ADu2021/efficient-dlm-ar-conversion
skills/skillxiv-v0.0.2-claude-opus-4.6/efficient-dlm-ar-conversion/SKILL.md
Systematically convert pretrained autoregressive models into efficient diffusion language models via block-wise attention and position-dependent masking. Efficient-DLM family (1.5B/4B/8B) maintains comparable accuracy to standard AR models while delivering 4.5× higher throughput.
$ install --global
skillsauthnpx skillsauth add ADu2021/skillXiv efficient-dlm-ar-conversionInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
Security Scan Results
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
3
Scanners Passed
9
Scanners in report
Clean
TrivyContainer and dependency vulnerability scanner
95%
Clean
SemgrepStatic code analysis for vulnerabilities
95%
Clean
mcp-scan (Snyk)Model Context Protocol security validation
95%
Skipped
Snyk (dep)Open source security scanning
50%
Skipped
Socket.devSupply chain security analysis
50%
Skipped
VirusTotalMulti-engine malware detection
50%
Skipped
CrowdStrikeAdvanced threat intelligence
50%
Skipped
OSV-ScannerOpen Source Vulnerability database check
50%
Skipped
OWASP Dep-Check
50%
Last scanned: Apr 17, 2026, 5:34 AM4.8s1 file scanned
SKILL.md
- name:
- efficient-dlm-ar-conversion
- title:
- Efficient-DLM: Converting Autoregressive Models to Diffusion Language Models with Superior Accuracy-Throughput Trade-offs
- version:
- 0.0.2
- engine:
- skillxiv-v0.0.2-claude-opus-4.6
- license:
- MIT
- url:
- https://arxiv.org/abs/2512.14067
- keywords:
- [diffusion-lm, autoregressive-conversion, parallel-generation, block-wise-attention, position-dependent-masking]
- description:
- Systematically convert pretrained autoregressive models into efficient diffusion language models via block-wise attention and position-dependent masking. Efficient-DLM family (1.5B/4B/8B) maintains comparable accuracy to standard AR models while delivering 4.5× higher throughput.
Skill Summary
Efficient-DLM presents a systematic framework for converting pretrained autoregressive models into parallel-decoding diffusion language models. The method combines block-wise attention with clean context, position-dependent token masking aligned with inference behavior, and comprehensive design analysis. Results show Efficient-DLM 8B maintains accuracy comparable to Qwen3 8B while delivering 4.5× higher throughput versus Dream 7B.
When To Use
- Converting existing AR checkpoints to parallel-decoding diffusion models
- Projects requiring dramatic throughput improvements without expensive retraining from scratch
- Scenarios balancing accuracy preservation with substantial speedup gains
- Research on efficient alternatives to autoregressive decoding
When NOT To Use
- Applications requiring maximum per-token quality over throughput
- Models where AR properties are fundamentally necessary
- Scenarios where the AR-to-DLM conversion introduces unacceptable artifacts
- Domains where single-pass AR inference already meets latency requirements
Core Technique
Three key technical components enable efficient AR-to-DLM conversion:
1. Block-wise Attention Pattern Employ "block-wise attention with clean context" where each corrupted block conditions only on previously decoded clean context. This better preserves pretrained AR model weights while enabling KV caching, avoiding full bidirectional attention that would substantially alter learned representations.
2. Position-Dependent Token Masking Identify training-test gap where uniform masking during training mismatches confidence-based sampling during inference (exhibiting left-to-right bias). Propose position-dependent masking assigning higher masking probabilities to later tokens:
w_i(t) = exp[β(1-t)i]
This aligns training with test-time behavior, improving sample efficiency.
3. Comprehensive Design Analysis Systematically study optimal block sizes, attention patterns, and training dynamics. Provide actionable guidelines for scalable AR-to-DLM conversion, enabling practitioners to apply the approach to different architectures.
Implementation Notes
Start with pretrained AR checkpoint. Implement block-wise attention maintaining context from previous blocks. Apply position-dependent masking with learned parameters aligned to inference-time confidence. Train progressively scaling block sizes. Validate accuracy preservation and measure throughput improvements over AR baselines.
References
- Original paper: Efficient-DLM (Dec 2025)
- Block diffusion language models
- Autoregressive model conversion techniques
Related Skills
ADu2021/flow-map-trajectory-tilting
testing
VerifiedTrustedCommunity
Uses flow maps as look-ahead operators to enable principled reward-guided diffusion by predicting trajectory endpoints at any denoising step. Deploy when applying rewards or preferences to diffusion trajectories with meaningful gradients throughout generation.
2SKILL.mdUpdated Apr 17, 2026
ADu2021/flexible-data-mixture-of-experts
testing
VerifiedTrustedCommunity
Train language models where each expert learns independently on closed datasets, enabling flexible inference with selective data inclusion or exclusion. 41% performance improvement while allowing users to opt out of specific data sources without retraining.
2SKILL.mdUpdated Apr 17, 2026
ADu2021/flexibility-trap-diffusion-reasoning
data-ai
VerifiedTrustedCommunity
Understand how token generation flexibility in diffusion LMs paradoxically constrains reasoning, as models exploit ordering flexibility to avoid uncertain tokens, and apply simplified approaches that preserve parallel decoding benefits. Use when optimizing diffusion-based language models for reasoning tasks.
2SKILL.mdUpdated Apr 17, 2026
ADu2021/flex-continuous-agent-evolution
devops
VerifiedTrustedCommunity
Enable LLM agents to improve continuously during deployment by constructing structured experience libraries through self-reflection on successes and failures—achieving 23% improvement on reasoning without gradient-based parameter updates or external training.
2SKILL.mdUpdated Apr 17, 2026