An experiment applies logit penalties for specific overthinking tokens to Qwen models using llama.cpp, improving accuracy on the MATH-500 benchmark across various quantizations, with notable accuracy gains and reduced reasoning tokens.
Meta came out with a banger paper https://arxiv.org/pdf/2606.00206, but it did not look at various quantizations supported in llama.cpp. So I did a run on 50 random MATH-500 questions (https://huggingface.co/datasets/HuggingFaceH4/MATH-500) and ran it on various quantizations of https://huggingface.co/bartowski/Qwen_Qwen3.5-4B-GGUF and tried --logit-bias 466-2 --logit-bias 694-2 --logit-bias 1362-2 \ --logit-bias 1412-2 --logit-bias 1921-2 --logit-bias 1990-2 \ --logit-bias 2086-2 --logit-bias 2361-2 --logit-bias 2441-2 \ --logit-bias 2493-2 --logit-bias 2892-2 --logit-bias 3222-2 \ --logit-bias 3315-2 --logit-bias 3384-2 --logit-bias 3404-2 \ --logit-bias 3482-2 --logit-bias 3655-2 --logit-bias 4213-2 \ --logit-bias 4370-2 --logit-bias 4598-2 --logit-bias 4611-2 \ --logit-bias 4808-2 --logit-bias 5752-2 --logit-bias 6970-2 \ --logit-bias 7014-2 --logit-bias 7643-2 --logit-bias 8106-2 \ --logit-bias 10179-2 --logit-bias 10451-2 --logit-bias 11746-2 \ --logit-bias 13264-2 --logit-bias 13428-2 --logit-bias 14673-2 \ --logit-bias 15029-2 --logit-bias 16036-2 --logit-bias 21143-2 \ --logit-bias 21979-2 --logit-bias 33955-2 --logit-bias 35999-2 \ --logit-bias 36563-2 --logit-bias 37201-2 --logit-bias 37781-2 \ --logit-bias 41484-2 --logit-bias 62586-2 --logit-bias 66073-2 \ --logit-bias 73071-2 --logit-bias 84485-2 --logit-bias 85152-2 \ --logit-bias 95500-2 these correspond to the paper's overthinking markers: [ " perhaps", " maybe", " wait", " Wait", " actually", " hold", " Hmm", " hmm", " Alternatively", " alternatively", " However", " however", " instead", " Instead", " But", " but", " though", " although", " yet", " rather", " unless", " otherwise", " nonetheless", " nevertheless", " regardless", " still", " anyway", " Or", " or", " either", " whether", " uncertain", " unsure", " possibly", " might", " could", " another", " different", " reconsider", " rethink", " backtrack", " retry", " revisit", " doubt", " confused", " wrong", " mistake", " error", " incorrect" ] Here are the results, surprisingly even BF16 leads to better accuracy. Caveats being this is one test on one model. Try it out and see it helps! Format Accuracy: baseline → penalty Reasoning tokens BF16 74% → 84% −19.4% Q8_0 76% → 80% −11.0% Q4_K_M 60% → 66% −14.8% Q3_K_M 52% → 66% −17.5% Q2_K 12% → 24% −11.5%
The author shares a quantization recipe for Qwen3.6 27B that makes the model use significantly fewer thinking tokens while still producing correct answers, leading to faster inference on math benchmarks.
The article argues that Qwen 3.8 27b's increased reasoning token usage is similar to other Chinese AI models like GLM and DeepSeek, with user frustration stemming from hardware limitations. It suggests using a reasoning budget can maintain performance over Qwen 3.6.
The article discusses Qwen 3.8 27B, a 27B parameter model that uses extensive reasoning tokens to compete with larger models, emphasizing trade-offs in token usage and benefits for local deployment.
Llama.cpp introduces adaptive speculation to dynamically adjust token prediction for faster inference, achieving up to 50% speed improvement, particularly for models like Qwen3.8.