Research map / Large language models / Retrieval and knowledge

Empowering Self-Learning of LLMs: Inner Knowledge Explicitation as a Catalyst

AAAI 2025 · Large language models › Retrieval and knowledge · approach: Knowledge editing in LLMs · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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 SKE-Learn, which has an LLM explicitly extract inner knowledge before reasoning and uses GPT-4-verified knowledge and self-assessment to filter self-synthesized data for iterative self-learning on Llama3-8B.

Problem

Self-learning LLMs amplify hallucinations in self-synthesized training data during iterative self-training, so reliable data selection is needed; the paper also asks whether explicitly extracting inner knowledge improves reasoning.

Approach

Inner Knowledge Explicitation has the model first output explicit natural-language knowledge from its memory and then reason with it; because such knowledge is verifiable against real-world reference chunks, the model's self-assessment can filter self-synthesized "question-knowledge-reasoning" traces across iterative self-training rounds.

Key results

• Average accuracy across six benchmarks rises from 52.79% (SKE-Learn M0 prompting) to 56.54% after four self-training rounds (M4), versus 49.88% for the self-learning baseline.
• SKE-Learn prompting (M0) averages 52.79%, outperforming the strongest prompting baseline Thread-of-Thought (51.36%).
• Meta-skill training model M1 reaches 54.35% average, a 4.47-point gain over the self-learning baseline while using about 90% less training data.

Measured effects reported in the paper

Surprising findings

• Zero-shot prompting outperforms several prompting techniques such as Chain-of-Thought on some tasks, attributed to the Magpie-trained model's long detailed answer style.
• Ablations that fine-tune on external scoring data, general data, or continued pre-training on the knowledge corpus perform worse than the first meta-skill round, suggesting gains are not from knowledge injection.

Limitations

• Relies on GPT-4 for scoring in the meta-skill training stage, so the pipeline is not fully self-contained.
• Self-assessment and verification depend on real-world knowledge chunks as references, limiting applicability to domains without a verifiable corpus.
• Evaluated on a single 8B backbone (Llama3-8B-Magpie) with temperature 0 decoding.

Experimental setup

Datasets: Magpie-Pro-300K-Filtered Wikipedia-based STEM knowledge corpus (wiki_stem) swti-stem-20k

Benchmarks: MMLU BBH ARC-E ARC-C AGIEval Natural Questions

Metrics: accuracy

Baselines: Zero-shot Chain-of-Thought Plan-and-Solve Rephrase-and-Respond System-2-Attention Thread-of-Thought Analogical Re-reading Self-learning Baseline

Builds on: Self-Rewarding language models Chain-of-Thought Magpie

Closest papers by meaning

• Toward Self-Improvement of LLMs via Imagination, Searching, and Criticizing (NeurIPS 2024 · similarity 0.88)
• Hypothesis, Verification, and Induction: Grounding Large Language Models with Self-Driven Skill Learning (AAAI 2024 · similarity 0.88)
• SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures (NeurIPS 2024 · similarity 0.88)
• Can LLMs Correct Themselves? A Benchmark of Self-Correction in LLMs (NeurIPS 2025 · similarity 0.88)
• S^3cMath: Spontaneous Step-Level Self-Correction Makes Large Language Models Better Mathematical Reasoners (AAAI 2025 · similarity 0.87)
• ReGenesis: LLMs can Grow into Reasoning Generalists via Self-Improvement (ICLR 2025 · similarity 0.87)