Line of inquiry
Inquiring lines›How can we train models to be prop…›What architectural and training in…›this line of inquiry
Do fine-tuning methods modify core model capabilities or just outputs?
A broader line of inquiry — a family of 83 specific questions the research asks around this. Follow one into its inquiring-line page, or move sideways to a related line below.
Questions in this line of inquiry 83
Specific inquiring lines the field asks around this — ordered from the most general framing down to the most specific angle.
- How does in-weights adaptation create spurious forgetting in models?
- Does fine-tuning actually change model capabilities or only output distribution?
- How do retrieval and fine-tuning trade off flexibility against training cost?
- Does LoRA's performance scale depend on model architecture or domain?
- Does parameter isolation per task enable online updates without retraining?
- How do finetuning and pretraining improvements differ in their effects on model capabilities?
- Does finetuning facts into weights overwrite existing model capabilities?
- How does sparsity in full-rank updates compare to explicit rank constraints?
- Does sparse parameter updating improve test-time training's computational cost?
- Why does full multi-task fine-tuning perform worse than sequential training?
- How do newly learned facts become accessible after gradient updates?
- What training cost tradeoffs exist between fine-tuning and other knowledge injection methods?
- What makes representation interventions more efficient than weight perturbations for finetuning?
- Why do sparse parameter subsets enable full-rank learning in RL?
- How much can externalized skills improve models before hitting diminishing returns?
- Why does specializing to one task make future task learning harder?
- Why do large language models outperform fine-tuned models once repeated items are removed?
- Can a trained decoder replace both search and parameter updates?
- How does behavioral fine-tuning differ from factual knowledge encoding in models?
- Can memory-based adaptation and gradient fine-tuning operate on complementary timescales?
- When does training a memory model beat RAG or fine-tuning?
- Can expert vectors learned offline transfer across multiple model architectures?
- Why does fine-tuning fail to remove temporal contamination from pretraining?
- Can finetuning sparse subnetworks alone match full parameter finetuning results?
- Why does fine-tuning change how models process retrieved context?
- When should model isolation be preferred over weight-averaging approaches?
- How tight should a textual learning rate be before it prevents skill escape?
- Why does parameter-efficient tuning scaling fail to improve finetuning performance?
- Why does adaptation concentrate in low-dimensional subspaces of weights or representations?
- Does pretraining data size matter less than base model scale for finetuning?
- Can filtering unknown examples during fine-tuning prevent hallucination increases?
- Which finetuning method works best across different task and data regimes?
- How do layer-wise versus parameter-wise merging strategies affect information retention?
- What makes frozen model reasoning different from weight-based parameter updates?
- What happens to base model capabilities when you apply finetuning?
- Why did prior multi-token prediction methods fail during fine-tuning?
- What hidden costs emerge when you fine-tune models for a single domain?
- Where does skill extraction fail compared to genuine model adaptation?
- How do weight perturbations reveal what performance benchmarks cannot measure?
- What performance trade-offs emerge when composing multiple independently trained model capabilities?
- Why do singular value experts compose better than low-rank adapter subspaces?
- How much do cross-model improvements compare to the original performance gains on the training benchmark?
- Why is the fast non-parametric loop vulnerable to overfitting differently than model weights?
- Can models adapt and combine search strategies beyond their training algorithm?
- How do optimal learning rates differ between LoRA and full fine-tuning?
- What are the distinct sources of catastrophic forgetting in sequential fine-tuning?
- Does cross-example gradient contamination explain finetuning-induced hallucination patterns?
- How much performance is lost when converting pretrained checkpoints versus training from scratch?
- Can we unlearn memorized text by finetuning only high-gradient weights?
- How does KL regularization prevent both forgetting and adaptation loss?
- How do trained weights differ from a stored library or text?
- How should skill libraries coordinate with gradient-based weight optimization?
- Does the Assistant Axis exist in pre-trained models before instruction tuning?
- Can dynamic variance weighting replace fixed objective combination weights?
- Does parameter composition work when adapter alignment is imperfect?
- Can temperature-adaptive sampling reduce the sharpening tax during training?
- How do 4-bit weights maintain accuracy during backpropagation through finetuning?
- When should full-parameter post-training be used instead of LoRA adaptation?
- How does task-oriented fine-tuning compare to preference tuning methods?
- What task structures benefit most from geometric parameter merging?
- Can split-softmax or other inference methods fully prevent instruction drift?
- What causes overfitting when forcing new facts into model weights?
- At what dataset size does LoRA performance start degrading?
- What happens to model capability as weight sparsity increases during training?
- How does upstream value embedding differ from downstream alignment patches?
- How does tournament selection without gradients compare to gradient-based hyperparameter tuning?
- What distinguishes fast non-parametric loops from slow parametric weight updates?
- How does pretrained knowledge constrain what adaptation strategies can achieve?
- Why does fine-tuning function as character training rather than capability training?
- What is the accuracy cost of enforcing temporal causality inside model parameters?
- Can per-user adapters remain consistent without drifting or leaking?
- How do cyclic learning rates anti-correlate with weight decay to create diversity?
- Does weight decay directly cause contractive behavior near training examples?
- How do fast skill injection and slow gradient updates work on different timescales?
- Why does LoRA fail when applied only to attention layers?
- Which hyperparameter theories best explain universal behaviors across neural networks?
- How does QLoRA differ from training ternary weights natively from scratch?
- How much does pretraining quality affect the modularity of fine-tuned models?
- Why does placing LoRA adapters at every layer matter for QLoRA performance?
- How does Easy Consistency Tuning accelerate consistency model training from diffusion checkpoints?
- How do aligned LoRA adapters compose through parameter-space arithmetic?
- Can different teams train specialist adapters that share one frozen base model?
- What makes two timescales better than one for minimizing weight movement?