Research map / Large language models / Language, safety and interpretability
The Shape of Adversarial Influence: Characterizing LLM Latent Spaces with Persistent Homology
Aideen Fay, Inés García-Redondo, Qiquan Wang, Haim Dubossarsky, Anthea Monod
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 |
|---|---|---|---|
| use PH barcode summaries instead of raw logistic regression | logistic regression · TaskTracker Mistral 7B layer 1 | accuracy: 0.8700 -> 1.0000 | Improved |
| use PH barcode summaries instead of linear SVM | linear SVM · TaskTracker Mistral 7B layer 1 | accuracy: 0.8875 -> 1.0000 | Improved |
| use PH barcode summaries instead of LDA | LDA · TaskTracker Mistral 7B layer 1 | accuracy: 0.995 -> 1.0000 | Improved |
| reduce activations with sparse autoencoder before logistic regression (ablation) | logistic regression · TaskTracker Mistral 7B layer 8 | accuracy: 0.9950 -> 0.6225 | Hurt |
| reduce activations with sparse autoencoder before linear SVM (ablation) | linear SVM · TaskTracker Mistral 7B layer 8 | accuracy: 1.0000 -> 0.6425 | Hurt |
| reduce activations with sparse autoencoder before LDA (ablation) | LDA · TaskTracker Mistral 7B layer 8 | accuracy: 1.000 -> 0.998 | No clear effect |
| permute neuron indices (control) in local information flow analysis (ablation) | local PH of consecutive layers · Mistral 7B | clean vs poisoned total persistence of 1-bars | Hurt |
| increase layer interval between compared layers to 10 (ablation) | local PH analysis · Mistral 7B | clean-poisoned topological difference | Hurt |
| poisoned (adversarial) inputs vs clean | PH of latent space · six LLMs 3.8B to 70B | mean death of 0-bars: increases | Improved |
| adversarial inputs vs clean on number of 1-bars | PH of latent space · six LLMs 3.8B to 70B | number of H1 bars: decreases | Hurt |
| LLaMA3 70B number of loops under attack | PH of latent space · LLaMA3 70B | number of H1 bars | No clear effect |
| random-split controls (clean vs clean, poisoned vs poisoned) for local dispersion ratio (ablation) | LDR · six LLMs | dispersion ratio difference: centered around zero | No clear effect |
Closest papers by meaning
- Remarkable Robustness of LLMs: Stages of Inference? (NeurIPS 2025 · similarity 0.87)
- Less is More: Local Intrinsic Dimensions of Contextual Language Models (NeurIPS 2025 · similarity 0.87)
- Invariant Features in Language Models: Geometric Characterization and Model Attribution (NeurIPS 2026 · similarity 0.86)
- When Witnesses Defend: A Witness Graph Topological Layer for Adversarial Graph Learning (AAAI 2025 · similarity 0.86)
- Topological Signatures of Adversaries in Multimodal Alignments (ICML 2025 · similarity 0.86)
- Towards Scalable Topological Regularizers (ICLR 2025 · similarity 0.86)
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.
