Polestar

NeurIPS 2026 / Training-free inference

Polestar

Drift-Aware Cache Calibration and Token Commitment
for Efficient Inference of Diffusion LLMs

Mingyu Lee1,*·Akshat Ramachandran1,*·Souvik Kundu2·Tushar Krishna1

1 Georgia Institute of Technology · 2 Intel AI Group

* Equal contribution

Code PaperWatch comparison

Representation drift becomes a shared signal for
selective cache calibration and early token commitment.

Up to

3.7×

higher throughput

Up to

+10.73pts

accuracy improvement

Reaching

3.67

tokens per forward pass

Reported across the paper’s mathematics and coding benchmarks. Accuracy gains are in percentage points. See results

Generation comparison

Compare Baseline, Fast-dLLM, and Polestar on the same prompt.

Illustrative responses; paper metrics and adjustable Replay TPS are shown separately.

Shared prompt


  

  

    

      
      
      
    

    

      
      
    

    
Ready to replay

  

  

  
Current-block commitPolestar suffix commitSeparated positions can commit together.

  

  
Authored examples, not recordings of a three-method run. Paper metrics are benchmark averages from the paper; Replay TPS controls tokenizer-token timing. Current and suffix commits share one total commit budget. Refresh selections and commit groups are schematic, without measured drift or confidence values.









  
The observation
The context changes.
So do the representations.

  
Diffusion LLMs refine a masked response with bidirectional attention. Revealing a new token changes the context available to other positions—even when those tokens stay fixed.
Polestar treats this representation drift as useful information: it reveals where cached states need calibration and helps identify predictions ready for commitment.

  

    
A new token reshapes its context.

    
48AprilclipsMay ✓

    
    
Inspired by the paper’s motivating example.

  

  
Polestar-Cache
Where should the cache be updated?
Rank the drift of cluster-level attention proxies and selectively refresh the relevant cached representations.
Polestar-Commit
Which tokens are ready to commit?
Combine current-token drift relative to recent history with confidence to increase decoding parallelism.






  
Interactive explanation
Follow the signal.
See how a sparse refresh step connects cache calibration and token commitment.

  

    
Polestar / method walkthroughIllustrative walkthrough of the paper’s method. Token positions are schematic.

    
Polestar method walkthroughCached context and a recomputed current block.
MaskedCached contextSelected / updatedCommitted

    
ContextStage 1 / 8
Reuse context. Recompute the block.
    
    
    
    
    
  
    
  
    Sparse refresh follows decoding progress. Drift prioritizes the update locations.
    
  
    
  The complete method
  
    
  
    Polestar-Cache & Polestar-Commit
    Connect the walkthrough to the original diagram.
    
  
    
  
    Original Polestar paper diagram: cluster summaries and sparse layer updates for Polestar-Cache; current and suffix token commitment for Polestar-Commit
    
  
    Polestar-Cache and Polestar-Commit.
    The full diagram from the paper.
    
  
    
  
    

    


    
  
    Results
    Accuracy and throughput,
    together.
    Mathematics, code, and vision-language reasoning.
    
  
    Accuracy–throughput trade-off.
    A better accuracy–throughput frontier.
    Cache calibration and token commitment address two linked dimensions of diffusion inference: the quality of cached context and the number of tokens decoded per forward pass.
    The paper evaluates Polestar across LLaDA, Dream, and LLaDA-V, using both mathematics and code benchmarks as well as visual reasoning tasks.
    Read the experiments 
    
  
    Explore the results
    Choose a model and generation length. Compare all three benchmarks.
    
  
    
  
    
    
    Model
    
    
    Generation length
    
  
    
  
    
  
    AccuracyTPSPolestar 
    
  
    Compare all three benchmarks. Select a method to highlight its measurements in the table.
    
  
    
  
    Paper measurements · NVIDIA A100 80GB · Batch size 1. Source data
    

    

  
    Results across all four models
    ModelBenchmarkAccuracy (%)TPFTPS
    

    


    
  
    Use Polestar
    Existing checkpoints.
    No additional training.
    Run text and image-conditioned generation with the released model presets.
    
  
    
  
    Examples
    
  
    Try it on your own inputs.
    Install Polestar, choose a model preset, and generate.
    Quickstart ↗Evaluation ↗
    

    


    Citation
    Cite Polestar
    Download
    

    


    Polestar method
    Enlarged Polestar figure