REVIEW 20 cited by
Dataless Knowledge Fusion by Merging Weights of Language Models
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
Dataless Knowledge Fusion by Merging Weights of Language Models
read the original abstract
Fine-tuning pre-trained language models has become the prevalent paradigm for building downstream NLP models. Oftentimes fine-tuned models are readily available but their training data is not, due to data privacy or intellectual property concerns. This creates a barrier to fusing knowledge across individual models to yield a better single model. In this paper, we study the problem of merging individual models built on different training data sets to obtain a single model that performs well both across all data set domains and can generalize on out-of-domain data. We propose a dataless knowledge fusion method that merges models in their parameter space, guided by weights that minimize prediction differences between the merged model and the individual models. Over a battery of evaluation settings, we show that the proposed method significantly outperforms baselines such as Fisher-weighted averaging or model ensembling. Further, we find that our method is a promising alternative to multi-task learning that can preserve or sometimes improve over the individual models without access to the training data. Finally, model merging is more efficient than training a multi-task model, thus making it applicable to a wider set of scenarios.
Forward citations
Cited by 20 Pith papers
-
Model Merging as Probabilistic Inference in Fine-Tuning Parameter Space
Model merging is cast as PoE inference with EBM experts, revealing Gaussian assumptions in prior work and proposing convergent Cauchy experts that improve empirical performance.
-
Predicting Mergeability of Parameter-Efficient Fine-Tuning Updates
MergeProbe forecasts LoRA adapter mergeability from first-few-percent training signals and outperforms interference-aware baselines on retention while adding low overhead on a five-domain benchmark.
-
Distributionally Robust Multi-Task Reinforcement Learning via Adaptive Task Sampling
DRATS derives a minimax objective from a feasibility formulation of MTRL to adaptively sample tasks with the largest return gaps, leading to better worst-task performance on MetaWorld benchmarks.
-
Generalizing the Geometry of Model Merging Through Frechet Averages
Model merging is generalized as Fréchet averaging on symmetry-invariant manifolds, containing Fisher merging as a special case and offering a new approach for LoRA adapters.
-
Generalizing the Geometry of Model Merging Through Frechet Averages
Model merging is reframed as Fréchet averaging on manifolds whose geometry respects architectural symmetries, generalizing Fisher merging and enabling better LoRA merges.
-
Differentially Private Model Merging
Post-processing via random selection or linear combination generates differentially private models for arbitrary privacy parameters from pre-trained models on the same dataset.
-
DA-MergeLoRA: Hypernetwork-Based LoRA Merging for Few-Shot Test-Time Domain Adaptation
A hypernetwork generates per-column merging weights to combine source LoRA modules on CLIP, achieving state-of-the-art few-shot test-time domain adaptation.
-
When Model Merging Breaks Routing: Training-Free Calibration for MoE
Merging breaks MoE routing via softmax sensitivity; HARC uses Hessian curvature for closed-form router calibration that improves merged model performance without retraining.
-
Dynamic Model Merging Made Slim
DiDi-Merging achieves dynamic model merging performance matching or exceeding prior methods while using only 1.24x to 1.4x the parameters of a single fine-tuned model.
-
Breaking Lock-In: Preserving Steerability under Low-Data VLA Post-Training
DeLock mitigates lock-in in low-data VLA post-training via visual grounding preservation and test-time contrastive prompt guidance, outperforming baselines across eight evaluations while matching data-heavy generalist...
-
Analytic Drift Resister for Non-Exemplar Continual Graph Learning
ADR achieves theoretically zero-forgetting class-incremental graph learning by combining backpropagation adaptation with ridge-regression-based layer-wise merging of GNN linear transformations.
-
ACE-Merging: Data-Free Model Merging with Adaptive Covariance Estimation
ACE-Merging estimates task input covariances from parameter differences to enable closed-form data-free merging that reduces interference and outperforms prior baselines on vision and language tasks.
-
Model Merging Scaling Laws in Large Language Models
Empirical scaling laws for LLM merging show a size-dependent floor and 1/k-like tail in cross-entropy loss that holds across architectures and merging methods.
-
DivMerge: A divergence-based model merging method for multi-tasking
DivMerge learns task-arithmetic merging weights by minimizing Jensen-Shannon divergence between each specialist model and the merged model, improving multi-task performance and scalability.
-
Rethinking Heterogeneous LLM Merging: A Weighted Model Averaging Perspective
After truncating or expanding checkpoints to a shared shape, small-ratio weight averaging slightly improves average benchmark scores over strong Qwen sources, but headline gains are inflated by per-task best-ratio selection.
-
DeltaMerge-LowRes: Composing Language and Task Deltas for Low-Resource Adaptation
A TIES-style merge across language-delta and task-delta axes (cross-axis TIES) beats additive and task-only composition on low-resource summarisation and QA.
-
ReLoRA: Knowledge-Reusing Adaptation for Fast Rollout of Evolving LLM Services
ReLoRA reduces time-to-readiness for LoRA adapters on updated LLMs by up to 8.9x through adaptive Bayesian initialization and scheduled regularization while improving accuracy by up to 4.6%.
-
Differentially Private Model Merging
Post-processing via random selection or linear combination of differentially private models allows meeting arbitrary target privacy parameters without additional training.
-
MAny: Merge Anything for Multimodal Continual Instruction Tuning
MAny addresses dual-forgetting in multimodal continual instruction tuning via CPM and LPM merging strategies, delivering up to 8.57% accuracy gains on UCIT benchmarks without additional training.
-
Expert Merging in Sparse Mixture of Experts with Nash Bargaining
Expert merging in MoE models can be improved by setting per-expert weights with the Nash bargaining solution instead of simple averaging.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.