ai-reviewer-diagnostics 0.1.1

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

Where Do LLMs Go Wrong? Diagnosing Automated Peer Review

The official repository for our CIKM 2025 paper, "Where Do LLMs Go Wrong? Diagnosing Automated Peer Review via Aspect-Guided Multi-Level Perturbation".

It also provides a pip-installable diagnostic toolkit for black-box evaluation of automated peer-review systems: run any automated review system on paired original/perturbed paper, review, or rebuttal inputs; export scores or decisions; then generate aspect-level reports measuring sensitivity to soundness, presentation, contribution, tone, factuality, completeness, and recommendation perturbations.

Companion code, prompts, examples, and reproducibility notes for the CIKM 2025 paper:

Where Do LLMs Go Wrong? Diagnosing Automated Peer Review via Aspect-Guided Multi-Level Perturbation
Jiatao Li, Yanheng Li, Xinyu Hu, Mingqi Gao, Xiaojun Wan. CIKM 2025.
DOI: https://doi.org/10.1145/3746252.3761274

If this repository helps your research, please cite the paper. Copy-paste BibTeX is below and in CITATION.bib; GitHub citation metadata is in CITATION.cff.

30-second quickstart

Install the diagnostic CLI from PyPI and run the toy report. This path needs no API keys, GPUs, model downloads, or companion dataset.

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

If you want the latest GitHub version before a PyPI release catches up:

python -m pip install "git+https://github.com/leejamesss/where-do-llms-go-wrong.git"

Expected package demo output:

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

For a repo checkout:

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

Expected checkout quickstart output:

AI-reviewer diagnostic release quickstart: OK
Validated 1 chat example(s), 3 OpenReview note(s).
Prompt rows: base=9, perturb=7.
Wrote outputs/quickstart/quickstart_summary.json

make quickstart validates the repo layout, example schemas, prompt files, and citation metadata, then writes a tiny demo artifact under outputs/quickstart/. make demo-report exercises the same packaged report engine exposed as ai-reviewer-diagnostics. These are format/schema demos, not model results.

What you can reuse

Goal	Start here	Requires
Check that the repo is healthy	make quickstart	Python only
Run the API-free code smoke test	uv sync && uv run make smoke-test	lightweight Python deps
Generate a toy diagnostic report	make demo-report or ai-reviewer-diagnostics ...	Python only
Reuse the prompt templates	prompts/	none
Run API-based model inference	scripts/run_openrouter.py, scripts/run_gemini.py	API key
Run local model inference	scripts/run_vllm.py	GPU + vLLM
Evaluate a new review system's outputs	ai-reviewer-diagnostics / ai_reviewer_diagnostics	JSONL outputs with shared id fields
Inspect released artifacts	docs/DATA.md + Hugging Face dataset	huggingface_hub
Recreate analysis tables/figures	analysis/ + docs/REPRODUCIBILITY.md	dataset + analysis deps

Command map

Command	Purpose
ai-reviewer-diagnostics --demo --output-md outputs/demo.md	Verify the pip-installed CLI with bundled toy fixtures.
ai-reviewer-diagnostics --baseline base.jsonl --perturbed pert.jsonl --condition paper/soundness --output-md report.md	Diagnose one baseline/perturbed pair from your own system.
ai-reviewer-diagnostics --scores-dir ai-reviewer-diagnostic-data/data/annotation_scores --output-md report.md	Summarize all paired score files in the released dataset format.
make quickstart	Check a repo checkout without installing dependencies.
make smoke-test	Run the API-free repository test suite.

Integrate your own review system

The toolkit only requires JSONL outputs with shared id values. Start with docs/INTEGRATIONS.md for schema examples, custom score fields, directory mode, and common pitfalls.

Dependencies

Dependencies and package metadata are managed in pyproject.toml. The install exposes two console commands, ai-reviewer-diagnostics and the shorter alias ai-reviewer-report. Analysis and vLLM dependencies are optional extras.

uv sync                    # default runtime dependencies
uv sync --extra analysis   # pandas/numpy/scipy plotting stack
uv sync --extra vllm       # optional local GPU inference stack

If you do not use uv, the pip-compatible fallback is:

python -m pip install -e .
python -m pip install -e ".[analysis]"
python -m pip install -e ".[vllm]"
ai-reviewer-diagnostics --help

API-free smoke test

uv sync
uv run make smoke-test

This compiles Python files, validates example JSON, checks all inference runners in --validate-only mode, and runs the OpenReview-cleaner fixture. Generated files go under outputs/ and can be removed with:

make clean

Dataset

The core paired pre-/post-perturbation content artifacts are hosted separately on Hugging Face so the GitHub repo stays lightweight:

https://huggingface.co/datasets/leejamesssss/ai-reviewer-diagnostic-data

Download and inspect:

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

Expected summary starts with file count, total size, file types, JSONL row counts, and largest files. In the current release, content_pairs/ is the core data contribution: before/after perturbation pairs. annotation_scores/ contains our experiment score outputs and summary tables. See docs/DATA.md for schema and naming notes.

Diagnostic toolkit workflow

To evaluate a new automated review system, export its baseline and perturbed outputs as JSONL with a shared id field and score/decision fields:

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

Then generate an aspect-level report:

make demo-report
# or, for your own system outputs:
uv 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

The report summarizes score deltas, decision-change rates, and top decision transitions. If you use the public dataset score-file naming convention, run directory mode:

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

Common commands

OpenAI-compatible / OpenRouter inference

export OPENROUTER_API_KEY=***
uv run python scripts/run_openrouter.py \
  --input examples/example.json \
  --output outputs/model_outputs.jsonl \
  --model mistralai/mistral-small-3.1-24b-instruct \
  --base-url https://openrouter.ai/api/v1 \
  --api-key-env OPENROUTER_API_KEY \
  --workers 1

Gemini inference

export GEMINI_API_KEY=***
uv run python scripts/run_gemini.py \
  --input examples/example.json \
  --output outputs/gemini_outputs.jsonl \
  --model gemini-2.0-flash \
  --workers 1

Optional local vLLM inference

vllm is intentionally kept out of the default install because it depends on your CUDA, PyTorch, and GPU setup.

uv sync --extra vllm
uv run python scripts/run_vllm.py \
  --input examples/example.json \
  --output outputs/vllm_outputs.jsonl \
  --model-path Qwen/Qwen2.5-72B-Instruct \
  --tensor-parallel-size 8 \
  --limit 1

Clean an OpenReview export

uv run python scripts/clean_openreview.py \
  --input examples/openreview_comments_minimal.json \
  --output outputs/openreview_conversations.json \
  --forum-id forum_example \
  --print-text

For your own data, replace examples/openreview_comments_minimal.json with an OpenReview comments export.

Repository layout

ai_reviewer_diagnostics/ # pip-installable diagnostic report package
scripts/              # wrappers and reusable CLIs: quickstart, inference, preprocessing, data summary
analysis/             # analysis scripts for released experiment score outputs
examples/             # tiny runnable fixtures for quickstart, smoke tests, and report generation
prompts/              # curated machine-readable prompt templates
  base_prompt.jsonl
  perturb_prompt.jsonl
data/README.md        # pointer to the external Hugging Face dataset
docs/                 # getting-started, data, and reproducibility notes
paper/README.md       # DOI, ACM PDF link, and citation pointer
CITATION.bib          # BibTeX citation
CITATION.cff          # GitHub citation metadata
CONTRIBUTING.md       # community contribution guide
MANIFEST.md           # full release inventory

Reproducibility path

The release is organized in tiers so users can get value from the public code, prompts, examples, and released artifacts:

1. Immediate check: make quickstart validates layout and schemas with Python only.
2. Code smoke test: make smoke-test checks scripts without API calls or GPUs.
3. Artifact inspection: download the Hugging Face dataset and inspect content_pairs/ first; this is the reusable before/after perturbation benchmark. perturbed_contents/ keeps perturbed-only artifacts for alignment with the original experiments.
4. Model inference: run OpenRouter, Gemini, or vLLM wrappers on your own prompt batches.
5. Diagnostic report: compare a system's baseline and perturbed outputs with ai-reviewer-diagnostics.
6. Analysis: use analysis/ scripts on downloaded experiment score artifacts when reproducing our reported score analyses.

See docs/GETTING_STARTED.md and docs/REPRODUCIBILITY.md for the longer guide.

Community contributions

Bug reports, integration requests, and metric ideas are welcome. Use the GitHub issue templates for reproducible CLI bugs, new automated-review-system integrations, or report-metric proposals. Pull requests should keep the default install lightweight and API-free smoke tests passing; see CONTRIBUTING.md.

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}
}

Documentation map

• docs/GETTING_STARTED.md: shortest path for a new user.
• docs/DATA.md: dataset contents, schemas, naming glossary, and rights notes.
• docs/REPRODUCIBILITY.md: tiered reproduction guide.
• docs/INTEGRATIONS.md: connect a new automated-review system.
• scripts/README.md: CLI map and examples.
• analysis/README.md: analysis script guide.
• prompts/README.md: prompt reuse notes.
• CONTRIBUTING.md: issue/PR workflow for community contributions.
• MANIFEST.md: release inventory.

Release status

The GitHub repository contains code, curated prompts, docs, examples, citation metadata, and pointers to the paper/data. The ACM paper PDF is linked from paper/README.md rather than redistributed here. Code is MIT licensed in LICENSE; the dataset is hosted separately on Hugging Face under its dataset-card terms.

Contact

Open a GitHub issue in the public repository for questions, reuse requests, or reproduction problems.