Research map / Large language models / Agents, code and math
Bohdi: Heterogeneous LLM Fusion with Automatic Data Exploration
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.
Bohdi fuses heterogeneous source LLMs into a target LLM using only synthetic data, organizing domains in a knowledge tree and adaptively sampling domains via a hierarchical multi-armed bandit with Thompson sampling and a sliding-window binomial likelihood-ratio change detector.
Problem
Existing heterogeneous LLM fusion relies on real data from a few fixed domains with fixed domain proportions, leaving the target LLM with narrow knowledge coverage and capability imbalance across domains.
Approach
Organize knowledge domains as a hierarchical tree that source LLMs expand via collaborative proposal (Sprout), generate QA data with a leader-judged best answer (Harvest), and allocate sampling proportions with a hierarchical MAB using Thompson sampling, plus an Introspection-Rebirth mechanism that resets arm posteriors when a sliding-window binomial likelihood-ratio test detects capability shifts
Key results
- On Llama3.2-3B-Instruct, Bohdi reaches 50.79 average accuracy vs 48.45 for the best baseline (FuseChat-3.0), a 4.12-point gain over base, using 1.7K data.
- On Gemma2-9B-IT, Bohdi averages 58.90 (+4.16 over base), raising TheoremQA from 15.62 to 26.88.
- Bohdi uses about 1.8% of the data of the baselines (1.7K vs 90K-95K).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fuse with Bohdi rather than the original model | Llama3.2-3B-Instruct · AVG across eight benchmarks | accuracy (%): +4.12% | Helped |
| Fuse with Bohdi rather than the original model | Gemma2-9B-IT · AVG across eight benchmarks | accuracy (%): +4.16% | Helped |
| Use Bohdi fusion | Llama3.2-3B-Instruct · All eight reported benchmarks | accuracy (%) | Helped |
| Use Bohdi fusion | Gemma2-9B-IT · TheoremQA | accuracy (%): 15.62 to 26.88 | Helped |
| Add adaptive sampling updates (DynaBranches) (ablation) | Bohdi · HumanEval and BBH | capability performance | Helped |
| Add the IR mechanism to DynaBranches (ablation) | Bohdi · Across benchmarks and fusion rounds | capability growth stability | Helped |
| Replace the Leader Model with a reward model (ablation) | Bohdi · GPQA and MBPP | performance stability | Hurt |
| Set SWBLRT parameters to u=0.1, w=30; u=0.2, w=20; or u=0.3, w=10 (ablation) | Bohdi · Average across benchmarks | accuracy | Helped |
| Use a Meditation-to-Enlightenment sampling ratio of 1:2 rather than 2:1 (ablation) | Bohdi on Llama3.2-3B-Instruct · Average across dimensions | performance | Helped |
| Increase the Meditation-to-Enlightenment sampling ratio to 2:1 (ablation) | Bohdi on Llama3.2-3B-Instruct · Overall across benchmarks | capability improvement | Hurt |
| Increase the number of source models from one to three (ablation) | Bohdi on Llama3.2-3B-Instruct · Across capabilities | performance | Helped |
| Use three smaller source models for Gemma2-9B-IT fusion | Gemma2-9B-IT · AVG across eight benchmarks | accuracy (%): +1.47% | Helped |
Surprising findings
- Baselines SFT and FuseChat-3.0 gain most on MMLU and MATH for Llama3.2-3B but drop by over 9 points on TheoremQA and BBH, showing capability imbalance.
- Weak-to-strong setting: Gemma2-9B-IT fused with smaller source models (2-3B) still improves average by 1.47 points.
Limitations
- Relies on source-model leader judgments and proposals, so quality of synthetic data depends on source LLM quality.
- Reward model (ArmoRM) substitution showed more fluctuation, indicating limited cross-domain generalization of reward models.
- Some baseline comparisons show Bohdi trailing on individual benchmarks (e.g., MMLU vs SFT on Llama3.2-3B).
Experimental setup
Datasets: FuseChat-Mixture
Benchmarks: MMLU GPQA GSM8K MATH HumanEval MBPP TheoremQA BBH
Metrics: accuracy
Baselines: FuseChat FuseChat-3.0 SFT Condor
Builds on: FuseChat FuseChat-3.0 Condor Thompson Sampling Wilks' theorem
Closest papers by meaning
- Knowledge Fusion of Large Language Models (ICLR 2024 · similarity 0.91)
- Weighted-Reward Preference Optimization for Implicit Model Fusion (ICLR 2025 · similarity 0.88)
- Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.88)
- Self-MoE: Towards Compositional Large Language Models with Self-Specialized Experts (ICLR 2025 · similarity 0.88)
- Knowledge Card: Filling LLMs' Knowledge Gaps with Plug-in Specialized Language Models (ICLR 2024 · similarity 0.87)
- Prismatic Synthesis: Gradient-based Data Diversification Boosts Generalization in LLM Reasoning (NeurIPS 2025 · 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.