MAGE: All-[MASK] Block Already Knows Where to Look in Diffusion LLM

Omin Kwon, Yeonjae Kim, Doyeon Kim, Minseo Kim, Yeonhong Park, Jae W. Lee · Feb 15, 2026 · Citations: 0

Abstract

Block diffusion LLMs are emerging as a promising next paradigm for language generation, but their use of KV caching makes memory access a dominant bottleneck in long-context settings. While dynamic sparse attention has been actively explored, existing methods designed for autoregressive LLMs rely on approximate importance estimation and perform poorly when adapted to block diffusion. This work identifies a key opportunity unique to block diffusion: attention at the first All-[MASK] denoising step reliably predicts important KV entries and budget requirements, enabling MAGE to perform a single exact attention pass per block and reuse it for training-free sparse denoising. Across long-context benchmarks including LongBench and Needle-in-a-Haystack, MAGE achieves near-lossless accuracy with a fraction of the KV budget while delivering up to 3-4x end-to-end speedup, consistently outperforming AR-oriented sparse attention baselines. A lightweight fine-tuning strategy further strengthens [MASK]-guided patterns with minimal cost, requiring only a few hours of training on a single NVIDIA H100 GPU for both 1.5B and 7B models.

HFEPX Relevance Assessment

This paper appears adjacent to HFEPX scope (human-feedback/eval), but does not show strong direct protocol evidence in metadata/abstract.

Eval-Fit Score

5/100 • Low

Treat as adjacent context, not a core eval-method reference.

Human Feedback Signal

Not explicit in abstract metadata

Evaluation Signal

Detected

HFEPX Fit

Adjacent candidate

If you are doing eval pipeline work, start here:

Human Eval Hub LLM-as-Judge Hub Pairwise Preference Hub Tool-Use Eval Hub

Human Data Lens

Uses human feedback: No
Feedback types: None
Rater population: Unknown
Unit of annotation: Unknown
Expertise required: General
Extraction source: Persisted extraction

Evaluation Lens

Evaluation modes: Automatic Metrics
Agentic eval: None
Quality controls: Not reported
Confidence: 0.45
Flags: low_signal, possible_false_positive

Protocol And Measurement Signals

Benchmarks / Datasets

LongBenchNeedle In A Haystack

Reported Metrics

accuracycost

Research Brief

Deterministic synthesis

Across long-context benchmarks including LongBench and Needle-in-a-Haystack, MAGE achieves near-lossless accuracy with a fraction of the KV budget while delivering up to 3-4x end-to-end speedup, consistently outperforming AR-oriented sparse… HFEPX signals include Automatic Metrics with confidence 0.45. Updated from current HFEPX corpus.

Generated Mar 5, 2026, 3:22 AM · Grounded in abstract + metadata only

Key Takeaways

Across long-context benchmarks including LongBench and Needle-in-a-Haystack, MAGE achieves near-lossless accuracy with a fraction of the KV budget while delivering up to 3-4x…

Researcher Actions

Treat this as method context, then pivot to protocol-specific HFEPX hubs.
Cross-check benchmark overlap: LongBench, Needle In A Haystack.
Validate metric comparability (accuracy, cost).

Caveats

Generated from title, abstract, and extracted metadata only; full-paper implementation details are not parsed.
Low-signal flag detected: protocol relevance may be indirect.

Recommended Queries

human-eval protocol design pairwise preference data quality inter-rater agreement adjudication

Research Summary

Contribution Summary

Across long-context benchmarks including LongBench and Needle-in-a-Haystack, MAGE achieves near-lossless accuracy with a fraction of the KV budget while delivering up to 3-4x end-to-end speedup, consistently outperforming AR-oriented sparse…

Why It Matters For Eval

Across long-context benchmarks including LongBench and Needle-in-a-Haystack, MAGE achieves near-lossless accuracy with a fraction of the KV budget while delivering up to 3-4x end-to-end speedup, consistently outperforming AR-oriented sparse…

Researcher Checklist

Gap: Human feedback protocol is explicit

No explicit human feedback protocol detected.
Pass: Evaluation mode is explicit

Detected: Automatic Metrics
Gap: Quality control reporting appears

No calibration/adjudication/IAA control explicitly detected.
Pass: Benchmark or dataset anchors are present

Detected: LongBench, Needle In A Haystack
Pass: Metric reporting is present

Detected: accuracy, cost

Category-Adjacent Papers (Broader Context)

These papers are nearby in arXiv category and useful for broader context, but not necessarily protocol-matched to this paper.

FlashEvaluator: Expanding Search Space with Parallel Evaluation Category Neighbor

Citations: 0 Relevance: 4.45
- Shared arXiv category (cs.LG, cs.CL)
- Shared metric mentions
- Shared terminology (accuracy)
Cross-Family Speculative Prefill: Training-Free Long-Context Compression with Small Draft Models Category Neighbor

Citations: 0 Relevance: 4.10
- Shared arXiv category (cs.CL)
- Shared metric mentions
- Shared terminology (accuracy, cost)
Faster, Cheaper, More Accurate: Specialised Knowledge Tracing Models Outperform LLMs Category Neighbor

Citations: 0 Relevance: 4.10
- Shared arXiv category (cs.CL)
- Shared metric mentions
- Shared terminology (accuracy, cost)
Think, But Don't Overthink: Reproducing Recursive Language Models Category Neighbor

Citations: 0 Relevance: 3.95
- Shared arXiv category (cs.CL)
- Shared benchmark mentions
- Shared metric mentions
- Shared terminology (accuracy)
Efficient Self-Evaluation for Diffusion Language Models via Sequence Regeneration Category Neighbor

Citations: 0 Relevance: 3.30
- Shared arXiv category (cs.CL)
- Shared metric mentions
- Shared terminology (accuracy, diffusion)
PonderLM-3: Adaptive Token-Wise Pondering with Differentiable Masking Category Neighbor

Citations: 0 Relevance: 3.30
- Shared arXiv category (cs.CL)
- Shared metric mentions
- Shared terminology (cost, mask)
MUSE: A Run-Centric Platform for Multimodal Unified Safety Evaluation of Large Language Models Category Neighbor

Citations: 0 Relevance: 3.20
- Shared arXiv category (cs.LG, cs.CL)
According to Me: Long-Term Personalized Referential Memory QA Category Neighbor

Citations: 0 Relevance: 2.85
- Shared arXiv category (cs.CL)
- Shared metric mentions
- Shared terminology (accuracy)

Need human evaluators for your AI research? Scale annotation with expert AI Trainers.

Post a Job Get a Quote