AI 模型和平台
iMerit 发布缩放专家主导的AI培训与评价的学者平台
在 Google 上的首选来源中添加 Unite.AI
模型可以产生令人信服的医学解释或对数学问题的流利回答，并且仍然犯一个只有专家才会认出的错误。 这种对专家判断的需要坐落在iMerit的"iMerit学者"平台的发布背后，这个系统旨在组织那些训练，评价并挑战高级AI模型的人。
EXL自有公司宣布"发布"于10月26日将招生，上岗，专家配对，项目任务，业务约定和支付等纳入统一平台。 根据公告，其现有学者方案包括118个国家和106个语言的30 000多名受邀和受审查的专业人员。
随着需求的增长，该平台旨在管理数十万名专家。 这是关于预定业务规模的说明，不同于社区目前报告的规模。 当务之急是使专门的人力投入更容易在整个模型开发和部署生命周期中进行协调。
建设示范改进背后的劳动力
学者社区跨越医学，数学，农业，语言学，自主系统等专业领域。 这些专家为监督微调，模型评价，红队和从人类反馈中强化学习做出了贡献，通常称为RLHF.
这些活动需要不同的判断。 微调可以要求一个模型应生成的反应的精心构建的例子。 评价询问某项产出是否符合既定标准。 红队特意探测出一个故障系统，而人类的反馈可以提供偏好和纠正，为后续培训提供参考。
因此，让合适的人胜任这项任务，与扩大人才库同样重要。 熟悉狭义专业的临床医生可能会发现一个知识丰富的审查者错过的间接错误。 语言学家可能识别出一种在文字翻译中消失的模糊。
新平台将项目与领域专业化和认知技能相匹配，同时提供行政基础设施，使这些专家保持参与。 iMerit将目标描述为与学习工作并参与反馈周期的专家的持续关系。
称 Radha Basu是EXL执行副总裁兼iMerit负责人。
Ango Hub 用培训数据连接专家判断
学者平台下的操作技术是Ango Hub,iMerit的数据注释和模型评价平台。 其作用是将专家工作转变为可进入培训或评价管道的有组织、可审查的文物。
Ango Hub的现有产品文档描述了可配置的工作流程，资产上的多个注释器，基准测试任务和审查器控制以接受，拒绝或纠正标签。 项目准则和问题系统有助于审查人员解决问题，而基于角色的接入控制则制约着谁可以处理特定数据。
该平台支持多种模式，包括图像，视频，音频，文本，PDF和医疗DICOM数据。 这一广度之所以相关，是因为模型开发越来越多地将语言与视觉、临床或物理世界信息结合起来。 接口和审查程序需要符合专家正在审查的材料。
两个平台的区别是有用的：学者协调专家队伍及其与项目的关系，而Ango Hub则提供工具和工作流程，通过这些工具和工作流程，专家的贡献成为可用AI数据。
从纠正答案到评价全部 Agent 工作流程
iMerit的"深理性实验室"新增更详细地介绍了这一过程。 公司介绍专家设计多步骤情景，审查模型产出，分出出出错，提供更正解释。 这些结构化的例子可以帮助开发者确定答案出错之处并创建有针对性的培训材料。
这种推理痕迹是专门制作和审查的文物。 不应假定它们揭示每个产生模型答案的内部计算。 其实际价值在于使评价路径及其更正足够明确，以便检查和再利用。
该公司的agent评价服务超越了对规划步骤，工具呼叫，中间状态和结果的最终回应。 报告说明了任务完成、工具选择、定位和政策界限等问题的结构化主题和专家判断。
当大赦国际能够采取行动时，这一点变得很重要。 agent可能会在使用错误的工具或者使基础任务不完全的同时产生被抛光的总结。 审查动作的顺序，使专家能够识别出这些失败，并将结构化的反馈还给开发者。
付款和项目连续性成为基础设施的一部分
发布还整合了"特罗莱"用于全球支付。 iMerit 称整合支持其国际专家的安全支付和合规工作流程。
付款业务可能似乎从模式业绩中去除，但影响专家能否一致参与。 当一个项目依赖许多国家的专家时，征聘、入职、任务准入和报酬都成为业务上的依赖。
带来平台与匹配和接触同时发挥作用，而不是让每个项目小组单独组装。 对客户来说，预期的好处是员工队伍和审查过程更加一致；对专家来说，从参加一个项目到完成工作和领取报酬的道路更加明确。
称 Pavel Kopach, M. D. 贝斯以色列Deaconess医疗中心的病理学专家，以及医疗AI的iMerit学者。
EXL 带来企业部署背景
发布号随EXL完成iMerit收购，宣布 August 3. EXL将结合企业数据、业务背景和AI执行能力，结合模式培训和评价专门知识。
这为学者创造了一条更广泛的市场路线。 Basu 称 iMerit正在从前沿模式发展到银行，保健和保险工作流程，其中域名定制，信任和安全影响AI能否在生产中成功运行。
新平台本身并不确定一个模式在每个专门环境下都是可靠的。 它涉及一个更具体的先决条件：组织必要的人员、工具和反馈，以根据这种环境的要求测试一个系统。 随着AI深入到专业工作流程中去，找到合适的专家，捕捉有用的修正并将知识带回发展的能力成为了技术堆栈中越来越重要的一环。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
AI Models & Platforms
iMerit Launches Scholars Platform to Scale Expert-Led AI Training and Evaluation
Add Unite.AI to your preferred sources on Google
A model can produce a convincing medical explanation or a fluent answer to a mathematics problem and still make a mistake that only a specialist will recognize. That need for expert judgment sits behind iMerit’s launch of the iMerit Scholars platform, a system designed to organize the people who train, evaluate and challenge advanced AI models.
The EXL-owned company announced the launch on October 26, bringing recruitment, onboarding, expert matching, project assignments, engagement and payments into a unified platform. Its existing Scholars program includes more than 30,000 invited and vetted professionals across 118 countries and 106 languages, according to the announcement.
The platform is designed to manage hundreds of thousands of experts as demand grows. That is a statement about intended operating scale, distinct from the community’s current reported size. The immediate proposition is to make specialized human input easier to coordinate throughout the model development and deployment lifecycle.
Building the Workforce Behind Model Improvement
The Scholars community spans medicine, mathematics, agriculture, linguistics, autonomous systems and other specialized fields. These experts contribute to supervised fine-tuning, model evaluation, red-teaming and reinforcement learning from human feedback, commonly called RLHF.
Those activities need different kinds of judgment. Fine-tuning can require a carefully constructed example of the response a model should produce. Evaluation asks whether an output meets defined criteria. Red-teaming deliberately probes a system for failures, while human feedback can provide preferences and corrections that inform subsequent training.
Matching the right person to the task therefore matters as much as expanding the pool. A clinician familiar with a narrow specialty may spot a consequential error that a broadly knowledgeable reviewer misses. A linguist may recognize an ambiguity that disappears in a literal translation.
The new platform matches projects to domain specialization and cognitive skills, while providing the administrative infrastructure to keep those experts engaged. iMerit describes the goal as a continuing relationship with specialists who learn the work and participate in feedback cycles.
“We need people who are knowledgeable enough not just to teach models, but to prove them wrong and correct them,” said Radha Basu, executive vice president and head of iMerit at EXL.
Ango Hub Connects Expert Judgment With Training Data
The operational technology underneath the Scholars platform is Ango Hub, iMerit’s data annotation and model evaluation platform. Its role is to turn expert work into organized, reviewable artifacts that can enter a training or evaluation pipeline.
Ango Hub’s existing product documentation describes configurable workflows, multiple annotators on an asset, benchmark tasks and reviewer controls to accept, reject or correct labels. Project guidelines and an issue system help reviewers resolve questions, while role-based access controls govern who can work with particular data.
The platform supports several modalities, including images, video, audio, text, PDFs and medical DICOM data. That breadth is relevant because model development increasingly combines language with visual, clinical or physical-world information. The interface and review process need to fit the material an expert is examining.
The distinction between the two platforms is useful: Scholars coordinates the expert workforce and its relationship with projects, while Ango Hub provides the tools and workflows through which expert contributions become usable AI data.
From Correcting Answers to Evaluating Entire Agent Workflows
iMerit’s Deep Reasoning Lab adds a more detailed view of that process. The company describes experts designing multi-step scenarios, reviewing model outputs, categorizing errors and supplying corrections with explanations. These structured examples can help developers identify where an answer goes wrong and create targeted training material.
Such reasoning traces are curated artifacts created and reviewed for a task. They should not be assumed to reveal every internal computation that produced a model’s answer. Their practical value is in making an evaluation path and its corrections explicit enough to inspect and reuse.
The company’s agent evaluation service extends review beyond a final response to planning steps, tool calls, intermediate states and outcomes. It describes structured rubrics and expert adjudication for issues such as task completion, tool selection, grounding and policy boundaries.
That becomes important when AI can take action. An agent may produce a polished summary while using the wrong tool or leaving the underlying task incomplete. Reviewing the sequence of actions gives specialists a way to identify those failures and return structured feedback to developers.
Payments and Project Continuity Become Part of the Infrastructure
The launch also integrates Trolley for global payouts. iMerit says the integration supports secure payments and compliance workflows for its international expert community.
Payment operations may appear removed from model performance, but they affect whether a specialist can participate consistently. Recruitment, onboarding, task access and compensation all become operational dependencies when a project relies on experts across many countries.
The platform brings those functions alongside matching and engagement rather than leaving each project team to assemble them separately. For customers, the intended benefit is a more consistent workforce and review process; for experts, it is a clearer path from joining a project to completing work and receiving payment.
“You’re not just labeling data, you’re helping shape how these systems reason and behave in your field,” said Pavel Kopach, M.D., a pathology specialist at Beth Israel Deaconess Medical Center and an iMerit Scholar in healthcare AI, in the announcement.
EXL Brings the Enterprise Deployment Context
The launch follows EXL’s completion of its iMerit acquisition, announced August 3. EXL positioned the combination around joining model training and evaluation expertise with its enterprise data, operational context and AI implementation capabilities.
That creates a broader route to market for Scholars. Basu said iMerit is expanding from frontier-model development into banking, healthcare and insurance workflows, where domain customization, trust and safety affect whether AI can operate successfully in production.
The new platform does not itself establish that a model is reliable in every specialized setting. It addresses a more concrete prerequisite: organizing the people, tools and feedback needed to test a system against the demands of that setting. As AI moves deeper into professional workflows, the ability to find the right expert, capture a useful correction and bring that knowledge back into development becomes an increasingly important part of the technology stack.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.