AI Paper Atlas

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

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Mechanistic analysis of attention and representations · 6 citations (OpenAlex)

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 changedWhereMeasuredEffect
use PH barcode summaries instead of raw logistic regressionlogistic regression · TaskTracker Mistral 7B layer 1accuracy: 0.8700 -> 1.0000Improved
use PH barcode summaries instead of linear SVMlinear SVM · TaskTracker Mistral 7B layer 1accuracy: 0.8875 -> 1.0000Improved
use PH barcode summaries instead of LDALDA · TaskTracker Mistral 7B layer 1accuracy: 0.995 -> 1.0000Improved
reduce activations with sparse autoencoder before logistic regression (ablation)logistic regression · TaskTracker Mistral 7B layer 8accuracy: 0.9950 -> 0.6225Hurt
reduce activations with sparse autoencoder before linear SVM (ablation)linear SVM · TaskTracker Mistral 7B layer 8accuracy: 1.0000 -> 0.6425Hurt
reduce activations with sparse autoencoder before LDA (ablation)LDA · TaskTracker Mistral 7B layer 8accuracy: 1.000 -> 0.998No clear effect
permute neuron indices (control) in local information flow analysis (ablation)local PH of consecutive layers · Mistral 7Bclean vs poisoned total persistence of 1-barsHurt
increase layer interval between compared layers to 10 (ablation)local PH analysis · Mistral 7Bclean-poisoned topological differenceHurt
poisoned (adversarial) inputs vs cleanPH of latent space · six LLMs 3.8B to 70Bmean death of 0-bars: increasesImproved
adversarial inputs vs clean on number of 1-barsPH of latent space · six LLMs 3.8B to 70Bnumber of H1 bars: decreasesHurt
LLaMA3 70B number of loops under attackPH of latent space · LLaMA3 70Bnumber of H1 barsNo clear effect
random-split controls (clean vs clean, poisoned vs poisoned) for local dispersion ratio (ablation)LDR · six LLMsdispersion ratio difference: centered around zeroNo clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy