Multiagent Debate

Orchestrates multiple LLM agents through structured debate rounds where agents iteratively build on each other's responses to refine answers. The system implements a generation phase that progresses from independent reasoning to collaborative refinement, with agents assigned distinct perspectives or roles across configurable debate rounds. Each round captures agent interactions as structured state, enabling systematic evaluation of how collaborative reasoning improves factuality and reasoning accuracy compared to single-agent baselines.

Provides modular generation modules for four distinct reasoning domains (Math, GSM, MMLU, Biography) that each implement specialized debate logic while accepting configurable parameters for agent count and debate round count. The generation phase processes domain-specific inputs through task-adapted prompts, manages agent state across rounds, and produces structured output files with naming conventions encoding experimental parameters (e.g., output_agents_N_rounds_R.json). This architecture enables systematic experimentation across different agent configurations without modifying core debate logic.

Implements evaluation modules (eval_gsm.py, eval_mmlu.py, eval_conversation.py) that systematically compare generated debate responses against ground truth data to measure improvements in factuality and reasoning accuracy. Each evaluation module encodes domain-specific metrics (e.g., exact match for math, factual accuracy for biography, multiple-choice accuracy for MMLU) and produces structured evaluation results. The framework enables quantitative comparison between single-agent baselines and multi-agent debate outputs, with results aggregated across test sets for statistical analysis.

Provides implementations for four distinct reasoning task domains (Math, Grade School Math, MMLU, Biography) with standardized generation and evaluation interfaces that enable systematic comparison across task types. Each task domain is implemented as a modular pair of generation and evaluation modules that follow consistent architectural patterns while accommodating domain-specific requirements. The system processes inputs through standardized pipelines, generating structured outputs with consistent naming conventions, enabling researchers to run identical debate experiments across different reasoning domains and compare relative improvements.

Provides abstraction layer for OpenAI API interactions, specifically integrating with the gpt-3.5-turbo-0301 model for all agent reasoning. The system manages API calls across multiple agents and debate rounds, handling request formatting, response parsing, and error handling. Integration points include agent prompt construction, response extraction from API outputs, and state management across sequential API calls. The abstraction enables swapping model versions or providers by modifying configuration, though current implementation is tightly coupled to OpenAI's API format.

Manages state across multiple debate rounds, tracking each agent's responses and building context for subsequent rounds. The system maintains agent response history, constructs prompts that reference previous round outputs, and ensures agents can build on each other's reasoning. State is stored in memory during execution and serialized to JSON output files for persistence and analysis. The architecture enables agents to see prior responses and refine their answers iteratively, implementing the core collaborative refinement mechanism of the debate approach.

Enables systematic experimentation by accepting configurable parameters (agent count, debate round count) and encoding them into output file names using standardized conventions (e.g., output_agents_N_rounds_R.json). This approach enables researchers to run multiple experiments with different configurations and automatically organize results by parameters. The naming convention makes it easy to identify which configuration produced which results without requiring separate metadata files. Configuration is passed as command-line arguments or function parameters, with minimal validation.

Loads and preprocesses task-specific datasets in different formats (GSM dataset, MMLU dataset, biography articles in JSON, generated math problems) and normalizes them into consistent input formats for debate generation. Each task domain has custom preprocessing logic that extracts questions, context, and ground truth from domain-specific file formats. The preprocessing layer abstracts format differences, enabling the debate generation pipeline to work with consistent input structures despite underlying dataset heterogeneity.

Constructs specialized debate prompts that guide agents through structured reasoning while optionally assigning distinct roles or perspectives. Prompts are engineered to encourage agents to build on previous responses, challenge assumptions, and refine answers through iterative rounds. The system encodes debate instructions, task context, and prior round responses into prompts that are sent to the LLM. Prompt engineering is task-specific, with different prompt templates for math, word problems, reading comprehension, and factual reasoning to optimize for domain-specific reasoning patterns.

Implements the methodology described in the paper 'Improving Factuality and Reasoning in Language Models through Multiagent Debate' with reproducible experimental setup, standardized datasets, and systematic evaluation. The system provides end-to-end pipelines for generating debate outputs and evaluating them against ground truth, enabling researchers to reproduce paper results and build on the methodology. The implementation includes task-specific configurations that match paper experiments, enabling direct comparison of results.