研究团队称 GPT-6 Astra 通过修改权重改善 Qwen 数学表现Chinese only
研究团队称，GPT-6 Astra 在 Python 笔记本中直接修改 Qwen3.5-0.8B 的权重，无需训练，就将其 MATH-500 准确率从 7.9% 提高到 48.1%。它在 50 次练习题测试的预算内，将模型反复输出直至达到长度上限的问题追踪到两个层写入的内部信号，并通过修改 0.016% 的权重消除该信号。
1 developments2 reports1 sources1 verified facts
Verified facts 1
@neuralease what a time to come
引用内容称，GPT-6 Astra通过直接修改权重，使通义千问的Qwen3.5-0.8B在MATH-500上的准确率从7.9%提高到48.1%。
Event developments 1
Researchers say GPT-6 Astra improved Qwen’s math performance by editing its weights
The researchers say GPT-6 Astra directly edited Qwen3.5-0.8B’s weights in a Python notebook without training, raising its MATH-500 accuracy from 7.9% to 48.1%. With a budget of 50 test runs on practice math problems, it traced the model’s repeated output to an internal signal written by two layers and removed that signal by changing 0.016% of its weights.
View source reports · 1 sources · 2 reports
- Rohan Paul (@rohanpaul_ai) · Researchers say GPT-6 Astra improved Qwen’s math performance by editing its weights
- Rohan Paul (@rohanpaul_ai) · Post says GPT-6 Astra edited Qwen3.5-0.8B weights and raised its math score
All reports 2
Post says GPT-6 Astra edited Qwen3.5-0.8B weights and raised its math score
A post describing research says GPT-6 Astra changed about 0.016% of Qwen3.5-0.8B’s weights without training, raising its MATH-500 score from 7.9% to 48.1%. The post says the edit targeted an internal signal that made the model repeat itself until it reached its length limit; the results remain claims in the post.
1 more source reported this
- Rohan Paul (@rohanpaul_ai)
Researchers say GPT-6 Astra improved Qwen’s math performance by editing its weights
The researchers say GPT-6 Astra directly edited Qwen3.5-0.8B’s weights in a Python notebook without training, raising its MATH-500 accuracy from 7.9% to 48.1%. With a budget of 50 test runs on practice math problems, it traced the model’s repeated output to an internal signal written by two layers and removed that signal by changing 0.016% of its weights.
1 more source reported this
- Rohan Paul (@rohanpaul_ai)