LFM2.5-8B tool use · V2

Model observation

Sigma5C observed LFM2.5-8B again after changes to its surrounding agent software. Some affected tasks took less time, while repeated tool use, incorrect answers and unsupported tool-use claims remained in the record.

Tool tasks: 12
Task classes: 5
Agent interfaces: 6

What was observed

Changes to file access and Python argument handling coincided with shorter recorded times on affected tasks. A completed status still did not establish a correct answer.

Repeated tool use and unsupported claims of tool execution remained in the record. The revised guards did not remove every observed failure pattern.

What we observed

Area	Recorded scope
Model	LFM2.5-8B by Liquid AI
Tool tasks	Search, browsing, file access and code execution
Agent interfaces	Coding, editing, execution and review prompts
Measures	Recorded process status and elapsed time

Reading the records

These records can support examination of the gap between process completion, observed tool execution and the requested outcome. The surrounding agent software is part of the observation.

Declared limitations

These are bounded observations of one model configuration in each version. Process status is not a task-satisfaction score, and tool access differed between agent interfaces. Timing includes the surrounding software and any retries; it does not isolate model inference time.

Tool-task status

Recorded tool-task outcomes

Step limit: 1
Completed: 11

Counts cover all 12 tool-task records. Completed is a process label; it does not establish task satisfaction. Each bar is scaled to the full task count.

Time by task class

Mean elapsed time by task class

Web search: 280.3 s
Search and browse: 76.2 s
File tools: 170.3 s
Python: 124.5 s
Multiple tools: 106.6 s

Arithmetic mean of recorded elapsed seconds in each class, including incomplete tasks. Bars share a zero baseline and scale to the largest mean on this page.

Companies

Liquid AI