Phi-3-medium-4k · CreativityNeuro

A CreativityNeuro (CN) variant of microsoft/Phi-3-medium-4k-instruct, with the weight edit already applied. It loads and runs exactly like the base model.

CreativityNeuro amplifies the parameters that matter for divergent generation but not for convergent generation, improving divergent thinking with no fine-tuning, no prompt changes, and no decoding changes.

📄 Paper · 💻 Code · 🤗 All optimal configs

Configuration

Parameter	Value
Base model	microsoft/Phi-3-medium-4k-instruct
ρ (keep ratio)	0.01
α (amplification)	2.0
Contrastive prompt set	`dat`
Mode	creative

This is the best-performing CreativityNeuro configuration for Phi-3-medium-4k.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("creativityschapiro/phi-3-medium-4k-instruct-cn-dat-kr0.01-a2.0-creative")
tokenizer = AutoTokenizer.from_pretrained("creativityschapiro/phi-3-medium-4k-instruct-cn-dat-kr0.01-a2.0-creative")

outputs = model.generate(...)

Method

Parameter importance is scored Wanda-style, S_ij = Σ_b |W_ij| · ‖X_j‖₂, under two contrastive prompt sets. The top ρ of each is taken, and the set difference — important for divergent generation, not for convergent generation — is amplified:

W_new = W × (1 + α × mask)

To build masks yourself, or apply CN to a model not published here, see samjschapiro/creativityneuro.

Results

Across six instruction-tuned models, CreativityNeuro improves scores on the Divergent Association Task and transfers to open-ended creativity tasks judged by human raters (N = 720) — the Alternative Uses Test and the Task Task — with gains in originality (avg. Cohen's d = +0.36 AUT, +0.40 TT) and surprise (+0.43 AUT). Full results in the paper.

Citation

@inproceedings{schapiro2026creativityneuro,
  title         = {CreativityNeuro: Steering Language Model Weights to Improve
                   Divergent Thinking and Reduce Mode Collapse},
  author        = {Schapiro, Samuel and Park, Core Francisco and Sosa, Felix
                   and Varshney, Lav R.},
  booktitle     = {Conference on Language Modeling (COLM)},
  year          = {2026},
  eprint        = {2607.01433},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2607.01433}
}