Back to feed
News Story
机器之心
1 sources

Ant Group Releases World's First Large-Scale Agentic Diffusion Model LLaDA2.2

Ant Group has released and open-sourced LLaDA2.2, the world's first large-scale agentic diffusion model. By introducing four atomic operations (keep, substitute, delete, insert), it overcomes the limitation of traditional diffusion language models in handling sequence length changes, enabling agentic tasks. This breakthrough marks a new era for diffusion language models and provides a new foundation for AI agent applications.

SynthePulse Insight · AI deep reading

LLaDA2.2: Diffusion Language Model Achieves True Agent Capabilities for the First Time, with Breakthroughs in Efficiency and Performance

Version 1 · 1 source

Ant Group has released and open-sourced the world's first large-scale Agentic diffusion model, LLaDA2.2. By leveraging Levenshtein editing and reinforcement learning, it addresses the structural deficiencies of diffusion models in long-horizon tasks, matching top autoregressive models on agent benchmarks while achieving 1.64x throughput.

  • LLaDA2.2 is the world's first large-scale Agentic diffusion model, now open-sourced.
  • Introduces Levenshtein editing (KEEP, SUBSTITUTE, DELETE, INSERT), enabling diffusion models to dynamically modify sequence length for the first time.
  • L-EBPO reinforcement learning method allows environmental feedback to drive editing decisions, blocking error propagation at the trajectory level.
  • Native support for 128K context, achieved through staged training and Block Routing MoE.
  • Averages 53.83 across 7 agent benchmarks, close to the autoregressive model Ling-2.6-flash's 55.74, and leads on tasks like τ²-Bench.
  • BF16 throughput reaches 1.64x that of Ling-2.6-flash, with an additional 18.6% improvement after FP8 quantization.
Open section navigationFrom Static Generation to Dynamic Correction: Levenshtein Editing Breaks Length Rigidity

From Static Generation to Dynamic Correction: Levenshtein Editing Breaks Length Rigidity

Traditional diffusion models (e.g., LLaDA2.1) are limited to equal-length replacements (KEEP and SUBSTITUTE), unable to handle structural deficiencies like deleting redundancies or inserting new content. LLaDA2.2 expands the action space to four atomic operations: KEEP, SUBSTITUTE, DELETE, and INSERT, where DELETE and INSERT change sequence length via edit control tokens. This is the first time Levenshtein editing has been integrated at scale into the denoising process of a diffusion model.

Supervision signals are derived by computing the longest common subsequence (LCS) between the model draft and ground truth: matching anchors are labeled KEEP, aligned positions in gaps are labeled SUBSTITUTE, extra tokens in the draft are labeled DELETE, and extra tokens in the target are labeled INSERT. Ablation experiments show that on SWE-bench Verified, enabling Levenshtein editing improves the score from 35.8 to 44.4, an absolute gain of 8.6 percentage points.

L-EBPO: Environment Feedback-Driven Editing Strategy Optimization

LLaDA2.2 proposes a Levenshtein editing evidence lower bound-based chunked policy optimization method (L-EBPO), which models editing decisions in multi-turn interactions as a reinforcement learning problem. The outer EBPO optimizes trajectory-level decisions, while the inner layer manages intra-chunk concatenation editing, unifying both layers under the same objective function by expanding the action space.

Reward signals come entirely from environmental feedback, composed of three components: correctness of tool call execution, validity of output format, and overall task completion. This calibrates the model's editing strategy to the environment rather than relying on preset rules, blocking error propagation in long-horizon interactions at the trajectory level.

128K Native Context and Block Routing MoE

LLaDA2.2 extends from LLaDA2.1's 8K context to 128K, a 16x increase. It employs staged training: first continuing training on 64K length for 300B tokens, then extending to 128K for another 200B tokens. Training data primarily consists of long documents and repository-level code, using document-aware packing and attention boundaries to avoid interference between unrelated documents. Agent data is only added during the 128K stage.

To address load balancing issues in block diffusion MoE, a Block Routing design is adopted: routing units align with diffusion generation units, computing block-level admission scores via a token racing strategy. Each block has a fixed-capacity expert pool with an upper bound of O(C). The technical report states that switching to block routing after post-training has almost no detrimental effect on model quality.

Agent and General Benchmark Evaluation: Competing Head-to-Head with Autoregressive Models

Evaluation covers 17 benchmarks (7 agent, 10 general), comparing with the same-parameter autoregressive model Ling-2.6-flash. On the 7 agent benchmarks, LLaDA2.2-flash averages 53.83, while Ling-2.6-flash scores 55.74. Specific leads include: τ²-Bench (80.33 vs 76.36), PinchBench (81.66 vs 81.30), MCP-Atlas (46.21 vs 41.12); Claw-Eval is roughly tied (64.22 vs 64.56).

On general benchmarks, LLaDA2.2-flash surpasses Ling-2.6-flash on LongBench v2 (45.13 vs 42.94), validating the effectiveness of the 128K context extension. In terms of efficiency, BF16 average throughput reaches 1.64x that of Ling-2.6-flash, with an additional 18.6% improvement after FP8 quantization.

Credibility boundary

This article's information primarily comes from Ant Group's official release and technical report, as reported by Machine Heart. Evaluation results and performance data are from the official report and have not been independently verified by third parties.

Insight takeaway

LLaDA2.2 demonstrates that diffusion language models can not only handle agent tasks but also offer significant efficiency advantages, providing a new foundational model option for cost-sensitive agent deployment.

Sources for this version

  1. 全球⾸个⼤规模 Agentic 扩散模型来了！机器之心

Primary report

机器之心

Primary source