Extract potential search queries from the input text

Analyze the quality of an IMO solution based on mathematical rigor criteria

Analyze the quality of a mathematical proof

Extract reasoning strategies from all agent solutions

Apply final polishing to the research report

Find a strategy of the same problem type with examples similar to the query. Args: problem_type: The problem typ

Attempt to fix an incomplete report by removing problematic sections and ensuring a coherent final document. This is

Format the prompt according to model's chat template

Format the prompt according to model's chat template

Format the DeepConf response with optional statistics. Args: answer: The final answer from weighted voting stats: Proces

Format a question for the benchmark dataset

Format search results into readable text

Create ServerConfig from a dictionary

Generates a completion based on the provided system prompt and task. Parameters: system_prompt (str): The system prompt to guide the

Generate the dataset and save it to a JSONL file.

Generate the dataset and save it to a JSONL file.

Ask LLM to fix the broken code.

Generate JSON based on the provided schema and prompt.

Generates n completions for the Best of N step of CePO. Parameters: system_prompt (str): The system prompt to guide the model.

Generate the preliminary draft (updatable skeleton) from LLM internal knowledge This serves as the initial state for the diffusion pr

Generate a comprehensive report comparing all approaches.

Generate a response using the specified approach.

Generate a response using the specified approach.

Generate a solution for the given problem using reasoning API

Async wrapper for agent solution generation

Generate a new problem-solving strategy using the LLM. Args: problem: The problem that needs a strategy problem_type: Th

Generate response using uncertainty-routed chain-of-thought. Args: prompt: The prompt to generate responses for

Return all 39 reasoning modules.

Get or create singleton AnalyzerEngine instance.

Get or create singleton AnonymizerEngine instance.

Get configuration for specific approach

Calculate average confidence across all tokens in the trace. Returns: Average confidence value

Get the best solution based on verification score and confidence

Calculate average confidence of bottom 10% groups. Returns: Bottom 10% group confidence

Get cached response with fuzzy matching

Get the complexity label and confidence score. Args: text: The query text Returns:

Get information about the DeepConf implementation. Returns: Dictionary with implementation details

Get the appropriate device (mps, cuda, or cpu).

Get reasoning effort profile based on specified level and max tokens. Args: reasoning_effort: 'low', 'medium', or 'high' max_

Get response from the LLM for a given problem. Args: problem (str): The problem text model (str): The model identifier

Get response from the LLM for a mathematical problem

Get response from the LLM for an IMO problem with extended timeout for complex reasoning

Get response from the LLM for a proof problem

Get response from the LLM for a given problem. If multiple choices are returned, formats them as attempt dictionaries. Args:

Get the minimum confidence across all groups. Returns: Lowest group confidence

Get just the descriptions for prompting.

Generate a unique ID for a new strategy.

Calculate optimal temperature based on prompt characteristics

Get optimized generation config

Get an existing session or create a new one for the given session ID.

Get or load adapter with enhanced caching.

Get inference pipeline - automatically chooses MLX or PyTorch based on model

Get or create a browser session for the given session ID.

Get the steering strength for a specific pattern. Args: pattern: The reasoning pattern Retu

Get the most appropriate steering vector for a context. Args: context: The current generation context.

Get summary of strategy network insights

Get a summary of the workspace state

Prepare input data for solution synthesis

Get token budget based on complexity. Args: complexity: Complexity label (HIGH or LOW) Retu

Grade response using SimpleQA methodology

Two-stage verification system from IMO25 repository: Stage 1: Detailed verification using comprehensive IMO grader prompt Stage 2: Simple

Improve a solution based on verification feedback

Increment the total query count.

Install steering hooks on a model. Args: model: The model to install hooks on manager: The steering vector manager

Check if response matches any of the target answers.

Check if the response is just a numeric value, possibly with whitespace and newlines. Args: response: The response text to check

Check if adding this token creates a thought switch sequence.

Check if adding this token creates a thought switch sequence. Args: token: Token ID to check

Run iterative improvement on solutions that failed verification with parallel execution

Limit the number of strategies per problem type to the specified maximum in the database. This controls storage limit, not the number

Call the LLM with retries on transient errors. Makes a chat completion request to the given client and extracts the response. Retries up

Load the 2024 dataset of problems. Returns: list[dict]: The dataset of problems.

Load the 2025 dataset of problems from math-ai/aime25. Returns: list[dict]: The dataset of problems.

Load a LoRA adapter with enhanced caching

Load steering vectors from the HuggingFace dataset.

Load dataset by year (2024 or 2025). Returns: list[dict]: The dataset of problems.