我的征尘是星辰大海。。。
The dirt and dust from my pilgrimage forms oceans of stars...
-------当记忆的篇章变得零碎,当追忆的图片变得模糊,我们只能求助于数字存储的永恒的回忆
作者:黄教授
手机视频列表
大模型深度思考推理侧现状分析
视频
音频
原始脚本
大模型深度推理能力现状全扫描,推理侧与训练侧技术路径、现状及问题。 大模型的深度推理本质是模拟人类解决复杂问题的逻辑链,如数学计算、因果分析、多步骤决策。 当前行业主要从推理侧优化,利用外部引导让模型会用能力和训练侧优化,通过技术让模型自带能力两大方向突破。 二者各有路径、现状与瓶颈,具体如 一、推理侧优化,靠外部引导激活模型推理。 核心是 Prompt Engineering 与协作框架。 推理侧不改变模型本身参数,而是通过设计引导策略,或多智能体协作流程。 让模型在现有能力基础上,更有序、更精准的输出逻辑链,避免直接跳步或答非所问。 核心是解决简单问题不用复杂流程,复杂问题不遗漏步骤的适配性问题。 一、核心技术路径与行业现状推理侧的技术进化,本质是从单一提示到动态协作的升级。 目前已形成三类主流路径,覆盖从简单到复杂的推理场景。 路径一,基础 Prompt Engineering,单轮提示,适配简单推理,核心逻辑。 通过指令模板或思维链示例,Few Shot COT,直接告诉模型该按步骤想,无需额外工具或多轮交互。 适用于步骤固定、无需外部信息的问题,如小学数学题、简单因果分析。 具体做法,对需要步骤的问题,在提示词中加入请分三步分析。 一、先明确题干条件。 二、计算关键数据。 三、推导结论。 对容易跳步的问题加入示例。 如例,已知 A 等于3,B 等于5,求 A 加 B 乘2。 第一步算 B 乘2等于10,第二步算 A 加10等于13。 你的解题步骤,括回。 代表模型,所有具备基础逻辑能力的大模型均支持。 如 GPT 3.5、Cloud 2、文心一言、通义千问,是目前最普及的推理方式,无需额外配置,仅靠提示词即可激活。 路径二,工具增强型推理,多轮调用,适配需外部信息的推理,核心逻辑,当推理需要实时 数据、私有知识或精准计算时,如2024年某城市 GDP 同比增长多少?根据公司内部合同条款,某条款是否合法?让模型自主判断需要调用工具,并通过多轮交互完成推理。 先调用搜索 RAG 工具获取数据,再基于数据推导结 论。 具体做法,一、模型先分析用户问题,判断是否需要外部信息。 如孔子是否活着,无需工具。 2024年某城市 GDP 需要工具。 二、若需要调用指定工具,如搜索引擎、向量数据库,获取数据后,再结合数据输出推理结论。 三、若工具返回数据不完整模型可二次发起工具调用。 如搜索结果未包含某区数据,请补充查询该区域 GDP。 代表模型与产品,GPT 4加插件,如 Browse with Bing、Code Interpreter,可自主调用搜索计算工具,解决需要实时专业数据的推理问题。 阿里云通义千问工具链,支持调用企业内部 RAG 知识库、计算器适配,企业及推理场景如财务报销规 则解读,产品参数对比,讯飞星火多轮工具调用,在数学计算中,可自动调用计算器,避免计算错误,在基于计算结果推导应用题答案。 路径三,多智能体协作推理,分工协作,适配超复杂推理,核心逻辑。 将复杂推理任务拆解为多个子任务,让不同角色模型分工完成。 如小模型负责拆解问题,检索模型负责找数据,大模型负责整合结论。 模拟人类团队写作,解决单模型无法覆盖全流程的问题。 如商业方案撰写、科研论文分析。 具体做法,一,任务拆解智能体,通常是轻量模型。 如拉玛尔杠7B,将用户需求拆解为子任务。 如写一份某产品市场分析报告,拆解为一、竞品销量数据查询二、目标用户画像分析三、市场规模测算。 二、工具调用智能体。 根据子任务调用对应工具,如调用电商平台 API 查竞品销量,调用调研数据库查用户画像。 3、结论整合智能体。 大模型如 GPT 4、DeepMind CAR1,将子任务结果整合为完整逻辑链,输出最终答案。 代表案例,Meta 的 Toolformer,通过训练让模型学会调用工具的时机,并与其他模型协作完成推理。 国内,智谱清言多智能体支持自定义角色。 如分析师、数据源、编辑、分工。 完成复杂任务,如分析师拆解市场任务、数据源查数据、编辑润色报告。 开源项目 AutoGPT 通过目标设定、子任务拆解、工具调用、结果反馈的闭环,自主完成长周期推理任务,如帮我规划一场欧洲旅行,包含签证流程。 形成路线,预算计算。 二,推理测的核心问题与瓶颈问。 题一,何时用流程的判断依赖外部框架,模型自主决策弱。 目前多数场景需要人工预设规则,如所有涉及实时数据的问题,必须调用搜索,而非模型自主判断。 例如,若未预设规则,模型可能对2024年 GDP 直接输出旧数据,依赖于训练知识,而非主动调用工具。 问题二,多轮协作的容错性差,若某一环出错,如工具返回错误数据、子任务拆解遗漏,后续推理会连锁出错。 且模型难以自主修正,如搜索到错误的 GDP 数据,模型会基于错误数据推导结论,无法识别数据本身有误。 问题三,效率与成本矛盾复杂协作流程,如多轮工具调用加多智能体分工,会导致响应速度变慢。 且大模型调用成本增加,不适用于实时性要求高的场景,如客服对话。 大模型深度推理能力现状全扫描,推理侧与训练侧 技术路径、现状及问题。 二、训练侧优化,让模型自带推理能力。 核心是从数据驱动到机制优化。 训练侧通过修改模型结构、优化训练目标或改进数据方式,让模型在预训练或微调阶段就掌握推理逻辑,无需依赖外部提示即可自主输出思维链。 目前行业仍处于从模仿推理 SFT 到自主推理算法优化的过渡阶段,尚未实现一劳永逸。 一、核心技术路径与行业现状训练册的技术进化,是从靠数据为模板,到靠算法改机制的突破。 目前有四类关键路径。 路径一,监督微调 SFT。 让模型模仿推理步骤,当前最成熟。 核心逻辑,用带思维链的标注数据,如解题步骤加答案、分析逻辑加结论,微调模型,让模型学会按步骤输出。 本质是照猫画虎式的推理模仿。 具体做法,构建数据集,收集大量问题加详细推理过程加答案的数据如 数学题解题步骤,法律案例分析逻辑,微调训练。 用该数据集对预训练模型,如 LLaMA 2、GPT 3进行微调,让模型学习输入问题输出推理链输出答案的映射关系。 代表模型 Deepseek Math,用200万加数学题加解题步骤数据微调。 在中小学数学竞赛题中,可自主输出推理步骤,无需提示分步骤。 百度文心一言 V4,通过中文思维链数据集微调,在中文因果分析、逻辑推理任务中,无需额外提示即可输出分析过程。 Anthropic 的 claude 2,用长文本推理数据集,如小说情节分析、论文逻辑拆解、微调,支持自主输出长链推理。 局限性,依赖高质量标注数据,且泛化性差。 用数学题数据微调的模型,在法律推理任务中可能失效。 同时,模型仅学会模仿数据中的推理模板,遇到数据中没有的新场景,如跨领域问题,仍会跳步或出错,无法真正自主推理。 路径二,思维链预训练,CoT Pre-training,让模型在预训练阶段学推理,探索中核心逻辑。 在模型预训练阶段,而非微调阶段,就加入思维链类数数据,如百科词条的逻辑解释、学术论文的推导过程。 让模型从学习 语言规律升级为学习逻辑规律,从源头掌握推理能力。 具体做法,扩展预训练数据。 在传统文本、新闻、书籍中加入带逻辑链的文本,如维基百科中的原理解释部分、学术论文的实验设计与推导章节。 调整预训练目标,除了预测下一个 token,额外加入预测推理链连贯性的目标,如判断步骤 A、步骤 B 是否符合逻辑。 代表模型,Google Palm 2,预训练阶段加入数学推导、代码逻辑、科学原理等思维链数据,在多语言推理任务中表现优 于前代,华为盘古大模型推理版,预训练时融入工程计算、物理公式推导数据,在工业场景推理,如设备故障排查逻辑,中可自主输出分析链。 局限性,四维炼域训练数据稀缺,高质量逻辑链文本远少于普通文本,且训练成本极高,需重新训练千亿参数模型。 目前仅少数巨头 Google、华为有能力尝试。 路径三,强化学习,RALF, RAL with feedback. 让模型修正推理错误,辅助优化核心逻辑。 通过人类反馈或机器反馈,让模型在训练中知道自己错在哪,并修正推理逻辑,本质是从错误中学习。 具体做法一、生成推理结果,让模型输出某问题的推理链,如解题步骤。 二、反馈打分,由人类或评判模型对推理链打分,如步骤是否完整、逻辑是否正确。 三、强化训练,用打分结果优化模型,让模型更倾向于输出高分推理链,减少错误步骤。 代表模型 GPT 4,通过人类标注员对推理链打分的 R L H F 训练。 减少推理跳步和逻辑矛盾,如之前算错的数学题,训练后可修正步骤。 商汤日日新大模型用机器评判模型,基于高质量推理数据训练,自动对推理链打分,降低人工标注成本,提升训练效率。 局限性,反馈质量决定效果。 若人类标注员对推理逻辑的判断不统一,如法律案例的不同解读,模型会陷入混乱。 且 RL 训练容易过拟合反馈数据,遇到新问题仍会出错。 路径四,算法机制优化,从统计学预测到逻辑化推理,根本性突破方向,核心逻辑。 不依赖数据,而是通过修改 Transformer 架构或训练目标,解决模型靠概率预测 token,而非逻辑推导的本质问题,让模型具备自主探索推理路径的能力。 目前最关键的技术是多路径推理,Multi-Path Reasoning。 具体做法,以多路径推理为例。 一,多路径生成模型,对同一个问题,生成多个不同的推理路径。 如解数学题时,生成三种不同的解题思路。 二、路径评估,通过逻辑一致性判断或结果验证,筛选出正确的推理路径,如三种解题思路中哪种步骤无矛盾,结果正确。 三、强化学习,让模型更倾向于生成正确路径,逐步掌握自主选择合理推理方向的能力。 代表技术与模型,Google,verification,COVE。 模型先输出推理链,再回头检查每一步是否正确。 若发现矛盾,则修正路径。 在事实性推理,如某事件时间线中降低幻觉。 开源模型 MiniCPM 2B,通过多路径采样加逻辑打分机制,在小参数模型20亿参数中实现自主推理,无需依赖大模型。 微软 ResNet 在 Transformer 中加入推理状态跟踪模块,让模型记录已完成的推理步骤,避免重复或遗漏,适用于长链推理,如科研论文推导。 现状,这类技术仍处于实验室阶段,尚未大规模落地,多路径生成会增加训练和推理成本。 且路径评估的准确性依赖复杂的逻辑判断模型。 目前仅在数学计算、简单逻辑推理等确定性场景中有效,在模糊性问题,如商业决策分析中表现不佳。 二、训练册的核心问题与瓶颈问题。 一、自主推理尚未实现,仍依赖数据模板或局部优化。 目前最成熟的 SFT 和 RLHF,本质仍是数据驱动,模型仅学会模仿已知推理模式,无法像人类一样面对全新问题时自主探索推理路径,如 AlphaGo Zero 那样仅靠规则自主学习。 问题二,Transformer 架构的概率本质与推理逻辑存在矛盾。 模型的核心是预测下一个 token 的概率,而非基于逻辑推导。 即使输出正确推理链,也可能是概率上最可能的序列,而非逻辑上最正确的序列,导致看似有逻辑,实则靠运气的幻觉。 问题三,训练成本与效果的平衡难题,根本性的机制优化,如多路径推理、思维炼狱训练。 需要千亿级参数模型、加海量高质量数据、加超高计算资源,仅少数巨头可承担,中小公司难以跟进。 且效果提升与成本投入不成正比。 部分技术如多路径推理,成本增加10倍,效果仅提升5%。 大模型深度推理能力现状全扫描,推理测与训练测技术路径、现状及问题。 三、总结当前大模型推理的核心结论。 一、推理测加训练测混合式主流,纯训练测一劳永逸尚未实现。 目前行业普遍采用训练测打底,SFT 加基础,RLHF 加推理测优化,Prompt 加工具调用的模式。 训练让模型具备基础推理意识,推理测通过引导和工具弥补泛化性不足,如 GPT 4,Deepseek CAR。 均采用此模式。 二、推理测的核心突破点是自主决策,训练测的核心突破点是机制优化。 未来推理测需要解决模型自主判断是否用工具协作,无需人工预设规则。 训练测需要解决 Transformer 概率本质与逻辑推理的矛盾,如通过多路径推理、逻辑模块嵌入等技术,才能实现模型自主、精准、高效推理。 3短期看混合模式够用,长期看,纯训练 测突破是关键。 对于多数场景,如客服、办公、教育、训练测 SFT 加推理测工具调用,也能满足需求。 但对于高可靠性场景,如医疗诊断、法律判决、自动驾驶决策,必须依赖训练测的根本性突破,让模型具备可解释、无幻觉、自主修正的推理能力。 这也是行业未来3~5年的核心方向。
修正脚本
大模型深度推理能力现状全扫描,推理侧与训练侧技术路径、现状及问题。 大模型的深度推理本质是模拟人类解决复杂问题的逻辑链,如数学计算、因果分析、多步骤决策。 当前行业主要从推理侧优化,利用外部引导让模型会用能力和训练侧优化,通过技术让模型自带能力两大方向突破。 二者各有路径、现状与瓶颈,具体如 一、推理侧优化,靠外部引导激活模型推理。 核心是 Prompt Engineering 与协作框架。 推理侧不改变模型本身参数,而是通过设计引导策略,或多智能体协作流程。 让模型在现有能力基础上,更有序、更精准的输出逻辑链,避免直接跳步或答非所问。 核心是解决简单问题不用复杂流程,复杂问题不遗漏步骤的适配性问题。 一、核心技术路径与行业现状推理侧的技术进化,本质是从单一提示到动态协作的升级。 目前已形成三类主流路径,覆盖从简单到复杂的推理场景。 路径一,基础 Prompt Engineering,单轮提示,适配简单推理,核心逻辑。 通过指令模板或思维链示例,Few Shot COT,直接告诉模型该按步骤想,无需额外工具或多轮交互。 适用于步骤固定、无需外部信息的问题,如小学数学题、简单因果分析。 具体做法,对需要步骤的问题,在提示词中加入请分三步分析。 一、先明确题干条件。 二、计算关键数据。 三、推导结论。 对容易跳步的问题加入示例。 如例,已知 A 等于3,B 等于5,求 A 加 B 乘2。 第一步算 B 乘2等于10,第二步算 A 加10等于13。 你的解题步骤)。 代表模型,所有具备基础逻辑能力的大模型均支持。 如 GPT 3.5、Cloud 2、文心一言、通义千问,是目前最普及的推理方式,无需额外配置,仅靠提示词即可激活。 路径二,工具增强型推理,多轮调用,适配需外部信息的推理,核心逻辑,当推理需要实时数据、私有知识或精准计算时,如2024年某城市 GDP 同比增长多少?根据公司内部合同条款,某条款是否合法?让模型自主判断需要调用工具,并通过多轮交互完成推理。 先调用搜索 RAG 工具获取数据,再基于数据推导结论。 具体做法,一、模型先分析用户问题,判断是否需要外部信息。 如孔子是否活着,无需工具。 2024年某城市 GDP 需要工具。 二、若需要调用指定工具,如搜索引擎、向量数据库,获取数据后,再结合数据输出推理结论。 三、若工具返回数据不完整,模型可二次发起工具调用。 如搜索结果未包含某区数据,请补充查询该区域 GDP。 代表模型与产品,GPT 4加插件,如 Browse with Bing、Code Interpreter,可自主调用搜索计算工具,解决需要实时专业数据的推理问题。 阿里云通义千问工具链,支持调用企业内部 RAG 知识库、计算器适配,企业及推理场景如财务报销规则解读,产品参数对比,讯飞星火多轮工具调用,在数学计算中,可自动调用计算器,避免计算错误,再基于计算结果推导应用题答案。 路径三,多智能体协作推理,分工协作,适配超复杂推理,核心逻辑。 将复杂推理任务拆解为多个子任务,让不同角色模型分工完成。 如小模型负责拆解问题,检索模型负责找数据,大模型负责整合结论。 模拟人类团队协作,解决单模型无法覆盖全流程的问题。 如商业方案撰写、科研论文分析。 具体做法,一,任务拆解智能体,通常是轻量模型。 如 LLaMA 7B,将用户需求拆解为子任务。 如写一份某产品市场分析报告,拆解为一、竞品销量数据查询二、目标用户画像分析三、市场规模测算。 二、工具调用智能体。 根据子任务调用对应工具,如调用电商平台 API 查竞品销量,调用调研数据库查用户画像。 三、结论整合智能体。 大模型如 GPT 4、DeepMind CAR1,将子任务结果整合为完整逻辑链,输出最终答案。 代表案例,Meta 的 Toolformer,通过训练让模型学会调用工具的时机,并与其他模型协作完成推理。 国内,智谱清言多智能体支持自定义角色。 如分析师、数据源、编辑分工。 完成复杂任务,如分析师拆解市场任务、数据源查数据、编辑润色报告。 开源项目 AutoGPT 通过目标设定、子任务拆解、工具调用、结果反馈的闭环,自主完成长周期推理任务,如帮我规划一场欧洲旅行,包含签证流程、行程路线、预算计算。 二、推理侧的核心问题与瓶颈。问题一,何时用流程的判断依赖外部框架,模型自主决策弱。 目前多数场景需要人工预设规则,如所有涉及实时数据的问题,必须调用搜索,而非模型自主判断。 例如,若未预设规则,模型可能对2024年 GDP 直接输出旧数据,依赖于训练知识,而非主动调用工具。 问题二,多轮协作的容错性差,若某一环出错,如工具返回错误数据、子任务拆解遗漏,后续推理会连锁出错。 且模型难以自主修正,如搜索到错误的 GDP 数据,模型会基于错误数据推导结论,无法识别数据本身有误。 问题三,效率与成本矛盾,复杂协作流程,如多轮工具调用加多智能体分工,会导致响应速度变慢。 且大模型调用成本增加,不适用于实时性要求高的场景,如客服对话。 大模型深度推理能力现状全扫描,推理侧与训练侧 技术路径、现状及问题。 二、训练侧优化,让模型自带推理能力。 核心是从数据驱动到机制优化。 训练侧通过修改模型结构、优化训练目标或改进数据方式,让模型在预训练或微调阶段就掌握推理逻辑,无需依赖外部提示即可自主输出思维链。 目前行业仍处于从模仿推理 SFT 到自主推理算法优化的过渡阶段,尚未实现一劳永逸。 一、核心技术路径与行业现状训练侧的技术进化,是从靠数据为模板,到靠算法改机制的突破。 目前有四类关键路径。 路径一,监督微调 SFT。 让模型模仿推理步骤,当前最成熟。 核心逻辑,用带思维链的标注数据,如解题步骤加答案、分析逻辑加结论,微调模型,让模型学会按步骤输出。 本质是照猫画虎式的推理模仿。 具体做法,构建数据集,收集大量问题加详细推理过程加答案的数据如 数学题解题步骤,法律案例分析逻辑,微调训练。 用该数据集对预训练模型,如 LLaMA 2、GPT 3进行微调,让模型学习输入问题输出推理链输出答案的映射关系。 代表模型 Deepseek Math,用200万加数学题加解题步骤数据微调。 在中小学数学竞赛题中,可自主输出推理步骤,无需提示分步骤。 百度文心一言 V4,通过中文思维链数据集微调,在中文因果分析、逻辑推理任务中,无需额外提示即可输出分析过程。 Anthropic 的 claude 2,用长文本推理数据集,如小说情节分析、论文逻辑拆解、微调,支持自主输出长链推理。 局限性,依赖高质量标注数据,且泛化性差。 用数学题数据微调的模型,在法律推理任务中可能失效。 同时,模型仅学会模仿数据中的推理模板,遇到数据中没有的新场景,如跨领域问题,仍会跳步或出错,无法真正自主推理。 路径二,思维链预训练,CoT Pre-training,让模型在预训练阶段学推理,探索中核心逻辑。 在模型预训练阶段,而非微调阶段,就加入思维链类数据,如百科词条的逻辑解释、学术论文的推导过程。 让模型从学习 语言规律升级为学习逻辑规律,从源头掌握推理能力。 具体做法,扩展预训练数据。 在传统文本、新闻、书籍中加入带逻辑链的文本,如维基百科中的原理解释部分、学术论文的实验设计与推导章节。 调整预训练目标,除了预测下一个 token,额外加入预测推理链连贯性的目标,如判断步骤 A、步骤 B 是否符合逻辑。 代表模型,Google Palm 2,预训练阶段加入数学推导、代码逻辑、科学原理等思维链数据,在多语言推理任务中表现优于前代,华为盘古大模型推理版,预训练时融入工程计算、物理公式推导数据,在工业场景推理,如设备故障排查逻辑中,可自主输出分析链。 局限性,思维链领域训练数据稀缺,高质量逻辑链文本远少于普通文本,且训练成本极高,需重新训练千亿参数模型。 目前仅少数巨头 Google、华为有能力尝试。 路径三,强化学习,RALF, RAL with feedback. 让模型修正推理错误,辅助优化核心逻辑。 通过人类反馈或机器反馈,让模型在训练中知道自己错在哪,并修正推理逻辑,本质是从错误中学习。 具体做法一、生成推理结果,让模型输出某问题的推理链,如解题步骤。 二、反馈打分,由人类或评判模型对推理链打分,如步骤是否完整、逻辑是否正确。 三、强化训练,用打分结果优化模型,让模型更倾向于输出高分推理链,减少错误步骤。 代表模型 GPT 4,通过人类标注员对推理链打分的 R L H F 训练。 减少推理跳步和逻辑矛盾,如之前算错的数学题,训练后可修正步骤。 商汤日日新大模型用机器评判模型,基于高质量推理数据训练,自动对推理链打分,降低人工标注成本,提升训练效率。 局限性,反馈质量决定效果。 若人类标注员对推理逻辑的判断不统一,如法律案例的不同解读,模型会陷入混乱。 且 RL 训练容易过拟合反馈数据,遇到新问题仍会出错。 路径四,算法机制优化,从统计学预测到逻辑化推理,根本性突破方向,核心逻辑。 不依赖数据,而是通过修改 Transformer 架构或训练目标,解决模型靠概率预测 token,而非逻辑推导的本质问题,让模型具备自主探索推理路径的能力。 目前最关键的技术是多路径推理,Multi-Path Reasoning。 具体做法,以多路径推理为例。 一,多路径生成模型,对同一个问题,生成多个不同的推理路径。 如解数学题时,生成三种不同的解题思路。 二、路径评估,通过逻辑一致性判断或结果验证,筛选出正确的推理路径,如三种解题思路中哪种步骤无矛盾,结果正确。 三、强化学习,让模型更倾向于生成正确路径,逐步掌握自主选择合理推理方向的能力。 代表技术与模型,Google,verification,COVE。 模型先输出推理链,再回头检查每一步是否正确。 若发现矛盾,则修正路径。 在事实性推理,如某事件时间线中降低幻觉。 开源模型 MiniCPM 2B,通过多路径采样加逻辑打分机制,在小参数模型20亿参数中实现自主推理,无需依赖大模型。 微软 ResNet 在 Transformer 中加入推理状态跟踪模块,让模型记录已完成的推理步骤,避免重复或遗漏,适用于长链推理,如科研论文推导。 现状,这类技术仍处于实验室阶段,尚未大规模落地,多路径生成会增加训练和推理成本。 且路径评估的准确性依赖复杂的逻辑判断模型。 目前仅在数学计算、简单逻辑推理等确定性场景中有效,在模糊性问题,如商业决策分析中表现不佳。 二、训练侧的核心问题与瓶颈问题。 一、自主推理尚未实现,仍依赖数据模板或局部优化。 目前最成熟的 SFT 和 RLHF,本质仍是数据驱动,模型仅学会模仿已知推理模式,无法像人类一样面对全新问题时自主探索推理路径,如 AlphaGo Zero 那样仅靠规则自主学习。 问题二,Transformer 架构的概率本质与推理逻辑存在矛盾。 模型的核心是预测下一个 token 的概率,而非基于逻辑推导。 即使输出正确推理链,也可能是概率上最可能的序列,而非逻辑上最正确的序列,导致看似有逻辑,实则靠运气的幻觉。 问题三,训练成本与效果的平衡难题,根本性的机制优化,如多路径推理、思维链训练。 需要千亿级参数模型、加海量高质量数据、加超高计算资源,仅少数巨头可承担,中小公司难以跟进。 且效果提升与成本投入不成正比。 部分技术如多路径推理,成本增加10倍,效果仅提升5%。 大模型深度推理能力现状全扫描,推理侧与训练侧技术路径、现状及问题。 三、总结当前大模型推理的核心结论。 一、推理侧加训练侧混合式是主流,纯训练侧一劳永逸尚未实现。 目前行业普遍采用训练侧打底,SFT 加基础 RLHF,加推理侧优化,Prompt 加工具调用的模式。 训练让模型具备基础推理意识,推理侧通过引导和工具弥补泛化性不足,如 GPT 4,Deepseek CAR。 均采用此模式。 二、推理侧的核心突破点是自主决策,训练侧的核心突破点是机制优化。 未来推理侧需要解决模型自主判断是否用工具协作,无需人工预设规则。 训练侧需要解决 Transformer 概率本质与逻辑推理的矛盾,如通过多路径推理、逻辑模块嵌入等技术,才能实现模型自主、精准、高效推理。 三、短期看混合模式够用,长期看,纯训练侧突破是关键。 对于多数场景,如客服、办公、教育、训练侧 SFT 加推理侧工具调用,也能满足需求。 但对于高可靠性场景,如医疗诊断、法律判决、自动驾驶决策,必须依赖训练侧的根本性突破,让模型具备可解释、无幻觉、自主修正的推理能力。 这也是行业未来3~5年的核心方向。
英文翻译
Comprehensive Scan of the Current State of Deep Reasoning Capabilities in Large Models: Technical Paths, Current Status, and Issues on the Reasoning Side and Training Side The essence of deep reasoning in large models is to simulate the logical chains humans use to solve complex problems, such as mathematical calculations, causal analysis, and multi-step decision-making. The industry is currently making breakthroughs in two major directions: optimization on the reasoning side, which uses external guidance to enable models to utilize their abilities, and optimization on the training side, which endows models with inherent capabilities through technology. Each direction has its own paths, current status, and bottlenecks. Specifically: I. Reasoning Side Optimization: Activating Model Reasoning through External Guidance The core lies in Prompt Engineering and collaboration frameworks. The reasoning side does not change the model’s parameters. Instead, it designs guidance strategies or multi-agent collaboration processes to make the model output logical chains in a more orderly and accurate manner based on its existing capabilities, avoiding direct leaps or irrelevant answers. The key is to address the adaptability issue: for simple problems, complex processes should not be used; for complex problems, no steps should be missed. I. Core Technical Paths and Industry Current Status The technological evolution on the reasoning side is essentially an upgrade from single prompts to dynamic collaboration. Three mainstream paths have been formed, covering reasoning scenarios from simple to complex. Path 1: Basic Prompt Engineering – Single-turn prompts, suitable for simple reasoning. Core logic: Use instruction templates or chain-of-thought examples (Few-Shot COT) to directly tell the model to think step by step, without the need for additional tools or multi-turn interactions. Applicable to problems with fixed steps that do not require external information, such as elementary math problems or simple causal analysis. Specific approach: For problems requiring steps, add in the prompt: "Please analyze in three steps: 1. Clarify the given conditions. 2. Calculate key data. 3. Derive the conclusion." For problems prone to skipping steps, add examples (e.g., "Given A=3, B=5, find A+B×2. Step 1: B×2=10, Step 2: A+10=13. Your solution steps"). Representative models: All large models with basic logical capabilities, such as GPT-3.5, Claude 2, ERNIE Bot, Tongyi Qianwen. This is the most common reasoning method, requiring no additional configuration—just prompts to activate. Path 2: Tool-Augmented Reasoning – Multi-turn calls, suitable for reasoning requiring external information. Core logic: When reasoning requires real-time data, private knowledge, or precise calculations (e.g., "What was the GDP growth rate of a certain city in 2024?" or "Is a certain clause legal according to internal company contracts?"), the model autonomously determines the need to call tools and completes reasoning through multi-turn interactions. First, call a search/RAG tool to obtain data, then derive conclusions based on the data. Specific approach: 1. The model first analyzes the user’s question to determine if external information is needed (e.g., "Is Confucius alive?" – no tool needed; "GDP of a certain city in 2024" – tool needed). 2. If a tool is needed, call a specified tool (e.g., search engine, vector database), obtain data, and then output the reasoning conclusion combined with the data. 3. If the data returned is incomplete, the model can initiate a second tool call (e.g., "Search results did not include data for a certain district; please supplement the query for that region's GDP"). Representative models and products: GPT-4 + plugins (e.g., Browse with Bing, Code Interpreter) can autonomously call search and calculation tools to solve reasoning problems requiring real-time professional data. Alibaba Cloud’s Tongyi Qianwen tool chain supports calls to internal enterprise RAG knowledge bases and calculators, suitable for enterprise reasoning scenarios (e.g., interpreting financial reimbursement rules, comparing product parameters). iFlytek’s Spark multi-turn tool calling: in mathematical calculations, it can automatically call a calculator to avoid calculation errors, then derive answers based on the results. Path 3: Multi-Agent Collaborative Reasoning – Division of labor, suitable for ultra-complex reasoning. Core logic: Break down complex reasoning tasks into multiple subtasks, with different role models working on them. For example, a small model handles problem decomposition, a retrieval model handles data searching, and a large model integrates conclusions. This simulates human team collaboration to solve problems that a single model cannot cover end-to-end, such as business plan writing or scientific paper analysis. Specific approach: 1. Task decomposition agent (usually a lightweight model, e.g., LLaMA 7B): Decomposes user requirements into subtasks (e.g., writing a market analysis report for a product: breakdown into (a) competitor sales data query, (b) target user profile analysis, (c) market size estimation). 2. Tool-calling agent: Calls corresponding tools based on subtasks (e.g., calls e-commerce platform API to query competitor sales, calls survey database to query user profiles). 3. Conclusion integration agent (large model, e.g., GPT-4, DeepMind): Integrates subtask results into a complete logical chain and outputs the final answer. Representative cases: Meta’s Toolformer trains models to learn when to call tools and collaborate with other models. Domestically, Zhipu’s ChatGLM multi-agent supports custom roles (e.g., analyst, data source, editor) to complete complex tasks (e.g., analyst decomposes market task, data source queries data, editor polishes report). Open-source project AutoGPT completes long-cycle reasoning tasks autonomously through a closed loop of goal setting, subtask decomposition, tool calling, and result feedback (e.g., "Plan a European trip for me, including visa process, itinerary, and budget calculation"). II. Core Problems and Bottlenecks on the Reasoning Side Problem 1: Judging when to use processes relies on external frameworks; the model’s autonomous decision-making is weak. Currently, most scenarios require manually preset rules (e.g., all questions involving real-time data must call search), rather than the model autonomously deciding. For example, without preset rules, the model might output outdated data for 2024 GDP based on training knowledge instead of actively calling a tool. Problem 2: Multi-turn collaboration has poor fault tolerance. If one link fails (e.g., tool returns incorrect data, subtask decomposition misses something), subsequent reasoning will chain-react with errors, and the model struggles to self-correct. For example, if it searches for incorrect GDP data, it will derive conclusions based on that wrong data without recognizing the error. Problem 3: Efficiency versus cost contradiction. Complex collaboration processes (e.g., multiple tool calls plus multi-agent division) lead to slower response times and increased costs for large model calls, making them unsuitable for real-time scenarios like customer service conversations. Comprehensive Scan of the Current State of Deep Reasoning Capabilities in Large Models: Technical Paths, Current Status, and Issues on the Reasoning Side and Training Side II. Training Side Optimization: Endowing Models with Inherent Reasoning Capabilities Core: From data-driven to mechanism optimization. The training side modifies model structure, optimizes training objectives, or improves data methods so that the model masters reasoning logic during pre-training or fine-tuning, enabling it to autonomously output chain-of-thought without relying on external prompts. Currently, the industry is still in a transitional phase from imitative reasoning (SFT) to autonomous reasoning algorithm optimization, and a one-size-fits-all solution has not yet been achieved. I. Core Technical Paths and Industry Current Status The technological evolution on the training side is a breakthrough from using data as templates to changing mechanisms through algorithms. Currently, there are four key paths. Path 1: Supervised Fine-Tuning (SFT) – Letting the model imitate reasoning steps; currently the most mature. Core logic: Use annotated data with chain-of-thought (e.g., problem-solving steps + answers, analysis logic + conclusions) to fine-tune the model, teaching it to output step by step. Essentially, it is copying reasoning like drawing a tiger according to a cat. Specific approach: Build a dataset (collect a large number of problems with detailed reasoning processes and answers, such as math problem solution steps, legal case analysis logic) and fine-tune pre-trained models (e.g., LLaMA 2, GPT-3) to learn the mapping from input problem to output reasoning chain and answer. Representative models: DeepSeek-Math – fine-tuned with 2M+ math problems and solution steps; can autonomously output reasoning steps for primary/secondary school math contest problems without prompting “step by step.” Baidu’s ERNIE Bot V4 – fine-tuned with Chinese chain-of-thought datasets; can output analysis processes for Chinese causal analysis and logical reasoning tasks without additional prompts. Anthropic’s Claude 2 – fine-tuned with long-text reasoning datasets (e.g., novel plot analysis, paper logic decomposition); supports autonomous long-chain reasoning. Limitations: Depends on high-quality annotated data and has poor generalization. A model fine-tuned on math data may fail on legal reasoning tasks. Also, the model only imitates reasoning templates in the data; encountering new scenarios not in the data (e.g., cross-domain problems) can still lead to skipping steps or errors, lacking true autonomous reasoning. Path 2: Chain-of-Thought Pre-training (CoT Pre-training) – Having the model learn reasoning during the pre-training phase; exploratory. Core logic: During pre-training (not fine-tuning), incorporate chain-of-thought-like data (e.g., logical explanations in encyclopedia entries, derivation processes in academic papers) so that the model upgrades from learning language patterns to learning logical patterns, acquiring reasoning ability at the source. Specific approach: Expand pre-training data: add logic-chain texts (e.g., principle explanations in Wikipedia, experimental design and derivation sections in academic papers) to traditional text (news, books). Adjust pre-training objectives: besides predicting the next token, add objectives to predict the coherence of reasoning chains (e.g., judging whether step A and step B are logically consistent). Representative models: Google PaLM 2 – included math derivation, code logic, and scientific principle chain-of-thought data in pre-training; outperformed previous generations in multilingual reasoning tasks. Huawei Pangu large model (reasoning version) – incorporated engineering calculations and physics formula derivation data in pre-training; can autonomously output analysis chains in industrial scenario reasoning (e.g., equipment fault diagnosis logic). Limitations: Scarcity of chain-of-thought domain training data (high-quality logic chain texts are far fewer than ordinary texts) and extremely high training costs (requires retraining models with hundreds of billions of parameters). Currently, only a few giants (Google, Huawei) can attempt this. Path 3: Reinforcement Learning with Human/Machine Feedback (RLHF/RALF) – Letting the model correct reasoning errors; assists optimization. Core logic: Through human or machine feedback, the model learns where it went wrong during training and corrects its reasoning logic. Essentially learning from mistakes. Specific approach: 1. Generate reasoning results: model outputs a reasoning chain for a problem (e.g., solution steps). 2. Feedback scoring: human or judge model scores the reasoning chain (e.g., step completeness, logical correctness). 3. Reinforcement training: Optimize the model using scores to favor high-scoring reasoning chains and reduce error steps. Representative models: GPT-4 – trained with RLHF where human annotators scored reasoning chains; reduces reasoning leaps and logical contradictions (e.g., previously incorrect math problems can be corrected after training). SenseTime’s SenseNova – uses a machine judge model (trained on high-quality reasoning data) to automatically score reasoning chains, reducing human annotation costs and improving training efficiency. Limitations: The quality of feedback determines effectiveness. If human annotators have inconsistent judgments on reasoning logic (e.g., different interpretations of legal cases), the model becomes confused. Also, RL training tends to overfit feedback data, leading to errors on new problems. Path 4: Algorithm Mechanism Optimization – From statistical prediction to logical reasoning; fundamental breakthrough direction. Core logic: Instead of relying on data, modify the Transformer architecture or training objectives to solve the fundamental problem that models predict tokens by probability rather than logical deduction, enabling the model to autonomously explore reasoning paths. The most critical current technology is Multi-Path Reasoning. Specific approach (taking multi-path reasoning as an example): 1. Multi-path generation: For the same problem, generate multiple different reasoning paths (e.g., three different solution ideas for a math problem). 2. Path evaluation: Use logical consistency checks or result verification to select the correct reasoning path (e.g., which of the three solution paths has no contradictions and yields the correct result). 3. Reinforcement learning: Train the model to favor generating correct paths, gradually mastering the ability to autonomously choose reasonable reasoning directions. Representative technologies and models: - Google’s COVE (Chain-of-Thought Verification): Model first outputs reasoning chain, then goes back to check each step; if a contradiction is found, it corrects the path. Reduces hallucinations in factual reasoning (e.g., timeline of events). - Open-source model MiniCPM-2B: Achieves autonomous reasoning in a small parameter model (2B parameters) through multi-path sampling plus logic scoring, without relying on large models. - Microsoft’s ResNet in Transformers: Adds reasoning state tracking modules to record completed reasoning steps, avoiding repetition or omission, suitable for long-chain reasoning (e.g., scientific paper derivation). Current status: These technologies are still in the laboratory stage and have not been widely deployed. Multi-path generation increases training and inference costs. The accuracy of path evaluation relies on complex logic judgment models. Currently effective only in deterministic scenarios like mathematical calculations and simple logical reasoning; performs poorly in fuzzy problems like business decision analysis. II. Core Problems and Bottlenecks on the Training Side Problem 1: Autonomous reasoning has not yet been achieved; still relies on data templates or local optimization. The most mature SFT and RLHF are essentially data-driven; models only imitate known reasoning patterns and cannot autonomously explore reasoning paths for entirely new problems like humans (e.g., AlphaGo Zero learning solely from rules). Problem 2: There is a contradiction between the probabilistic nature of the Transformer architecture and reasoning logic. The core of the model is predicting the probability of the next token, not logical deduction. Even if the model outputs a correct reasoning chain, it may be the most probable sequence rather than the logically most correct one, leading to seemingly logical but actually lucky hallucinations. Problem 3: Difficulty balancing training cost and effectiveness. Fundamental mechanism optimizations (e.g., multi-path reasoning, chain-of-thought training) require models with hundreds of billions of parameters, massive high-quality data, and ultra-high computing resources, affordable only by a few giants. Small and medium companies cannot keep up. Moreover, the improvement in effectiveness is not proportional to cost input: some technologies (e.g., multi-path reasoning) increase cost by 10 times while improving effectiveness by only 5%. Comprehensive Scan of the Current State of Deep Reasoning Capabilities in Large Models: Technical Paths, Current Status, and Issues on the Reasoning Side and Training Side III. Summary: Key Conclusions on Current Large Model Reasoning 1. The hybrid model (reasoning side + training side) is mainstream; a pure training-side one-shot solution has not yet been achieved. The current industry generally adopts a pattern: training side as the foundation (SFT + basic RLHF) plus reasoning side optimization (prompts + tool calling). Training gives the model basic reasoning awareness; the reasoning side compensates for insufficient generalization through guidance and tools. For example, GPT-4 and DeepSeek R1 both use this pattern. 2. The core breakthrough point on the reasoning side is autonomous decision-making; on the training side, it is mechanism optimization. In the future, the reasoning side needs to solve the problem of the model autonomously deciding whether to use tools or collaborate, without manual preset rules. The training side needs to resolve the contradiction between the probabilistic nature of Transformers and logical reasoning, e.g., through multi-path reasoning, embedding logic modules, to achieve autonomous, accurate, and efficient reasoning. 3. In the short term, the hybrid model is sufficient; in the long term, breakthroughs on the pure training side are key. For most scenarios (customer service, office work, education), training-side SFT plus reasoning-side tool calling can meet requirements. However, for high-reliability scenarios (medical diagnosis, legal judgments, autonomous driving decisions), fundamental breakthroughs on the training side are necessary to equip models with explainable, hallucination-free, and self-correcting reasoning abilities. This is the core direction for the industry over the next 3–5 years.
back to top