Qwen3.5 `</think>` Bug Triggers Endless Loops

// 116d agoNEWS

Qwen3.5 `</think>` Bug Triggers Endless Loops

A SillyTavern user reports Qwen3.5 27B Q3_XS suddenly emitting empty `<think></think>` blocks and stray `</think>` tags, then spiraling into repetitive outputs until context runs out. Similar reports point to a template or reasoning-parser mismatch in the local stack, not just a sampling issue.

// ANALYSIS

This looks more like a chat-template regression than a model “going bad.”

–Qwen3.5 has a known edge case where disabling thinking can still leave a dangling `</think>` tag in the output
–If the backend strips or rewrites those tags between turns, the prompt can be reprocessed and the model starts echoing itself
–Old GGUF conversions or custom Jinja templates are prime suspects when tag handling drifts
–The fact that penalties, DRY, and prompt instructions did nothing points away from sampling and toward serialization/parsing
–For local RP setups, reasoning-tag hygiene is now part of the runtime contract, not just cosmetic output formatting

// TAGS

qwen3-5llmreasoningprompt-engineeringself-hosted

DISCOVERED

116d ago

2026-04-04

PUBLISHED

116d ago

2026-04-04

RELEVANCE

7/ 10

AUTHOR

VerdoneMangiasassi

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

OPEN SOURCE26m ago

Cloudflare open-sources pvcli privacy proxy CLI

Cloudflare has open-sourced pvcli, a command-line utility that collapses multi-party privacy proxy flows—such as Oblivious HTTP and MASQUE—into a curl-like interface. By exposing binary HTTP framing, HPKE encryption, and intermediate trace logs, pvcli simplifies diagnosing network issues across relays, gateways, and origins.

NEWS3h ago

Tencent Cloud Developer Breaks Down Graph Engineering

Tencent Cloud shared an educational breakdown by developer Lukiexing examining Graph Engineering in AI agent architectures. As AI systems shift from single loops to graph-based structures, Graph Engineering addresses key challenges in orchestrating reliable multi-agent workflows.

UPDATE3h ago

Cursor adds local Bugbot and Security Review slash commands

Cursor developers can now run automated code quality and security audits locally on branch or uncommitted changes using in-editor review slash commands. Running Bugbot and Security Review locally helps developers identify logic flaws and security risks before pushing code to CI.