Research map / Probabilistic and scientific modeling / Treatment effects
Spiked-CFR: Causal Representation Learning from LLMs via Wasserstein Projection Pursuit
Fan Wang, Hengyu Yue, Yu Bowen, Weiming Liu, Zongxin Yang, Xuyun Zhang, Xiaolin Zheng, Chaochao Chen, Shuiguang Deng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace full-space Wasserstein balancing with projection-pursuit balancing | Llama-CFR-Wass · ACIC-post | epsilon_PEHE: 2.32 -> 2.18 | Improved |
| replace full-space Wasserstein balancing with projection-pursuit balancing | Llama-CFR-Wass · IHDP-post | epsilon_PEHE: 2.65 -> 2.47 | Improved |
| replace full-space Wasserstein balancing with projection-pursuit balancing | Llama-CFR-Wass · Retail Hero | RMSE(ATE): 2.15 -> 0.72 | Improved |
| replace full-space Wasserstein balancing with projection-pursuit balancing | Llama-CFR-Wass · Sema vs. Lira (clinical) | RMSE(ATE): 6.17 -> 0.56 | Improved |
| replace full-space Wasserstein balancing with projection-pursuit balancing | Llama-CFR-Wass · Onab vs. Topi (clinical) | RMSE(ATE): 8.14 -> 1.38 | Improved |
| use learned WPP subspace instead of random projection or PCA of the same dimension k (ablation) | SPIKED-CFR · semi-synthetic and real-world datasets | relative error | Improved |
| learn projection U with WPP rather than fixed projection (ablation) | SPIKED-CFR · ACIC, Eren vs. Topi | seed variance | Improved |
| extract features from the L-4 layer rather than other layers or last-token hidden state (ablation) | SPIKED-CFR · semi-synthetic and clinical | normalized error | Improved |
| use Llama-3-8B embeddings instead of BERT features | CFR/forest estimators · all benchmarks | effect error | Improved |
| full-space Wasserstein instead of MMD in CFR on Llama embeddings | Llama-CFR · IHDP-post | epsilon_PEHE: 2.51 -> 2.65 | Hurt |
| apply SPIKED-CFR balancing to raw LLM embeddings | LLM embeddings · ACIC-post | treatment classification AUC | Improved |
| add NATURAL-style LLM probability IPW instead of SPIKED-CFR | NATURAL · Eren vs. Topi (clinical) | RMSE(ATE): 1.68 -> 3.99 | Hurt |
Closest papers by meaning
- R-Sparse: Rank-Aware Activation Sparsity for Efficient LLM Inference (ICLR 2025 · similarity 0.85)
- A Sparsity Principle for Partially Observable Causal Representation Learning (ICML 2024 · similarity 0.85)
- Dynamic Sparse No Training: Training-Free Fine-tuning for Sparse LLMs (ICLR 2024 · similarity 0.84)
- Identifying Representations for Intervention Extrapolation (ICLR 2024 · similarity 0.84)
- NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models (ICLR 2025 · similarity 0.84)
- Leveraging semantic similarity for experimentation with AI-generated treatments (NeurIPS 2025 · similarity 0.84)
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.
