ARCH Launches Custom Evaluation Consultations for AI Agent Benchmark 'LLMSnare'

ARCH Inc. (Headquarters: Chiyoda-ku, Tokyo; Chairman & CEO: Yonglong Wei) has begun accepting consultations for enterprise-level custom evaluations using 'LLMSnare,' a benchmark tool designed to evaluate Large Language Models (LLMs) used in AI Agents across different operational cases.

LLMSnare is a behavior-based benchmark that measures whether LLMs, when functioning as Agents, read necessary context, use tools appropriately, and recover from incorrect instructions or missing information. In addition to the public version of LLMSnare and LLMSnare Arena, ARCH provides support for designing custom cases aligned with a company’s specific business workflows, access permissions, tool integrations, and failure conditions.

Background
The number of models available for AI Agents is growing. Companies are increasingly combining commercial LLMs, open-weight models, various cloud services, OpenAI-compatible endpoints, and self-managed models.

However, for companies deploying AI Agents, determining which model is 'smartest' is not enough. Operations such as inquiry classification, pre-publication review, internal document verification, code correction, data updates, and periodic report generation each require different behaviors.

To use AI Agents effectively, companies must move beyond general benchmark rankings and verify how a model behaves under conditions similar to their own operational cases. ARCH believes that models used for AI Agents should be evaluated based on behavior in specific business cases, not general rankings.

What is LLMSnare?
LLMSnare is a benchmark tool that measures the behavior of an LLM when acting as an Agent.

While many benchmarks measure only final answers or products, LLMSnare additionally confirms whether the model read necessary information before acting, followed existing helpers and rules, recovered to the correct context from incorrect instructions, and maintained stable behavior when repeating cases.

LLMSnare evaluates Agent behavior primarily based on:
- Whether it reads the necessary context before execution.
- Whether it verifies required files, rules, and existing helpers before writing or executing.
- Whether the order and frequency of tool calling are appropriate.
- Whether it can recover from incorrect paths, ambiguous instructions, or missing information.
- Whether the final output adheres to case-specific regulations.
- Whether its behavior remains stable when repeating the same case.

The public version of LLMSnare offers CLI access, a public Arena, and documentation for creating custom cases.

Why Custom Enterprise Cases are Necessary
While public benchmarks are useful for gauging general model trends, companies need to know if a model is applicable under their specific operational conditions. For example, in pre-publication reviews, natural language quality is secondary to referencing correct materials, adhering to forbidden terms, legal compliance, and passing data to the correct approver. These cannot be judged by general rankings.