新智元报道
通过接口关掉大模型的「思考模式」（如将Qwen3的enable_thinking设为False），再告诉它「不要推理，直接给出答案」，它就会直接回答吗？实际输出并不符合我们日常直觉。
清华大学、牛津大学和斯坦福大学的研究者系统测试了这个问题。他们发现，即使关闭思考模式，模型仍可能在最终答案前写出计算或推导；题目越开放，这种现象越明显。研究者把这种现象称为Thinking Inertia（思考惯性）。
论文链接：https://arxiv.org/abs/2610.11765
项目主页：https://thinking-inertia.github.io/
代码链接：https://github.com/thinking-inertia/code
轨迹数据：https://huggingface.co/datasets/thinking-inertia/llm-nothinking-trajectories
图1 即使明确要求只输出选项，模型仍可能写出解题过程。
判断模型是否在推理，不能只看思考开关或回答长度。一个关闭了思考开关的模型，仍然可能在普通回答区写出计算过程；一段很长的文字，也可能只是重复题目、套用格式，未必包含真正的推理。
于是，研究者把回答拆成前置文字与最终答案，再分三层检查。ETR看模型有没有真的只给答案：答案格式必须合法，前后都不能有实质性文字。QRel.衡量前置文字与问题的相关性，复述题目也可能得分较高。EIR则进一步判断其中有没有计算、比较、排除等显式推理步骤。
负责评判的模型看不到被测模型的身份和实验设置。在600条抽样输出上，五名标注者与自动评判的「是否包含显式推理」二分类平均一致率为98.8%。
图2 三层指标分别检查是否只给答案、文字是否相关、是否包含显式推理。
实验覆盖Qwen3、DeepSeek、GPT和Gemini四个家族的六个模型，在六个基准上测试不同思考设置和作答指令对输出的影响。问题分为判断题、选择题和需要自行生成答案的开放式数学题。
总体而言，只需回答是或否时，模型更容易直接作答；选择题的显式推理比例居中；到了开放题，答案前的可见推理明显增多。更严格的「只给答案」指令会降低显式推理比例，但三类任务之间的总体排序依然存在：开放题 > 选择题 > 是非题。
一个直观例子来自DeepSeek-V4-Flash：关闭思考模式后，它在GSM8K和MATH两项基准上的总体显式推理率仍达99.9%。
这意味着，关闭思考开关，并不能保证模型在最终答案前省略推理过程。
即使最终答案正确，模型仍可能没有遵循「只给答案、不输出推理」的要求；答对题与按要求直接作答，是两个需要分别评价的维度。
差异是否来自题目本身？
研究者从MMLU和MMLU-Pro中选取一批数值答案题，保留题干，改变候选答案和作答要求，并在Qwen3-32B关闭思考模式的设置下比较。
原始选择题的准确率为77.8%，只给答案的比例为10.5%；改成判断题，让模型核对一个给定答案后，准确率仍为78.9%，只给答案的比例却升到62.5%。去掉候选项，让模型自行生成结果时，准确率降到52.6%，只给答案的比例变为0，显式推理比例达到90.1%。
有候选答案时，模型可以核对或选择；没有候选答案时，它需要自己构造结果，也更容易在答案前写出推理过程。
图3 同一批题目改成不同回答形式后，准确率和输出行为随之变化。
要求模型少写过程，还会牵动最终准确率。对五个支持关闭思考模式的模型，在开放式数学题上进一步加入严格的只给答案指令后，平均只给答案比例从接近0提高到约40%，平均准确率下降约15个百分点。判断题和选择题的准确率变化则相对平稳。
因此，能否按要求不输出推理、直接作答，也应成为一项受到重视、被独立评估的模型能力。
这种能力不会随着思考开关的关闭自动获得，评测时需要同时考察输出合规性与答案正确率。
这项工作也引起了社区的后续验证。George Shostakovych 受论文启发，对Gemini-4-Argon、GPT-6-Astra、Kimi-K3、DeepSeek-V4-Flash和Muse-Wonder-1.0进行了小规模独立测试，在其中观察到关闭思考后仍可能输出显式推理的现象，并在博客中报告了结果[2]。
第一作者Dianqiao Lei 是清华大学大四本科生，Kevin Qinghong Lin 与 Philip Torr、James Zou 共同担任通讯作者。
Kevin Qinghong Lin 是牛津大学博士后学者；Pan Lu 是斯坦福大学博士后学者；Philip Torr 是牛津大学工程科学系教授；James Zou 在斯坦福大学计算机科学系和电气工程系任副教授。
论文已被 NeurIPS 2026 接收，论文、代码和数据已公开。
参考资料：
[1] Thinking Inertia: LLMs Keep Thinking When Told Not To https://arxiv.org/abs/2610.11765
[2] George Shostakovych. 'Answer only' is a request, not a switch. 2026年10月15日。 https://gshost.dev/writing/answer-only-is-a-request/
编辑：LRST
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
Turning off LLM's "thinking mode" through the interface (e.g. setting the Qwen3 environment thinking as False) and telling it, "Don't reason, just give the answer," will it answer? The actual output does not match our daily instincts.
The researcher system at Tsinghua University, Oxford University and Stanford University tested the problem. They found that even when thinking patterns are closed, models may write calculations or extrapolations before the final answer; the more open the topic, the more obvious it becomes. Researchers call this phenomenon Thinking Inertia.
Paper link: https://arxiv.org/abs/2610.11765
Project homepage: https://thinking-inertia.github.io/
Code link: https://github.com/tinking-inertia/code
Track data: https://huggingface.co/datasets/tinking-inertia/llm-nothining-travers
Figure 1 Even if only the output options are explicitly required, the model may have a problem-solving process.
Whether the model is deduced cannot be judged solely by thinking about the switch or the length of the answer. A model that shuts down the thought switch may still be able to write calculations in the ordinary response area; a long text may simply repeat the subject, format and not necessarily contain true reasoning.
The researchers then split the answers into pretexts and final answers, with three layers of examination. ETR to see if the model really gives only the answer: the answer must be in a legal format, without substantive text. QRel. Measuring the relevance of the prefix to the problem, the recapitulation of the title may also score high. EIR further determined that there were no obvious reasoning steps such as calculation, comparison, exclusion.
The model responsible for the evaluation does not see the identity and experimental settings of the detected model. For the 600 sample outputs, the average consistency rate between the five labelers and the self-judgmentation of the two categories "includes explicit reasoning" was 98.8 per cent.
Figure 2 Three tiers of indicators examine whether answers are given only, whether the text is relevant or whether it contains explicit reasoning.
The experiment covered six models for the families of Qwen3, DeepSeek, GPT and Gemini, testing the impact of different thinking settings and response instructions on output on six benchmarks. The questions are divided into judgments, selections and open mathematical questions that require self-generated answers.
In general, it is easier for the model to respond directly if the answer is yes or no; the obvious reasoning of the choice is in the middle; and the visible reasoning before the answer is significantly increased when the open question is reached. The stricter "Give the answer only" directive reduces the proportion of obvious reasoning, but the overall ranking between the three types of tasks still exists: the open question > selection > is not.
An intuitive example comes from DeepSeek-V4-Flash: after closing down the thinking pattern, its overall apparent reasoning on the GSM8K and MATH benchmarks is still 99.9 per cent.
This means that closing the thought switch does not guarantee that the model omits the reasoning process before the final answer.
Even if the final answer is correct, the model may not comply with the requirement of "only answer, not output reasoning"; answering the question and responding directly to the request are two dimensions to be evaluated separately.
Is the difference from the title itself?
The researchers select a number of numerical answers from MMLU and MMLU-Pro, keep them dry, change candidate answers and answer requests, and compare them with the Qwen3-32B closed-think mode settings.
The accuracy of the original selection was 77.8 per cent, whereas the proportion of answers given was 10.5 per cent; the ratio of accuracy that allowed the model to check a given answer was still 78.9 per cent, while that of answers only rose to 62.5 per cent. When the candidate is removed and the model produces its own results, the accuracy rate is reduced to 52.6 per cent, the proportion given to the answer is zero and the proportion of the obvious reasoning is 90.1 per cent.
When there are candidate answers, the model can check or select; when there are no candidate answers, it needs to construct the results itself and it is easier to write the reasoning process before the answer.
Figure 3 Change in accuracy and output behaviour following the change of the same title to a different form of response.
Requires the model to be underwritten and will also influence the final accuracy rate. For the five models that support the closure of thinking patterns, the average answer rate increased from close to 0 to about 40 per cent, and the average accuracy rate decreased by about 15 percentage points, following the further addition of strict answer-only instructions to open mathematical questions. The change in the accuracy of the judgments and selections is relatively smooth.
Therefore, the ability not to output reasoning and respond directly as required should also be a valued and independently evaluated modelling capability.
This capability will not automatically be acquired with the thought switch closing, and the assessment will need to be accompanied by an examination of output compliance and the correctness of the answer.
This has also led to community follow-up validation. George Shostakovych, inspired by the paper, conducted small-scale independent tests of Gemini-4-Argon, GPT-6-Astra, Kimi-K3, DeepSeek-V4-Flash and Muse-Wonder-1.0, in which he observed that it was still possible to export the obvious reasoning after closing for thought and reported the results in his blog [2].
The first author, Diangqiao Lei, is a senior undergraduate student at the University of Qinghua, Kevin Qinghong Lin, and Philip Torr and James Zou are co-authors of the newsletter.
Kevin Qinghong Lin is a post-doctoral scholar at Oxford University; Pan Lu is a post-doctoral scholar at Stanford University; Philip Torr is a professor at the Faculty of Engineering Science at Oxford University; and James Zou is an associate professor at the Faculty of Computer Science and Electrical Engineering at Stanford University.
The papers have been received by NeuurIPS 2026 and the papers, codes and data have been made public.
References:
[1] Thinking Inertia: LLMs Keep Thinking When Told Not To https://arxiv.org/abs/2610.11765
George Shostakovych. 'Answer only' is a request, not a switch. https://gshost.dev/wring/answer-only-is-a-request/
Edit: LRST
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.