ai-reviewer-diagnostics 0.1.4

Diagnostic toolkit for black-box evaluation of automated peer-review systems.

Where Do LLMs Go Wrong? Diagnosing Automated Peer Review

Official repository for the CIKM 2025 paper “Where Do LLMs Go Wrong? Diagnosing Automated Peer Review via Aspect-Guided Multi-Level Perturbation.”

Use this repo to evaluate automated peer-review systems on paired original/perturbed inputs and generate aspect-level diagnostic reports. The data lives on Hugging Face; the installable reporting CLI lives on PyPI.

Start fast

Install the CLI and run the bundled demo. No dataset, API key, GPU, or model download is needed.

python -m pip install ai-reviewer-diagnostics
ai-reviewer-diagnostics --demo --output-md outputs/demo_diagnostic_report.md

Expected output:

Compared 1 condition pair(s).
Wrote outputs/demo_diagnostic_report.md

For a repo checkout:

git clone https://github.com/PKU-ONELab/where-do-llms-go-wrong
cd where-do-llms-go-wrong
make quickstart
make demo-report

What is included

Need	Use
Diagnostic report CLI	ai-reviewer-diagnostics / ai-reviewer-report
Main paired perturbation data	HF dataset → data/content_pairs/*.jsonl
Released score artifacts	HF dataset → data/annotation_scores/*.jsonl
Prompt templates	prompts/
API / local inference wrappers	scripts/
Paper-analysis scripts	analysis/
Reproduction notes	docs/REPRODUCIBILITY.md

Use the dataset

The primary dataset is before/after perturbation pairs:

from datasets import load_dataset
pairs = load_dataset("leejamesssss/ai-reviewer-diagnostic-data", split="train")
print(pairs[0].keys())  # id, source, aspect, content_before, content_after

Download the full HF repo when you need manifests or score artifacts:

hf download leejamesssss/ai-reviewer-diagnostic-data   --repo-type dataset   --local-dir ai-reviewer-diagnostic-data
python scripts/summarize_release_data.py --data-dir ai-reviewer-diagnostic-data/data

content_pairs/ is the canonical benchmark surface. annotation_scores/ is for reproducing or auditing the paper’s released scoring outputs. The duplicated perturbed-only view is intentionally excluded because content_pairs.content_after already contains it.

Evaluate your own review system

Export baseline and perturbed outputs as JSONL with shared id values and any score/decision fields:

{"id":"paper_001","overall_score":8,"soundness_score":4,"final_decision":"Accept as Poster"}

Then run:

ai-reviewer-diagnostics   --baseline outputs/my_system_baseline.jsonl   --perturbed outputs/my_system_soundness_perturbed.jsonl   --condition paper/soundness   --output-md reports/my_system_soundness_report.md   --output-json reports/my_system_soundness_report.json

Directory mode works for released score files:

ai-reviewer-diagnostics   --scores-dir ai-reviewer-diagnostic-data/data/annotation_scores   --output-md reports/released_scores_report.md

The report summarizes score deltas, decision-change rates, and top decision transitions. See docs/INTEGRATIONS.md for schemas and custom fields.

Development commands

uv sync                    # default runtime deps
uv sync --extra analysis   # analysis deps
uv sync --extra vllm       # optional local GPU inference deps
uv run make smoke-test     # API-free checks
make clean                 # remove generated outputs

Inference wrappers:

python scripts/run_openrouter.py --input examples/example.json --output outputs/openrouter.jsonl --model <model> --api-key-env OPENROUTER_API_KEY
python scripts/run_gemini.py --input examples/example.json --output outputs/gemini.jsonl --model gemini-2.0-flash
python scripts/run_vllm.py --input examples/example.json --output outputs/vllm.jsonl --model-path <hf-or-local-model>

Repository map

ai_reviewer_diagnostics/  # pip package and diagnostic report CLI
scripts/                  # inference, preprocessing, quickstart, data-summary CLIs
analysis/                 # scripts for released score artifacts
examples/                 # tiny fixtures for demos/tests
prompts/                  # prompt templates
data/README.md            # pointer to HF dataset
docs/                     # data, integrations, reproducibility
paper/README.md           # DOI / citation pointer

Citation

@inproceedings{li2025where,
  title     = {Where Do LLMs Go Wrong? Diagnosing Automated Peer Review via Aspect-Guided Multi-Level Perturbation},
  author    = {Li, Jiatao and Li, Yanheng and Hu, Xinyu and Gao, Mingqi and Wan, Xiaojun},
  booktitle = {Proceedings of the 34th ACM International Conference on Information and Knowledge Management (CIKM '25)},
  year      = {2025},
  publisher = {ACM},
  doi       = {10.1145/3746252.3761274},
  url       = {https://doi.org/10.1145/3746252.3761274}
}

Docs: GETTING_STARTED, DATA, REPRODUCIBILITY, INTEGRATIONS, CONTRIBUTING.