Archer

Adaptive Reuse of Cached Hidden States for Efficient Rollback in Diffusion Language Models

Paper  |  PDF

Archer is a training-free prompt KV-caching method for rollback-capable diffusion language models (DLMs). It reuses the fixed prompt states, recomputes the response states that remain subject to revision, and refreshes the prompt cache when the response moves sufficiently far from its cached anchor. This repository provides the minimal path needed to reproduce the Archer rows in Table 1.

Quick start

1. Requirements

• Linux and git;
• Python 3.10 or 3.11;
• uv;
• one NVIDIA GPU with at least 40 GB of memory.

2. Install the environment

git clone https://github.com/Hxnng/Archer.git
cd Archer
./setup.sh

setup.sh creates the locked Python environment, downloads the pinned LLaDA-8B-Instruct checkpoint, verifies the bundled MBPP and HumanEval snapshots, and prepares the pinned LiveCodeBench evaluator and dataset.

If Hugging Face is unavailable from your network, set a compatible endpoint before setup, for example:

HF_ENDPOINT=https://hf-mirror.com ./setup.sh

An existing model checkout can be reused:

ARCHER_LLADA_MODEL_PATH=/path/to/LLaDA-8B-Instruct ./setup.sh

3. Reproduce the Archer rows

Run all three Table 1 benchmarks with their paper-default refresh radii:

CUDA_VISIBLE_DEVICES=0 ./reproduce.sh

Run a single benchmark:

CUDA_VISIBLE_DEVICES=0 ./reproduce.sh mbpp
CUDA_VISIBLE_DEVICES=0 ./reproduce.sh livecodebench
CUDA_VISIBLE_DEVICES=0 ./reproduce.sh humaneval

The default values are read from configs/table1_archer.json: MBPP uses $K=11$, LiveCodeBench uses $K=10$, and HumanEval uses $K=8$.

Choose another refresh radius

Pass --k to override the paper default without editing a configuration file:

CUDA_VISIBLE_DEVICES=0 ./reproduce.sh mbpp --k 16

Custom-radius outputs use a distinct run name such as table1_archer_mbpp_k16, so they cannot overwrite the default Table 1 run. The same override can be applied to all three benchmarks:

CUDA_VISIBLE_DEVICES=0 ./reproduce.sh all --k 16

Verify before a full run

Print the exact commands without downloading or loading a model:

./reproduce.sh --dry-run
./reproduce.sh humaneval --k 16 --dry-run

After setup, run two examples as an end-to-end smoke test:

CUDA_VISIBLE_DEVICES=0 ./reproduce.sh mbpp --limit 2

Interrupted runs resume by default. Add --no-resume to request a fresh run.

Every run records its arguments, model revision, GPU, per-problem generations, timing, and aggregate evaluation. Outputs are written below:

outputs/table1/table1_archer/
├── table1_archer_mbpp/
├── table1_archer_livecodebench/
└── table1_archer_humaneval/

Each run directory contains the raw JSONL generations, evaluation details, manifest, and summary.json.

Repository layout

configs/table1_archer.json       # Table 1 Archer defaults
reproduce.sh                     # user-facing reproduction entry point
scripts/run_experiment.py        # versioned config runner
src/archer/state_distance/       # Archer cache controller and sampler
src/archer/benchmark.py          # Archer generation and evaluation pipeline
benchmarks/                      # versioned benchmark snapshots and checksums

Evaluation executes model-generated Python programs. Run the artifact in an isolated environment without credentials or access to sensitive files.

Citation

If Archer is useful for your research, please cite our paper:

@misc{he2026archer,
  title         = {Archer: Adaptive Reuse of Cached Hidden States for Efficient Rollback in Diffusion Language Models},
  author        = {Xuning He and Zinan Sheng and Yongding Tao and Huanyu Liu and Ge Li and Xue Jiang and Yihong Dong},
  year          = {2026},
  eprint        = {2608.08086},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2608.08086}
}