Research map / Architectures and efficiency / RNNs and normalization
Negative Pre-activations Differentiate Syntax
Linghao Kong, Angelina Ning, Micah Adler, Nir N Shavit
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 |
|---|---|---|---|
| zero only negative pre-activations of top 1% Wasserstein neurons per layer (ablation) | Llama 3.1 8B · TSE | accuracy: 84.4 -> 55.2 | Hurt |
| zero only negative pre-activations of top 1% Wasserstein neurons per layer (ablation) | Llama 3.1 8B · BLiMP | accuracy: 81.2 -> 77.5 | Hurt |
| zero negative pre-activations of top 1% WD neurons in Qwen3 8B Base (ablation) | Qwen3 8B Base · TSE | accuracy: 84.8 -> 78.6 | Hurt |
| perplexity-matched ablation of many low-WD neurons instead of top Wasserstein neurons (ablation) | Llama 3.1 8B · BLiMP and TSE | accuracy: BLiMP 81.2 -> 81.5; TSE 84.4 -> 85.9 | No clear effect |
| ablate 1% random neurons' negative pre-activations (ablation) | Llama 3.1 8B · BLiMP and TSE | accuracy: TSE 84.4 -> 84.1 | No clear effect |
| perplexity-matched low-WD ablation vs Wasserstein ablation on non-grammatical benchmarks (ablation) | Llama 3.1 8B · ARC, BoolQ, HellaSwag, PIQA, SciQ, WinoGrande, TruthfulQA | accuracy: approx 11% vs approx 4% absolute error increase | Hurt |
| Wasserstein ablation on TruthfulQA (outlier) (ablation) | Llama 3.1 8B · TruthfulQA | accuracy: 45.2 -> 39.8 vs perplexity-matched 42.9 | Hurt |
| ablate early layers (first four) Wasserstein neurons (ablation) | Llama 3.1 8B · TSE NPI licensing | error: 20% increase in error | Hurt |
| cumulative layer-group ablations of Wasserstein neurons (ablation) | Llama 3.1 8B · BLiMP and TSE | error | Hurt |
| replace SiLU positive branch with ReLU in top-WD neurons (ablation) | Llama 3.1 8B · WikiText 2 validation | perplexity | No clear effect |
| flip sign of negative post-activations of top 2% WD neurons (ablation) | Llama 3.1 8B · WikiText 2 validation | perplexity: 6.50 -> 2277 | Hurt |
| select neurons by MD instead of WD (ablation) | Llama 3.1 8B · BLiMP and TSE | accuracy | No clear effect |
Closest papers by meaning
- Do LLMs Signal When They’re Right? Evidence from Neuron Agreement (ICML 2026 · similarity 0.86)
- Sparsing Law: Towards Large Language Models with Greater Activation Sparsity (ICML 2025 · similarity 0.86)
- A Polar coordinate system represents syntax in large language models (NeurIPS 2024 · similarity 0.85)
- Task-Specific Data Selection for Instruction Tuning via Monosemantic Neuronal Activations (NeurIPS 2025 · similarity 0.85)
- Hierarchical Frequency Tagging Probe (HFTP): A Unified Approach to Investigate Syntactic Structure Representations in Large Language Models and the Human Brain (NeurIPS 2025 · similarity 0.85)
- The Emergence of Abstract Thought in Large Language Models Beyond Any Language (NeurIPS 2025 · similarity 0.85)
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.
