Agentic Inference Systems

This project implements advanced agentic systems for complex inference tasks, focusing on Deep Research capabilities and Self-Refining generation strategies.

Project Structure

The repository focuses on two main agentic workflows:

1. Deep Research Agent (deep_research_agent/)

A comprehensive research agent capable of executing multi-step research tasks.

• Core capabilities:
  • Automated information gathering and synthesis
  • Multi-step reasoning and planning
  • Integration with external tools (Search, Browser)
• Key components:
  • react_agent.py: Implementation of the ReAct (Reasoning + Acting) paradigm.
  • mcp_agents/: Modular Component Protocol (MCP) agents for extensible tool use.
  • graph/: Graph-based reasoning utilities.

2. Self-Refining Agent (self_refine/)

An agentic system that iteratively improves its own outputs through self-correction.

• Core capabilities:
  • Self-evaluation of generated content
  • Iterative refinement loops
  • Performance analysis on benchmarks (MMLU, Graph tasks)
• Key components:
  • self_refine.py: Main logic for the self-refining loop.
  • refine_modal.py: Modal integration for scalable execution.
  • analyze_accuracy.py: Tools for evaluating refinement performance.

3. Reranking & Evaluation (rerank_outputs.py)

Tools for evaluating and selecting the best generations from multiple candidates.

• Implements various scoring mechanisms:
  • Scalar Reward Models using Skywork/Reward-Llama
  • Pairwise Reward Models using LLM-Blender (PairRM)
  • MBR (Minimum Bayes Risk) decoding with BLEU and BERTScore
  • Log-probability analysis using Qwen models

Setup & Installation

1. Install Dependencies:

   pip install -r requirements.txt

2. Environment Variables: Create a .env file with the following keys:

   OPENAI_API_KEY=your_key
   ANTHROPIC_API_KEY=your_key
   # Add other provider keys as needed

Usage

Running the Deep Research Agent

Navigate to the deep_research_agent directory and configure the agent in react_agent.yaml.

python deep_research_agent/react_agent.py

Running Self-Refinement

To run the self-refinement experiments:

python self_refine/self_refine.py --task mmlu --model qwen3-4b

Reranking Outputs

To evaluate generated outputs using the reranking system:

python rerank_outputs.py

This will process all_results_processed.json and compute scores for all candidates.

Analysis

Use calculate_stats_reranking.py to generate statistical analysis and plots comparing different reranking strategies against gold-standard evaluations.