Research map / Large language models / Reasoning and chain of thought

InfiGFusion: Graph-on-Logits Distillation via Efficient Gromov-Wasserstein for Model Fusion

NeurIPS 2025 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling · 0 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 InfiGFusion, which fuses several source LLMs into a Phi-4 pivot using a Graph-on-Logits Distillation loss that aligns top-k logit co-activation graphs via a sorting-based approximation of Gromov-Wasserstein distance.

Problem

Logit-based model fusion treats vocabulary dimensions independently and ignores semantic dependencies among logits; exact Gromov-Wasserstein alignment of logit graphs is O(n^4) and impractical at LLM vocabulary sizes.

Approach

Build top-k logit feature-wise graphs (inner-product co-activation matrices over sequence positions), compress each graph to degree features, and align them by sorting, giving an O(n log n) closed-form GW approximation used as a distillation loss alongside ULD and SFT.

Key results

• InfiGFusion achieves 83.85 average over 11 benchmarks vs 81.36 for Pivot-SFT (Phi-4) and 82.20 for InfiFusion.
• On the 18-task complex reasoning subset, average rises from 77.94 (Phi4 SFT) to 84.16 (table) / 83.79 (text), with +35.6 on Multistep Arithmetic and +37.06 on Causal Judgement vs the SFT model.
• Approximate GW error relative to exact GW on synthetic Gaussians is about 0.39-0.59 across n=50 to 3000, while running in ~0.3 s at n=3000 versus ~53 s for exact.

Measured effects reported in the paper

Surprising findings

• The approximate GW has relative error of roughly 39-59% against exact GW in the authors' own synthetic experiment, yet is used as a training loss.
• Reported main-table averages disagree across sections (84.16 in Table 2 vs 83.79 in text; +35.6 Multistep Arithmetic claimed as +34.8 in text).

Limitations

• Gains are marginal on factoid QA and literal token-matching tasks where a few logit dimensions dominate.
• No explicit mechanism for resolving factual conflicts among source models.
• Approximation error bound is loose and relative error against exact GW is large in the authors' experiment; no statistical significance tests reported.

Experimental setup

Datasets: Infinity-Instruct NuminaMath-1.5 KodCode-V1-SFT-R1 AM-DeepSeek-R1-Distilled-1.4M

Benchmarks: GSM8K MATH TheoremQA MBPP HumanEval BBH ARC-C MMLU IFEval DROP HellaSwag BIG-Bench Hard subtasks (84 tasks in appendix)

Metrics: accuracy average accuracy

Baselines: MiniLogit FuseLLM FuseChat InfiFusion Pivot-SFT (Phi-4) Qwen2.5-Instruct Mistral-Small Qwen2.5-Coder

Builds on: ULD FuseLLM Gromov-Wasserstein distance Sinkhorn

Closest papers by meaning

• Beyond Parameter Arithmetic: Sparse Complementary Fusion for Distribution-Aware Model Merging (NeurIPS 2026 · similarity 0.89)
• Knowledge Fusion of Large Language Models (ICLR 2024 · similarity 0.86)
• Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.86)
• Entropy-Gated Latent Recursion (NeurIPS 2026 · similarity 0.86)
• Prismatic Synthesis: Gradient-based Data Diversification Boosts Generalization in LLM Reasoning (NeurIPS 2025 · similarity 0.86)
• Metastable Dynamics of Chain-of-Thought Reasoning: Provable Benefits of Search, RL and Distillation (ICML 2025 · similarity 0.86)