Qwen3-VL Syntax Constraint Causes Infinite Repetition
mattjcoles · reddit · 2026-07-14
The author discovered that when using GBNF grammar-constrained decoding, Qwen3-VL-30B-A3B gets stuck in a repetition loop, endlessly outputting the same valid JSON entry until the context window is exhausted.
Multiple tests were conducted:
- Tweaking quantization, temperature, flash attention, context length, and repetition penalty failed to fix the root cause.
- Cranking up the repetition penalty broke the loop, but resulted in grammatically valid JSON with zero entries.
- The same issue occurred on the smaller 8B model, failing on 15/37 pages.
- gemma-4-31b showed similar failures with long string fields.
Conclusion: Simply checking for JSON validity is insufficient; extraction tasks must also verify the actual count and content of the generated entries.
Related event: Qwen3-VL Gets Stuck in Repetition Loops Under GBNF Constraints(2 posts)→
More from Models
- Sam Altman is headed to Washington to brief Congress on OpenAI’s GPT-6 line — inductionheads · 2026-07-22
- Benchmark chart pits GPT-5.6 Luna, Grok 4.5 and Gemini 3.6 Flash on price and scores — iruletheworldmo · 2026-07-22
- Claim says Kimi was distilled from Fable, sparking a model-attribution jab — cephaloform · 2026-07-22
- Gemini 3.6 Flash is now available in Antigravity and chat — MartianOnJupiter · 2026-07-22
- Gary Marcus says LLM math skills are like knowing only a car’s engine size — GaryMarcus · 2026-07-22
- OpenAI’s Codex + GPT-5.6 Sol hits 99% recall in Project APE verification tests — soumitrashukla9 · 2026-07-22