SFT plus GRPO training combined with renderer-in-the-loop inference improves LLM Manim code generation, with Qwen 3 Coder 30B reaching 94% render success and 85.7% visual similarity, beating GPT-4.1 by 3 points.
Direct Preference Optimization: Your Language Model is Secretly a Reward Model
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Training and Agentic Inference Strategies for LLM-based Manim Animation Generation
SFT plus GRPO training combined with renderer-in-the-loop inference improves LLM Manim code generation, with Qwen 3 Coder 30B reaching 94% render success and 85.7% visual similarity, beating GPT-4.1 by 3 points.