Research map / Large language models / Reasoning and chain of thought
ReGenesis: LLMs can Grow into Reasoning Generalists via Self-Improvement
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 ReGenesis, a self-improvement method that synthesizes reasoning paths from general reasoning guidelines via task-specific guidelines and reasoning structures, then fine-tunes the same LLM on filtered paths to improve in-domain and out-of-domain reasoning.
Problem
Self-synthesized reasoning-path training (e.g., STaR, LMSI) improves in-domain reasoning but degrades out-of-domain generalization, because the generated paths are task-specific and lack task-agnostic reasoning diversity.
Approach
Generate reasoning paths by progressing from abstract to concrete: adapt 25 general seed reasoning guidelines to each task, derive reasoning structures without solving, then produce solutions; filter with ground truth and add ground truth as a hint only for instructions lacking correct paths.
Key results
- ReGenesis achieves average in-domain improvement of 16.56% over the non-fine-tuned Mistral-7B-Instruct-v0.3, outperforming all baselines.
- On six OOD tasks, ReGenesis yields about 6.1% average improvement while STaR and LMSI-GT decline on average (about 4.6% drop for existing methods).
- On NumGLUE with Meta-Llama-3-8B-Instruct, ReGenesis raises in-domain accuracy from 58.9% to 79.2% and improves six OOD tasks by 6.46% on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune with ReGenesis-generated reasoning paths rather than no further fine-tuning | Mistral-7B-Instruct-v0.3 · OOD tasks (six-task average) | accuracy: 6.1% improvement | Helped |
| Fine-tune with existing self-improvement methods rather than no further fine-tuning | Mistral-7B-Instruct-v0.3 · OOD tasks (six-task average) | accuracy: 4.6% | Hurt |
| Use reasoning paths with ground-truth filtering instead of majority-vote filtering | LMSI · In-domain tasks (five-task average) | accuracy: 1.02% to 6.26% | Helped |
| Fine-tune with ground-truth answers only, without reasoning paths | Mistral-7B-Instruct-v0.3 · GSM8K test | accuracy: 44.8% | Hurt |
| Fine-tune with ground-truth answers only, without reasoning paths | Mistral-7B-Instruct-v0.3 · ARC-c test | accuracy: 3.4% | Hurt |
| Add Reasoning Structure Generation (ablation) | ReGenesis · GSM8K test | accuracy: $66.2\%$ -> $76.6\%$ | Helped |
| Add Guidance Adaptation (ablation) | ReGenesis · GSM8K test | accuracy: $70.0\%$ -> $76.6\%$ | Helped |
| Add Reasoning Structure Generation (ablation) | ReGenesis · ASDIV | accuracy: $76.1\%$ -> $81.2\%$ | Helped |
| Add Guidance Adaptation (ablation) | ReGenesis · ASDIV | accuracy: $77.5\%$ -> $81.2\%$ | Helped |
| Use majority-vote filtering with ReGenesis instead of LMSI (ablation) | Mistral-7B-Instruct-v0.3 · GSM8K | accuracy: $62.3\%$ -> $62.8\%$ | Helped |
| Use majority-vote filtering with ReGenesis instead of LMSI (ablation) | Mistral-7B-Instruct-v0.3 · StrategyQA | accuracy: $79.0\%$ -> $83.1\%$ | Helped |
| Add ground-truth hints at every generation step, including Reasoning Path Generation (ablation) | ReGenesis on GSM8K · GSM8K test | accuracy: $75.7\%$ -> $73.7\%$ | Hurt |
Surprising findings
- Fine-tuning on ground-truth answers only (FT w/ GT) causes a 44.8% drop on GSM8K in-domain performance.
- Existing self-improvement methods (STaR, LMSI with GT) generally degrade OOD performance after fine-tuning.
Limitations
- Evaluated only on 7-8B models (Mistral-7B-Instruct-v0.3 and Llama-3-8B-Instruct) and a limited set of tasks.
- Filtered-out reasoning paths are not exploited; future work needed on using them.
- Some OOD results are mixed (e.g., ReGenesis drops slightly on GSM8K-trained ASDIV and ReClor-trained StrategyQA-related results).
Experimental setup
Datasets: GSM8K NumGLUE ReClor ARC-c StrategyQA ASDIV SVAMP AQUA-RAT BBH ANLI OpenbookQA
Benchmarks: GSM8K NumGLUE ARC-c ReClor StrategyQA ASDIV SVAMP AQUA-RAT BBH ANLI OpenbookQA
Metrics: zero-shot exact match accuracy
Baselines: w/o FT FT w/ GT LMSI LMSI w/ GT STaR
Builds on: STaR LMSI Chain-of-thought prompting Self-consistency
Closest papers by meaning
- Toward Self-Improvement of LLMs via Imagination, Searching, and Criticizing (NeurIPS 2024 · similarity 0.90)
- Can LLMs Correct Themselves? A Benchmark of Self-Correction in LLMs (NeurIPS 2025 · similarity 0.90)
- GPO: Learning from Critical Steps to Improve LLM Reasoning (NeurIPS 2025 · similarity 0.90)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.90)
- General-Reasoner: Advancing LLM Reasoning Across All Domains (NeurIPS 2025 · similarity 0.90)
- Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.90)
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.