Ant Group Open-Sources LLaDA 2.2: A Self-Correcting Diffusion LLM for Agents
Ant Group's inclusionAI team released and open-sourced LLaDA 2.2, a large-scale diffusion language model designed for agentic tasks. Positioned as the first true diffusion LLM capable of agentic work, it supports planning and tool calling. It breaks the traditional left-to-right generation constraint of autoregressive models by enabling mid-generation self-correction.
Confirmed
Performance and Speed: LLaDA 2.2-flash supports native 128K context (using a progressive training strategy expanding from 8K). Its throughput averages 1.64 times that of Ling-2.6-flash, with an additional 18.6% improvement when combined with FP8 quantization.
Benchmarks: LLaDA 2.2-flash outperforms Ling-2.6-flash across multiple benchmarks. It scored 49.28 on SWE-bench Verified (vs. 31.88) and achieved 592.80 on τ²-Bench (vs. 334.90), with its fast mode reaching 705.30. It also maintains a lead in tests like SWE-bench Pro and SWE-bench Multilingual.
Self-Correction and Training: The model can self-correct during generation—identifying and fixing errors—while maintaining block-parallel speed. Its underlying mechanism uses Levenshtein editing, allowing four primitive operations: KEEP, SUBSTITUTE, DELETE, and INSERT. The training process has three stages: pre-training with Levenshtein editing, SFT with long-context and agent data, and reinforcement learning using L-EBPO to block error propagation in long tasks.
Why it matters
Traditional autoregressive LLMs generate strictly left-to-right, making it hard to backtrack and fix errors. By introducing a diffusion architecture and block-level editing, LLaDA 2.2 overcomes this limitation, enabling dynamic mid-generation correction. This not only significantly boosts inference throughput but also shows distinct advantages in complex agentic tasks requiring long-horizon planning and tool calling (such as code repair and multilingual software engineering), providing a vital reference for the practical deployment of non-autoregressive LLMs.
2026-07-25 ~ 2026-07-26 · 10 related posts
- Episode 1: Ant Group Open-Sources LLaDA 2.2: A Self-Correcting Diffusion LLM for Agents(2026-07-25, 10 posts)
- Episode 2: Ant's LLaDA2.2 Brings Diffusion LLMs to Long-Context Agents(2026-07-28, 2 posts)
Primary sources
- [source] Ant Group’s LLaDA2.2-flash edits its own generations and claims 1.64× speed — dr_cintas · 2026-07-25
- Ant Group releases LLaDA2.2-flash, a diffusion model that edits its own output mid-generation — dr_cintas · 2026-07-25
- [source] LLaDA 2.2 aims at real agents with 1.64× throughput and stronger SWE-bench scores — omarsar0 · 2026-07-26
- LLaDA 2.2 can delete, insert, and substitute tokens mid-generation — omarsar0 · 2026-07-26
- LLaDA 2.2-flash shows broad benchmark gains over Ling-2.6-flash — omarsar0 · 2026-07-26
- LLaDA 2.2 uses L-EBPO to cut error propagation in long agent runs — omarsar0 · 2026-07-26
- LLaDA 2.2’s block-diffusion MoE edits its own outputs during training and inference — omarsar0 · 2026-07-26
- LLaDA 2.2 claims 1.64× throughput and native 128K context — omarsar0 · 2026-07-26
- LLaDA 2.2-flash posts 592.8 on τ²-Bench and 705.3 in fast mode — omarsar0 · 2026-07-26
- [source] Ant Group’s inclusionAI team says LLaDA 2.2 weights and code are open — omarsar0 · 2026-07-26