“Recursive Self-Aggregation Unlocks Deep Thinking in Large Language Models.” implemented with DSPy.

Setup

Set your API key (required for OpenAI models):

export OPENAI_API_KEY="..."

Usage

No-install (using uvx from GitHub):

uvx --from git+https://github.com/eaubin/recselfagg recselfagg \
  --population 3 \
  --subset 2
  --steps 3
  --show-completions \
  "Argue for the best fictional character."

CLI Flags

• --model: DSPy model id (default: openai/gpt-4o-mini)
• --population: population size N
• --subset: subset size K sampled each aggregation
• --steps: number of RSA steps
• --temperature: sampling temperature
• --population-temperature: temperature for population sampling (defaults to --temperature)
• --aggregate-temperature: temperature for aggregation steps (defaults to --temperature)
• --max-tokens: per-call token limit
• --show-completions: print every completion at each step
• --seed: random seed for reproducibility
• --rollout-base: base rollout id (defaults to random)
• --no-progress: disable progress logging
• --debug: print rollout ids and debug info for each call
• --trace-json: write JSON trace of populations at each step to a file
• --final-population-file: write final population (JSON) to a file
• --parallel: number of concurrent model calls (default: 1)

Output

A random member of the final population is printed.