GradCuit

KO | EN

gitlite — search

#macos #css #python #javascript #graphql #postgresql #list #cloud-native #svg #pytorch

(★ 9)

No description available.

파일 탐색기

ZIP 다운로드

- - pages.yml
- - index.html
- - __next._full.txt
  - __next._head.txt
  - __next._index.txt
  - __next._not-found.txt
  - __next._tree.txt
  - index.html
  - index.txt
- .gitkeep
- .nojekyll
- 404.html
- __next.__PAGE__.txt
- __next._full.txt
- __next._head.txt
- __next._index.txt
- __next._tree.txt
- gradient-attribution.png
- index.html
- index.txt
- layer-sensitivity.png
- learning-rate.png
- main-teaser.png
- nlco-logo.png
- og.png
- random-vs-latentseek.png
- run_example.sh
- start_verifier.sh
- - __init__.py
  - grader.py
  - math_equivalent_MATH.py
  - parse_utils_qwen.py
  - utils.py
- - __init__.py
  - gradcuit.py
- - __init__.py
  - single_verifier_prompts.py
  - solver_prompts.py
- - __init__.py
  - single_verifier_reward.py
- data.py
- evaluation.py
- main.py
.gitignore
README.md
requirements.txt

# CDN으로 사용하기

jsDelivr

jsDelivr는 공개 GitHub 리포지토리를 별도 설정 없이 CDN으로 즉시 서빙합니다. 버전과 파일을 고르면 웹페이지에 바로 붙일 수 있는 링크와 예시 코드가 만들어집니다.

버전

파일

링크

예시

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

Zhaoxin Yu&#42;, Qi Shen&#42;, Hengli Li&#42;†, Zhaowei Zhang, Song-Chun Zhu, Chi Zhang✉, and Zilong Zheng✉

&#42; Equal Contributions. † Project lead. ✉ Equal Supervision.

[![Project Page](https://img.shields.io/badge/Project-Page-BA8E9F?style=for-the-badge)](https://yuzhaoxin946.github.io/GradCuit/)
[![Paper](https://img.shields.io/badge/arXiv-2608.02585-B31B1B?style=for-the-badge)](https://arxiv.org/abs/2608.02585)

GradCuit (*gradient through circuit*) is a test-time latent reasoning method for frozen
autoregressive transformers. It places a small set of continuous latent tokens directly in the
transformer's input embedding sequence, alongside the embedded prompt. Generated tokens attend to
these latents through standard self-attention, while token-probability gradients flow backward
through the same attention connectivity to assign credit to each latent.

For each problem, GradCuit evaluates the generated answer with a single verifier and applies a
policy-gradient-style update to the latent tokens. The language model parameters remain frozen;
only the instance-specific latents are updated before regenerating the answer.

This implementation supports GSM8K, MATH-500, and GPQA-Diamond.
Answer extraction and final-answer judging use the same evaluation pipeline as the
original single-verifier experiments.

## Installation

Python 3.10 or newer is required. Install a CUDA-compatible PyTorch build for your system, then run:

```bash
pip install -r requirements.txt
```

MATH-500 evaluation uses `latex2sympy2_extended==1.11.0` with
`antlr4-python3-runtime==4.11.1`, matching the parser used by `math-verify==0.9.0`.

The verifier uses a vLLM OpenAI-compatible server. The solver and verifier normally run in separate
processes and may use different GPUs.

## Quick start

Edit the static model and GPU settings in the two example scripts. Start the verifier first:

```bash
bash scripts/start_verifier.sh
```

Then run the GradCuit example:

```bash
bash scripts/run_example.sh
```

The example uses the public model identifier `meta-llama/Llama-3.2-3B-Instruct`, GSM8K, the relative
output directory `./output`, a learning rate of `0.001`, and `2048` maximum new tokens. It optimizes
the inserted prefix at decoder block input `14`, the middle of the model's 28 decoder blocks.

GPQA Diamond is gated. Accept its dataset terms and authenticate with Hugging Face before using
`gpqa_diamond`.

## Main arguments

`src/main.py` requires all experiment-defining arguments explicitly. Run:

```bash
python src/main.py --help
```

`--solver_prompt_idx` accepts `0` for boxed answers and `1` for JSON answers.
`--optimize_layer_idx 0` optimizes prefix embeddings; a positive value optimizes the prefix hidden
state at the input to that decoder block. `--optimizer` accepts `adam`, `sgd`, or `muon`.
`--max_num_steps` must be positive; the paper uses `10`. Set `--grad_clip 0` to disable clipping and
`--num_data -1` to process all remaining examples.

Use `--resume` with the same experiment arguments to continue the newest matching output directory.
The API key, when needed, is read from `VLLM_API_KEY`.

## Outputs

Each run writes:

- `run_args.json`: sanitized experiment settings.
- `test/<index>_data.json`: prompt, original output, optimization history, and final correctness.
- `logistics.pt`: resume state and aggregate counters.

Generated outputs are ignored by Git.

## Validation

```bash
python -m compileall src
python src/main.py --help
vllm --help
```

## Citation

```bibtex
@misc{yu2026gradcuitcreditassignedgradientflow,
      title={GradCuit: Credit-Assigned Gradient Flow Enables Robust and Interpretable Test-Time Latent Reasoning},
      author={Zhaoxin Yu and Qi Shen and Hengli Li and Zhaowei Zhang and Song-Chun Zhu and Chi Zhang and Zilong Zheng},
      year={2026},
      eprint={2608.02585},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2608.02585},
}
```

**Contact:** [Zhaoxin Yu](mailto:yuzhaoxin2024@ia.ac.cn), [Qi Shen](mailto:shenqi@bupt.edu.cn), [Hengli Li](mailto:lihengli@stu.pku.edu.cn), [Chi Zhang](mailto:chizhang.cz@pku.edu.cn), and [Zilong Zheng](mailto:zlzheng@bigai.ai) (corresponding authors: Hengli Li, Chi Zhang, and Zilong Zheng)

GradCuit

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

Download README