Skip to content
 
 

Repository files navigation

UnpaidRA

Give it a research field, get back what took your RA three weeks.

Powered by Cheap Shot, a confidence-based prospective routing framework grounded in CP-Router (Su et al., 2025) and the Confidence-Driven LLM Router (Zhang & Dimitriadis, 2025).


What it does

UnpaidRA takes a research field as input and returns:

  • Recent papers fetched via Semantic Scholar
  • Structured summaries of each paper
  • Identified gaps in the literature
  • Novel research directions with high-level experiment sketches

The routing layer (Cheap Shot) assigns models before any agent runs using a single-token logprob probe:

  • granite4:micro for simple tasks
  • claude-haiku-4-5 for structured reasoning
  • claude-sonnet-4-6 for complex synthesis

Research grounding

  • CP-Router (Su et al., 2025) — Introduces logprob-based routing between standard LLMs and Large Reasoning Models.
  • Confidence-Driven LLM Router (Zhang & Dimitriadis, USC/Amazon, 2025) — Uses semantic entropy as an uncertainty signal for routing decisions.
  • OI-MAS (arXiv 2601.04861, Jan 2026) — Proposes a confidence-aware conductor architecture for multi-agent model selection.
  • RouteLLM (Ong et al., ICLR 2025) — Uses preference data to train routers and achieves substantial cost reductions while maintaining quality.
  • Self-Evaluating LLMs for Multi-Step Tasks (arXiv 2511.07364) — Evaluates verbalized confidence estimates for multi-step reasoning tasks.
  • Dynamic Model Routing and Cascading Survey (arXiv 2603.04445) — Surveys routing methods and finds probe-based approaches outperform verbalized confidence.
  • Semantic Uncertainty (Kuhn et al., ICLR 2023) — Introduces semantic entropy as a principled uncertainty measure for natural language generation.

Results

Metric Value
Cost Reduction vs All-Sonnet Baseline 39.1%
Actual Cost per Run $0.014
Total Tokens 7,728
Evaluation Status 3/3 evals passing

Routing Distribution

Model Count
granite4:micro ×1
claude-haiku-4-5 ×5
claude-sonnet-4-6 ×2

Setup

uv sync
ollama pull granite4:micro
cp .env.example .env

Add the following variables to .env:

ANTHROPIC_API_KEY=your_key_here
SEMANTIC_SCHOLAR_KEY=your_key_here

Run pipeline

uv run python projects/unpaid-ra/agent.py "your research field"

Run UI

uv run python projects/unpaid-ra/ui/app.py

Then open:

http://localhost:5000

Run evals

uv run python projects/unpaid-ra/evals/run_evals.py

Lane

Lane 3 — Most Innovative Multi-Agent System

TOA Agent Build Day, June 27 2026


License

MIT

About

Starter agents + skills for TOA Agent Build Day (June 27, 2026) — 9 fork-ready projects, agentskills.io-spec skills, measured evals, local-first

Resources

Contributing

Stars

1 star

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages