# Ubicloud says decoding change eliminates malformed tool calls in benchmark

_Published Monday, October 26, 2026 at 10:07 AM EDT · AI, Infrastructure · Latest · Tier 2 — Notable_

![Ubicloud says decoding change eliminates malformed tool calls in benchmark — Primary](https://cdn.prod.website-files.com/64f9d9b4e737e7b37d4e39a4/6ace25a7011f2d0bca7a06ea_LLM-inference.png)

Ubicloud says it deployed a vLLM configuration change after testing showed tool-call schema accuracy improved from 92% to 100% on the K2VV ToolCall benchmark. The company serves Kimi and DeepSeek models and verified the result across all model versions it serves.

The change enables constrained decoding, which blocks tokens that would violate a tool's required structure during generation. Ubicloud traced a previous accuracy decline to vLLM disabling the feature by default for compatibility across models. The benchmark measures whether calls follow the required schema, rather than whether their contents solve a user's task.

## Sources

- [ubicloud.com](https://www.ubicloud.com/blog/fixing-tool-call-accuracy-in-vllm)

---
Canonical: https://techandbusiness.org/newswire/1ujNY1pRFzgwFdcNhMHOTF
Published: 2026-10-26T14:07:56.178Z
Story chronology: 2026-10-26T12:59:48.000Z
Retrieved: 2026-10-26T16:58:17.761Z
Publisher: Tech & Business (techandbusiness.org)