Research map / Large language models / Reasoning and chain of thought
K-STaR: Knowledge-Aware Self-Taught Reasoner
Guozheng Li, Xinyu Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes K-STaR, a self-training method for LLM reasoning that uses knowledge-aware signals beyond answer correctness to filter noisy reasoning paths.
Problem
Self-training LLMs on self-generated reasoning paths typically uses answer correctness as supervision, which treats correct-answer samples with flawed intermediate steps as high quality, producing noisy training signals.
Approach
Replaces purely correctness-based selection of self-generated reasoning paths with a knowledge-aware criterion intended to assess the quality of intermediate reasoning steps, not just the final answer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add knowledge-based process reward on top of outcome filtering (ablation) | K-STaR (Llama-3-8B) · Average of four benchmarks | accuracy: 66.7% -> 72.3% | Improved |
| add outcome (answer correctness) filter on top of process reward (ablation) | K-STaR (Llama-3-8B) · GSM8K | accuracy: 52.6% -> 62.7% | Improved |
| add process reward to outcome filter on commonsense QA (ablation) | K-STaR (Llama-3-8B) · CommonsenseQA | accuracy: 72.3% -> 78.6% | Improved |
| K-STaR knowledge-aware filtering instead of ReSTEM w/ PRM | Mistral-7B-Instruct · Average of four benchmarks | accuracy: 67.1% -> 70.4% | Improved |
| K-STaR instead of ReSTEM w/ PRM | Llama-3-8B-Instruct · Average of four benchmarks | accuracy: 69.8% -> 72.3% | Improved |
| supervised fine-tuning on ground-truth answers only | Mistral-7B-Instruct · GSM8K | accuracy: 42.3% -> 11.8% | Hurt |
| cross-domain transfer after self-training with STaR | Llama-3-8B-Instruct · GSM8K to CommonsenseQA | accuracy: 69.0% -> 65.1% | Hurt |
| cross-domain transfer after self-training with K-STaR | Llama-3-8B-Instruct · cross-domain transfers | accuracy: 69.0% -> 70.2% on GSM8K to CSQA; 78.0% -> 77.7% on NumGLUE to StrategyQA | Mixed |
| increase knowledge sets n from 1 to 3 (m=10) (ablation) | K-STaR (Llama-3-8B) · Average of four benchmarks | accuracy: 69.1% -> 72.1% | Improved |
| increase n from 5 to 7 or m from 10 to 15 (ablation) | K-STaR (Llama-3-8B) · Average of four benchmarks | accuracy: 72.3% -> 72.2% (n=7,m=10) | No clear effect |
| run more self-training iterations (R3 to R5) (ablation) | K-STaR (Llama-3-8B) · CommonsenseQA | accuracy: 78.1 -> 78.6 | No clear effect |
| first self-training iteration (ablation) | K-STaR (Llama-3-8B) · CommonsenseQA | accuracy: 69.0 -> 73.2 | Improved |
Experimental setup
Baselines: STaR
Builds on: STaR
Closest papers by meaning
- Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models (ICML 2026 · similarity 0.88)
- ReGenesis: LLMs can Grow into Reasoning Generalists via Self-Improvement (ICLR 2025 · similarity 0.88)
- Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.88)
- B-STaR: Monitoring and Balancing Exploration and Exploitation in Self-Taught Reasoners (ICLR 2025 · similarity 0.88)
- Native Reasoning Models: Training Language Models to Reason on Unverifiable Data (ICLR 2026 · similarity 0.87)
- Verifying Meta-Awareness via Predictive Rewards in Reasoning Models (ICML 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
