filipe(@filicroval):Can you make a reasoning model stop looping without just cutting it off? I used an MLX-LM logits processor to penalize rumination continuations inside <think>: “Wait…”, “Actually…”, “double-check…”, “reconsider…” No finetune, no weight changes. Runtime steering. DeepSeek-R1-Distill-Qwen-14B-4bit · GSM8K n=100: baseline: 90% acc · 524 avg tokens · 250 marker hits penalty: 91% acc · 439 avg tokens · 7 marker hits hard cap near same budget: 48% acc · 410 avg tokens · 35 missing </think> It changes how the model gets shorter: fewer self-doubt loops, not chopped reasoning. training-free, repo:

2026.07.09 14:08

Can you make a reasoning model stop looping without just cutting it off? I used an MLX-LM logits processor to penalize rumination continuations inside : “Wait…”, “Actually…”, “double-check…”, “reconsider…” No finetune, no weight changes. Runtime steering. DeepSeek-R1-Distill-Qwen-14B-4bit · GSM8K n=100: baseline: 90% acc · 524 avg tokens · 250 marker hits penalty: 91% acc · 439 avg tokens · 7 marker hits hard cap near same budget: 48% acc · 410 avg tokens · 35 missing It changes how the model gets shorter: fewer self-doubt loops, not chopped reasoning. training-free, repo: