Diffusion large language models (dLLMs) re-encode the entire prefix at every denoising step, causing recomputation that scales quadratically with context length and becomes prohibitive for long-context scenarios. We propose Prefilling-dLLM, a training-free prefill-decode disaggregation framework for dLLMs that partitions the prefix into N chunks, caches their KV representations once, and selects the top-K most relevant chunks with intra-chunk token sparsity for decoding, showing that sparse prefilling can outperform dense attention while reducing per-step complexity from quadratic in the full sequence length to quadratic only in the decode length. On LongBench and InfiniteBench, Prefilling-dLLM achieves state-of-the-art quality among dLLM acceleration methods, and an attention kernel that parallelizes decoding over the non-contiguously cached chunk KV yields 9.1–28.0× speedup at 8K–32K contexts. We further show that beginning-of-sequence tokens prepended to each chunk act as periodic attention anchors that eliminate the lost-in-the-middle phenomenon. Our code is available at https://github.com/menik1126/Prefilling-dLLM.

Diffusion language models (LMs) enable parallel decoding, but still lag behind autoregressive (AR) LMs in accuracy, and behind multi-token prediction (MTP) methods in efficiency. This raises three questions: should diffusion LMs compete with AR LMs or be harmonized with them; can diffusion provide a stronger acceleration mechanism than MTP; and does diffusion decoding have sufficient long-term potential to justify further exploration? We study these questions through Nemotron-Labs-Diffusion, a tri-mode LM that unifies three decoding modes—AR, diffusion, and self-speculation (diffusion drafts, AR verifies)—within a single model to sustain high throughput across all concurrency levels. Our study shows that (1) AR and diffusion objectives are mutually beneficial rather than competing, (2) diffusion-based self-speculation achieves higher acceptance lengths and better system vs. per-user throughput trade-offs than MTP baselines, and (3) diffusion decoding shows strong potential for parallel decoding under an optimal sampler in a speed-of-light analysis. Based on these insights, we develop the Nemotron-Labs-Diffusion family at 3B, 8B, and 14B scales. Our models outperform state-of-the-art open-source AR and diffusion LMs in both accuracy and speed, e.g., achieving 6x tokens per forward over Qwen3-8B with comparable benchmark accuracy, translating to 4x throughput on SPEED-Bench measured with SGLang on an NVIDIA GB200 GPU.

Multi-modal Large Language Models have shown remarkable progress in visual contexts, yet their ability to convert visual figures into executable code remains underexplored. To address this, we introduce Plot2Code, a comprehensive benchmark designed to assess MLLMs’ visual coding capabilities. Plot2Code includes 132 high-quality matplotlib plots across six plot types, as well as an additional 150 and 86 plots from Python’s and R’s plotly libraries respectively, totaling 368 plots. Each plot is paired with its source code and a descriptive instruction generated by GPT-4, enabling thorough evaluation across diverse inputs. Furthermore, we propose three automatic evaluation metrics—code pass rate, text-match ratio, and GPT-4V rating judgement—to assess the quality of generated code and rendered images. Notably, the GPT-4V rating demonstrates strong reliability, as it correlates well with human evaluations, particularly for datasets of a certain size. Cross-validation across MLLMs (GPT-4V, Gemini-1.5-Pro, and Claude-3-Opus) also shows high consistency in ratings, which likely stems from the fact that ratings are based on rendered images rather than direct MLLM outputs, indicating minimal bias for this metric. Our evaluation of 14 MLLMs, including both proprietary and open-source models, highlights significant challenges in visual coding, particularly for text-dense plots, where MLLMs heavily rely on textual instructions. We believe these findings will advance future development of MLLMs.

Humans generally acquire new skills without compromising the old; however, the opposite holds for Large Language Models (LLMs), e.g., from LLaMA to CodeLLaMA. To this end, we propose a new post-pretraining method for LLMs with an expansion of Transformer blocks. We tune the expanded blocks using only new corpus, efficiently and effectively improving the model’s knowledge while mitigating forgetting. In this paper, we experiment on the corpus of code and math, yielding LLaMA Pro-8.3B, a versatile foundation model initialized from LLaMA2-7B, excelling in general tasks, programming, and mathematics. LLaMA Pro and its instruction-following counterpart (LLaMA Pro - Instruct) achieve advanced performance among various benchmarks, demonstrating superiority over existing open models in the LLaMA family and the immense potential of reasoning and addressing diverse tasks as an intelligent agent. Our findings provide valuable insights into integrating natural and programming languages, laying a solid foundation for developing advanced language agents that operate effectively in various environments.