 305 to 41; the patch and scripts are open.
A methodology called SGNE-CC v2 has been posted on Habr: rotating 205 columns of mlp.down_proj (0.0177% of parameters, 0.58 MiB) in DeepSeek-R1-Distill-Qwen-1.5B without fine-tuning. The author published open artifacts: a weights patch, traces, a SHA-256 certificate, an archive on Zenodo and an Apache-2.0 license, so the claimed numbers can be checked against the files. With greedy generation (T=0) on 1,054 tasks (MATH-500, MBPP, HumanEval, ARC-Challenge, BBH Logic), correct answers rose to 733 from 638. Responses without a closing reasoning marker numbered 41 versus 305, and average length fell from 10.8 thousand to 4.0 thousand tokens. On MATH-500, accuracy rose from 65.4% to 76.4%, and unclosed reasoning chains dropped to 16 from 139. At T=0.6, the accuracy gain is not statistically significant (73.3% versus 72.3%, p = 0.4472); the effect shows up as a 33.6% reduction in length. This is a solo author's independent work on Habr, done, by his own account, in collaboration with SI, rather than a peer-reviewed paper. The figures are claimed by the author, there is no independent replication, and the original model was not measured at T=0.3.
Sources 1