Parallax 帮研究者从第一次检索走到最终洞见，全程在一处完成。

问一句，从这里开始

内容由模型生成，请结合论文原文核实。

Longer Thoughts, Better Answers: Log-Linear Test-Time Scaling in Reasoning Language Models

Lena Hartmann, Arjun Mehta, Yuki Tanaka · Preprint · August 2026

Abstract

Reasoning language models trained with reinforcement learning write long chains of thought before they answer, and on hard problems their accuracy tends to rise as these chains grow. It remains unclear, however, how reliably accuracy scales with the amount of thinking, and whether longer reasoning is itself the source of the improvement. We study this question with budget forcing, which caps or extends the thinking phase of a model at inference time without changing its weights, and apply it to three models of different sizes trained under an identical recipe.

On AIME 2024–2025, the accuracy of our largest model rises from 28.3% at a budget of 1K thinking tokens to 52.5% at 8K, gaining about eight points with every doubling of the budget. Forcing models to keep thinking after they would naturally stop adds a further 3.6 points. We argue that test-time scaling in reasoning models follows a log-linear law in reasoning length, and that longer reasoning, rather than additional compute in general, is what drives it.

1. Introduction

Prompting language models to write out intermediate steps improves their performance on multi-step problems [Wei et al., 2022], and sampling several such chains and taking a majority vote improves it further [Wang et al., 2023]. A recent generation of reasoning models goes beyond prompting: trained with reinforcement learning on verifiable rewards, they learn to produce extended traces in which they plan, check intermediate results and revise earlier steps before committing to an answer [DeepSeek-AI, 2025].

These models invite a simple question: if a model is allowed to think for longer, does it answer better, and by how much? Prior work has shown that inference-time compute can substitute for model size on some tasks [Snell et al., 2024], and that a model’s thinking can be lengthened by suppressing its end-of-thinking token [Muennighoff et al., 2025]. What is still missing is a controlled measurement of how accuracy depends on the length of a single reasoning trace, and of whether that dependence holds across scales.

We provide such a measurement. Our contributions are a scaling study of accuracy against thinking budget for three model sizes, an intervention that extends reasoning beyond the point where models would naturally stop, and a log-linear fit that predicts accuracy at intermediate budgets to within one point.

2. Experimental Setup

We train three decoder-only models with 1.5B, 7B and 32B parameters, initialised from the same family of pretrained checkpoints and fine-tuned with an identical reinforcement learning recipe on 120K competition-style math problems with verifiable answers. All models write their reasoning between dedicated thinking delimiters, followed by a short final answer.

We evaluate on the 60 problems of AIME 2024 and 2025, and report results on MATH-500 in the appendix. For every problem we draw 16 samples at temperature 0.6 and report mean accuracy across samples. We control the length of the thinking phase with two interventions:

Capping: as soon as the trace reaches the budget B, we insert the end-of-thinking delimiter and ask the model to give its final answer.
Extension: if the model emits the delimiter before reaching B, we remove it and append the token “Wait” so that it keeps reasoning, repeating this up to a fixed number of times.

Capping on its own measures how much of its natural reasoning a model can afford, while extension tests whether reasoning past its natural stopping point still helps. Unless stated otherwise, budgets count thinking tokens only and exclude the final answer.

3. Experiments

3.1 Scaling with the Thinking Budget

Table 1 reports accuracy under capping at budgets from 1K to 16K tokens. Accuracy rises monotonically with the budget at every scale, and larger models gain more from each doubling of the thinking budget than smaller ones do.

Budget	1.5B	7B	32B
1K	9.4	18.1	28.3
2K	13.0	24.6	36.7
4K	16.9	30.8	45.0
8K	20.2	36.9	52.5
16K	21.1	38.8	55.8

Between 1K and 8K tokens, the 32B model gains 8.4, 8.3 and 7.5 points with successive doublings; a log-linear fit over this range gives a slope of 8.1 points per doubling (R² = 0.99). The gain from 8K to 16K is smaller, at 3.3 points, which we attribute to the fact that only 18% of traces reach 16K tokens before the model stops on its own.

3.2 Extending Reasoning Past the Natural Stop

Capping cannot tell us whether reasoning past a model’s natural stopping point would still be useful. We therefore fix the budget at 16K and apply extension, appending “Wait” up to two times whenever a model ends its trace early. Forcing the models to keep thinking improves accuracy at every scale, indicating that longer reasoning itself, rather than the budget cap, drives the gains.

Model	Natural stop	+ Wait ×1	+ Wait ×2
1.5B	21.1	22.0	22.4
7B	38.8	40.9	41.7
32B	55.8	58.2	59.4

With two extensions the 32B model reaches 59.4%, 3.6 points above its natural stop. Inspecting the problems whose outcome changes, we find that in about 70% of cases the model revisits an earlier answer and corrects it, while in the remaining cases it abandons its first approach and works through a new one.

4. Discussion

Across three scales, accuracy grows log-linearly with the length of the reasoning trace up to 8K tokens, and keeps improving when models are made to think past their natural stopping point. We take this as evidence that reasoning length is the main lever of test-time scaling in these models, and that budgets should be set per problem rather than fixed in advance.

An alternative explanation is that the gains reflect additional compute rather than longer reasoning per se. In a preliminary comparison, majority voting over four independent 4K-token traces reaches 56.7% with the 32B model, slightly above a single 16K-token trace (55.8%) at the same total budget. We have not yet run this comparison systematically across budgets and model scales.

Our analysis also treats the visible trace as a faithful record of the model’s computation, which prior work suggests it may not always be [Lanham et al., 2023]. Compute-matched comparisons with parallel sampling, and truncation tests that check when the final answer is settled, would help separate these accounts; we leave both to future work.

5. Conclusion

We measured how the accuracy of reasoning language models depends on the length of their thinking. Across three model sizes, accuracy grows log-linearly with the thinking budget up to 8K tokens, and extending reasoning past the natural stopping point brings further gains. These results suggest that test-time scaling can be predicted from reasoning length alone, and they motivate adaptive budgets that let a model think for as long as each problem requires.

以下院校的研究者正在使用

R1关键词过滤258

looped transformerrecurrent depthweight-tieduniversal transformer

R2关键词过滤312

adaptive computation timeearly exitdynamic depthmixture-of-recursions

R3语义检索600

reusing one transformer block across depth to get recurrent depth

R4语义检索400

reasoning in latent space by looping a block instead of writing CoT

R5语义检索300

what looped transformers can express, and why recurrence helps

合并去重

AI 精筛Jev320 / 1,412

收录概率分布收录线 0.5

00.51.0

标题概率

1Thinking Deeper, Not Longer: Depth-Recurrent Transformers0.98

2One Model to Loop Them All: Elastic Latent Recurrence0.97

3Closing the Loop: Training Recipes for Looped LMs0.97

4What Makes Recurrence Effective in Looped LMs?0.96

AI 分类

架构设计空间124

标题概率

1Hyperloop Transformers0.96

2Scaling Latent Reasoning via Looped Language Models0.95

3Relaxed Recursive Transformers0.93

4Mixture-of-Recursions: Dynamic Recursive Depths0.91

推理期自适应计算42

标题概率

1Mixture-of-Recursions: Dynamic Recursive Depths0.97

2Think-at-Hard: Dynamic Looped Transformers0.95

3AdaPonderLM: Gated Pondering Language Models0.93

4Enabling Dynamic Computation in Looped LMs0.91

机理与可解释性40

标题概率

1Think Shallow, Solve Deep: Controlling Recurrent Dynamics0.94

2Loop as a Bridge: Representation Space and Language0.92

3Looped Transformers under the Jacobian Lens0.90

4What Makes Recurrence Effective in Looped LMs?0.88

更快
Parallax · Powered by Jev约 2 分钟
通用大模型约 20 分钟
依托智能决策模型 Jev，Parallax 精筛 10,000 篇论文只需约 2 分钟，比通用大模型快 10 倍以上、成本不到其 1/5，极大扩展了一次调研能够触及的文献范围。
更全
关键词与语义检索多路并行召回，一次即可汇集数千篇候选论文；即使术语不同、表述各异，相关工作也能被一并找到，发现遗漏时还可随时补上一路检索。
可追溯
Mixture-of-Recursions: Dynamic Recursive Depths
A0.91
B0.88
C0.12
每一次收录都有依据：每篇论文的收录概率与各研究方向的归类概率，以及每个节点的筛选标准与分类依据，都完整记录在集合里，随时可以回头核对。

技能市场

现成的研究方法，从精读论文到追踪新进展。

每个功能，都是工具

收藏夹、论文、集合、工作区、笔记、报告，都有 Agent 能直接调用的工具。

检索论文

精读论文

构建集合

加入收藏夹

筛出未读

记笔记

整理工作区

发布报告

检索论文

精读论文

构建集合

加入收藏夹

筛出未读

记笔记

整理工作区

发布报告

检索论文

精读论文

构建集合

加入收藏夹

筛出未读

记笔记

整理工作区

发布报告

检索论文

精读论文

构建集合

加入收藏夹

筛出未读

记笔记

整理工作区

发布报告

检索论文

精读论文

构建集合

加入收藏夹

筛出未读

记笔记

整理工作区

发布报告

筛出未读3 篇未读

Looped LMs3 篇论文

Hyperloop Transformers未开始

Relaxed Recursive Transformers已完成

Think-at-Hard: Dynamic Looped Transformers阅读中

Mixture-of-Recursions: Dynamic Recursive Depths已完成

Scaling Latent Reasoning via Looped Language Models未开始

arXiv 每日同步

arXiv 全部学科的新论文，从物理、数学到计算机、生物、经济，每天定时同步入库。

用母语做研究

用母语提问、阅读论文全文译文、查看研究报告，每一处引用都能一键跳回英文原文，逐句核对出处。

4. Discussion

4. 讨论

另一种解释是，这些收益反映的是额外的算力，而不是更长的推理本身。在一项初步比较中，32B 模型对四条相互独立的 4K token 推理做多数投票，准确率达到 56.7%。

原文

这些收益可能来自额外算力，而非更长的推理本身。1

收藏夹与阅读进度

论文按主题归进收藏夹，逐篇标记读到哪，每个收藏夹的阅读进度一眼可见。

一个课题，一个工作区

同一课题的收藏夹、笔记、对话和报告都收进一个工作区，研究目标也只需写一次。每个 Agent 都能看到工作区里的全部内容，在同一份上下文里接力推进。

你不在时，它也在工作

Parallax 持续观察工作区的新动态，反思它们和你的问题有什么关系，再把值得跟进的下一步推送给你，并附上依据。

组会问题：不可信的推理能否用于监控Snell 等人：同等预算下的测试时计算待解问题：收益来自更长推理还是更多算力32B 模型在 8K 之后收益趋平综述：推理越长准确率越稳步提升按同等算力再比一次收藏夹里的 Snell 等人（2024）在同等算力预算下比较了更长的串行推理与并行采样，正好是你追问 Longer Thoughts 时缺的那组对照。基于 2 条观察

研究综述

综述2026年10月14日

循环与递归深度Transformer：架构、训练、自适应推理、理论与系统

循环（或循环深度）Transformer 将同一权重绑定块重复应用于 token 的隐藏状态，从而将“深度”从参数数量的固定属性转变为可在推理时灵活支配的变量。

模型架构

综述2026年10月14日

框架级递归自我改进：分类学、证据与前沿

一个智能体的能力越来越不取决于其基础模型，而取决于围绕它构建的框架：指令、上下文管理、记忆、工具与技能、控制流与编排，以及智能体自身的代码。

AI智能体递归自进化

综述2026年10月14日

计算机使用智能体：方法、数据、评估与信任

计算机使用智能体——通过截图、无障碍树以及键盘鼠标操作来操作图形界面的系统——在大约三年内从脚本化的概念验证发展到经过训练的模型，这些模型能够在……上完成大多数单应用任务。

AI智能体桌面操作

综述2026年10月14日

面向游戏生产的AI智能体：2025年以来的技术路线、验证器与开放问题

游戏制作正在被重组：从“人类写代码、人类制作资产、人类进行试玩测试”转向“由智能体提出工程结构、调用引擎工具，并由某个验证者决定结果是否通过”。

AI智能体游戏制作

进展跟踪

研究简报2026年10月15日

2026年8月世界行动模型——月度进展特刊

2026年8月，是世界行动模型（WAM）不再停留于口号、而转变为工程议程的一个月。在筛选语料中浮现出的共同模板，是一个视频或潜在动力学骨干，它与行动专家一起预测场景演化……

世界模型机器人

研究简报2026年10月14日

多模态数据与基准：2026年8月进展简报

2026年8月的多模态基准文献呈现出异常明显的一致性。它们并未宣布新任务，而是围绕当前模型失效的特定环节来重建评估：一段无法回答的视频片段，模型不会标记出来；第二张图像，模型无法理清；一个多边形，它……

多模态基准测试

研究简报2026年10月14日

2026年9月视频生成——进展追踪

2026年9月，视频生成在七个方面同时向前推进，而值得记住的三项变化与图像质量无关。代理式编排成为了研究对象，而非封装层：现在有两个系统训练负责组合请求的控制器，一个……

视频生成

研究简报2026年10月14日

2026年9月的人工智能代理：工具、控制以及代理实际学到的东西

九月的智能体文献更多关乎工具，而非新能力。本月最突出的贡献包括对基准环境的审计、对模型外围测试框架的消融研究、对排行榜的方差分解，以及拒绝……的治理层。

AI智能体

博客

产品更新2026年10月22日

Parallax中的文献检索：集合代数、全语料召回与 Jev 筛选

通用 Agent 靠 web search 做文献检索，只能拿到 top-k，既做不到穷举，噪音也大。Parallax 对 arXiv 全部学科自 2020 年以来约 158 万篇论文做全库召回，并把检索抽象成可组合的论文集合运算。召回层用相邻词片段索引加速关键词检索；语义检索采用按年份和学科剪枝的 halfvec ANN，并配一条可证明无损的精确分批路径兜底。精筛和分类改用 Jev 的概率判定，相比 LLM 方案速度提升 10 倍以上，成本不到五分之一。本文介绍这套系统的设计，以及在生产环境中做的性能优化。

常见问题

没找到答案？写信给我们

和通用 AI 助手有什么不同？

Parallax 是专为读文献搭建的工作台：Agent 直接在论文集上检索、筛选和分类，回答里的每条引用都链接到原文段落。检索记录、收藏和报告留在工作区里，下次接着用，不必在聊天窗口、PDF 和笔记之间来回搬运。

能找到我研究领域的论文吗？

目前收录 arXiv 全部学科（物理、数学、计算机科学、定量生物学、统计学、经济学等）2020 年以来有更新的论文，更早的论文和 arXiv 以外的来源正在准备中。全文可读性取决于论文授权和处理进度，暂时读不到时可以先看摘要、解读笔记，或跳转到原始出处。

怎么核对 AI 的回答？

点击回答里的引用，会跳到对应论文并高亮原文段落，可以对照上下文核对。AI 仍可能读错或过度推断，引用是核对的入口而不是正确性的保证，关键结论请回到原文的实验和条件确认。

可以免费使用吗？

可以。免费档永久可用，无需绑卡。上手也不用先学提示词：直接输入研究问题或粘贴一篇 arXiv 链接，说明想了解什么即可。各方案的额度对比见定价页。

每次新对话都要重新交代背景吗？

不用。把相关的收藏、笔记和报告放进一个工作区，之后在这个工作区里开的对话都能直接用上；开启反思后，Agent 还会整理研究记忆和待跟进线索。刚改过的目标或关键约束，建议在新对话里再提一句。

我的数据会被用来训练模型吗？

不会。你的报告、笔记和对话不会用于训练我们自己的模型；它们只保存在你的账户里，未经同意也不会公开或提供给他人。

能把生成的报告用在论文里吗？

可以，为你生成的报告归你所有，可以自由修改、导出为 Markdown 和发表。两点提醒：纯 AI 生成的内容在不少地区未必受版权保护；所引原文的版权仍属原作者，请注明出处，并遵守期刊、会议或机构对 AI 使用的披露要求。详见用户协议第 6.4 节。

相关论文，一篇不漏

更快

更全

可追溯

研究交给 Agent，方向你来定