Neura Market
STaR: Bootstrapping Reasoning With Reasoning logo

STaR: Bootstrapping Reasoning With Reasoning

Free

Bootstrapping Language Model Reasoning Through Self-Generated Rationales

FreeFree tier
Inputs: textOutputs: text
Type
Open Source
Visit Website GitHub (40 stars)

About STaR: Bootstrapping Reasoning With Reasoning

STaR (Self-Taught Reasoner) is a technique proposed by researchers from Stanford University and Google AI that enables language models to improve their reasoning capabilities through an iterative bootstrapping process. It starts with a small set of rationale examples and a large dataset without rationales, then generates chain-of-thought rationales for many questions. If the generated answer is incorrect, the model tries again given the correct answer. All rationales that led to correct answers are collected and used to fine-tune the model, and the process repeats. This allows the model to learn from its own generated reasoning and progressively tackle more complex reasoning tasks, achieving performance comparable to fine-tuning a 30x larger state-of-the-art model on CommensenseQA.

Key Features

Iterative bootstrapping loop: generate rationales, fine-tune on correct ones, repeat
Leverages a small set of rationale examples and a large unlabeled dataset
Improves performance on complex reasoning tasks (mathematics, commonsense QA)
Self-taught mechanism: model learns from its own generated reasoning
Comparable to fine-tuning a 30x larger model on CommensenseQA

Pros & Cons

Pros
  • Reduces dependency on massive rationale datasets
  • Enables self-improvement of reasoning through iterative learning
  • Demonstrates significant accuracy gains over direct answer prediction baselines
  • Open source and freely available research
Cons
  • Requires an initial small set of rationale examples to bootstrap
  • May not generalize to all types of reasoning tasks
  • Iterative fine-tuning can be computationally expensive

Best For

Enhancing language model performance on arithmetic and mathematical reasoningImproving commonsense question-answering capabilitiesReducing the need for large manually-annotated rationale datasets

FAQ

What does STaR stand for?
Self-Taught Reasoner.
How does STaR work?
STaR uses a loop: generate chain-of-thought rationales for many questions using a few examples; if the answer is wrong, attempt to generate a rationale given the correct answer; fine-tune the model on all rationales that yielded correct answers; then repeat the process to progressively improve reasoning.
What tasks has STaR been evaluated on?
The paper evaluates STaR on multiple reasoning datasets including mathematics and commonsense question-answering (e.g., CommensenseQA).
Does STaR require large curated rationale datasets?
No, STaR only needs a small set of rationale examples and a large dataset without rationales, bootstrapping the model's ability to generate its own rationales.