title	Citation Benchmark
colorFrom	blue
colorTo	indigo
sdk	docker
app_port	7860
pinned	false

Citation-Agent — OpenEnv Environment

Citation-Agent is an OpenEnv-compliant environment for training and evaluating agents on academic literature review and citation selection.

Given a claim, the agent must navigate a localized research database to search across papers, read abstracts, traverse citation graphs, and ultimately select the correct paper that supports the claim.

Motivation

Most agent benchmarks focus on web scraping or terminal navigation. This environment tests an agent's ability to reason over complex scholarly graphs — a real-world task that researchers actually perform daily. The agent must:

• Differentiate between ArXiv IDs and Semantic Scholar Corpus IDs
• Follow citation intents and contexts
• Run SQL-backed queries to find papers in a 891MB database
• Avoid getting stuck in infinite search loops

Action Space

All actions are JSON objects with an action_type field:

Action	Parameters	Description
search	query: str	Search papers by title (SQL LIKE). Returns up to 10 results with corpus_id, arxiv_id, title, year.
read_abstract	paper_id: str	Read the abstract of a paper by its arxiv_id.
get_citations	paper_id: str	Get citation context for a paper by its corpus_id. Returns up to 5 citing papers with context, intent, title, abstract.
submit	paper_id: str	Submit the answer. Graded against ground truth. Terminal action.

Observation Space

JSON containing:

Field	Type	Description
current_claim	str	The claim/task the agent must resolve
search_results	list[dict]	Results from the last search action
last_abstract	str	Abstract from the last read_abstract action
citations_data	list[dict]	Citation data from the last get_citations action
message	str	Feedback message from the environment
step_count	int	Current step in the episode

Reward Function

The reward function provides dense signal over the full trajectory:

• Step penalty: -0.03 per step (encourages efficiency)
• Search rewards: +0.08 to +0.18 for finding relevant papers; -0.12 for no results; -0.22 for repeated no-op searches
• Read abstract: +0.06 to +0.12 based on abstract quality; -0.1 for invalid ID
• Get citations: +0.06 to +0.15 based on number found; -0.08 for no results
• Submit correct: +1.0 plus efficiency bonus (up to +0.2 for solving quickly)
• Submit wrong: -0.6 penalty
• Max steps exceeded: -1.0

Tasks (50 total)

Tasks span three difficulty levels with deterministic graders:

Difficulty	Count	Description	Example
Easy	~17	Direct title search → submit	"Search for paper X. Submit its ID."
Medium	~17	Search + verify metadata → submit	"Find paper X. Verify metadata and return ID."
Hard	~16	Search + check citations → submit	"Search for paper X, check its citations, and submit its ID."

Each grader checks the submitted paper_id against ground truth via exact title match (case-insensitive). Score is binary: 1.0 (correct) or 0.0 (wrong).

Setup and Usage

Prerequisites

• Python 3.10+
• Docker (for containerized deployment)

Install Dependencies

pip install -r requirements.txt

Set Environment Variables

export HF_TOKEN="your_huggingface_api_key"
export MODEL_NAME="llama-3.3-70b-versatile"
export API_BASE_URL="https://api.groq.com/openai/v1"

Run Inference (Baseline)

python inference.py

This runs the LLM agent against all 50 tasks, producing structured [START], [STEP], [END] logs.

Run via Docker

docker build -t citation-agent .
docker run -p 7860:7860 citation-agent

The environment server will be accessible at http://localhost:7860.

Deploy to Hugging Face Space

Push to your HF Space repository. The Dockerfile will:

1. Install dependencies
2. Download the 891MB SQLite database from ruby56/Citation-Database
3. Start the FastAPI OpenEnv server on port 7860

Baseline Scores

Using llama-3.3-70b-versatile via Groq API:

Difficulty	Avg Score	Avg Steps
Easy	~0.8	~2-3
Medium	~0.6	~3-5
Hard	~0.4	~5-8

Architecture

• environment.py: Core RL environment with actions, transitions, and dense reward logic
• tasks.py: 50 tasks with deterministic grader checks
• server/app.py: FastAPI OpenEnv server entry point
• server/citation_environment.py: OpenEnv-compatible environment wrapper
• inference.py: Baseline LLM inference script using OpenAI client
• openenv.yaml: OpenEnv manifest
• Dockerfile: Container definition for HF Spaces deployment

Database

The environment uses a 891MB SQLite database (citation_db.sqlite) containing:

• s2_papers: Semantic Scholar paper metadata (corpus_id, arxiv_id, title, year)
• arxiv_metadata: ArXiv paper abstracts
• citations: Citation graph with contexts and intents

The database is automatically downloaded from the public HF Dataset ruby56/Citation-Database at container build time.