GLM-4.6 model fails to use web search tools after processing complex mathematical problems, and subsequently loses all tool use capability even when explicitly prompted.

🔧 Reproduction Steps

 Initiate a complex mathematical problem-solving task
 Follow up with a simple web search request (e.g., "what is today's date?")
 Observe that the model disregards tool use and provides knowledge-based responses
 Explicitly prompt for web search tool usage
 Observe complete tool use failure and simulation of tool calls instead of actual execution

🎯 Expected Behavior

 Model should execute web search tool when requested
 Tool use should remain available after complex tasks
 Model should not simulate tool calls but actually execute them

🌍 Environment

 Model: GLM-4.6
 Platform: chat.z.ai
 Browser: Floorp
 Date: 2025-10-01

📝 Additional Context

 Issue occurs consistently after complex problem-solving tasks
 Model appears to enter a broken state where tool use is completely disabled
 Simulation of tool calls instead of actual execution suggests a systemic failure
 This affects core functionality and reliability of the model
 Chat URL: https://chat.z.ai/s/5c08bd6d-ecf7-426f-a909-350f4c1c6af2

This issue might be a duplicate of existing issues. Please check:

GLM Coding Plan - Frequent Errors #2898: GLM Coding Plan - Frequent Errors - Reports regular "Invalid API Parameter" and timeout errors with GLM models, specifically mentions GLM 4.6 having more pronounced errors than 4.5
Timeout then tool errors with Devstral-Medium #2865: Timeout then tool errors with Devstral-Medium - Describes tool execution failures followed by tool errors, including "AI_APICallError: Unexpected role 'user' after role 'tool'" which indicates broken tool state similar to your simulation issue
AI_APICallError: tool_use blocks found without corresponding tool_result blocks #2720: AI_APICallError: tool_use blocks found without corresponding tool_result blocks - Reports tool use blocks without proper tool results in longer conversations, which aligns with your "broken state where tool use is completely disabled"
Refused to use tool right after a rejection. #1807: Refused to use tool right after a rejection - Describes models refusing to use tools after certain interactions, similar to your "loses all tool use capability" issue
Tool use with Ollama models. #1068: Tool use with Ollama models - Reports tool execution being aborted with various models, with similar symptoms of tool calls failing to execute properly

The common pattern across these issues involves models losing tool functionality or entering broken states after certain interactions, which matches your description of GLM-4.6 losing tool capability after complex mathematical problems.

Feel free to ignore if none of these address your specific case.

Navigation Menu

Repository navigation

GLM-4.6 model fails to use web search tools after processing complex mathematical problems, and subsequently loses all tool use capability even when explicitly prompted. #2903

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

GLM-4.6 model fails to use web search tools after processing complex mathematical problems, and subsequently loses all tool use capability even when explicitly prompted. #2903

Description

Activity

github-actions commented on Oct 1, 2025

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions