Research map / Large language models / Language, safety and interpretability

PersonaManifold: Revealing and Exploiting Curved Geometry in LLM Persona Representations

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: Task vectors and steering · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Models LLM persona activations as a curved Riemannian submanifold and uses geodesic distances and geodesic steering (plus a new BST behavioral-similarity benchmark) to improve persona similarity prediction and interpolation over Euclidean methods.

Problem

Inference-time persona control via persona vectors assumes linear, isotropic geometry in activation space, yet existing methods report non-orthogonal traits, asymmetric effects, and composition deviations. The paper asks whether persona representations have curved geometry that Euclidean operations

Approach

Estimate a local metric tensor, kNN-graph geodesic distances, and Ollivier-Ricci curvature over ~2K persona activations, then steer by following the manifold geodesic between two personas instead of the Euclidean straight line; evaluate with a behavior-grounded Behavioral Similarity Triplet benchmark.

Key results

• Geodesic distance achieves TCR-combined of 68.2% on Llama-3.1-8B vs 62.1% for Euclidean on BST (+6.1 points).
• On high-deviation steering pairs, geodesic steering reaches IC 0.72 and TS 0.69 on Llama vs 0.58 and 0.52 for Euclidean-Linear.
• Human naturalness ratings rise to 3.82 for geodesic steering vs 3.41 for linear steering (Krippendorff's alpha 0.73).

Surprising findings

• Geodesic steering gains are concentrated in high-deviation pairs, with near-zero gain on near-flat pairs, so the benefit depends on how far the manifold deviates from flatness.

Limitations

• Experiments limited to 7-8B open models; no 70B+ evaluation.
• Each persona is represented by a single mean activation vector, collapsing intra-persona variation.
• BST covers six psychological constructs and relies on GPT-5.2 generated questions and responses.

Experimental setup

Datasets: PersonaHub CoSER BFI-44 HEXACO-60 TIPI PersonaGym

Benchmarks: BST PersonaGym BFI-44

Metrics: Triplet Consistency Rate (TCR) Interpolation Coherence Transition Smoothness Manifold Adherence MRR naturalness Likert

Baselines: Euclidean-Linear SLERP Mahalanobis Isomap-Euclidean Diffusion Map UMAP Graph-NN Chain

Builds on: Isomap Ollivier-Ricci curvature BFI-44 PersonaHub

Closest papers by meaning

• LLM Generated Persona is a Promise with a Catch (NeurIPS 2025 · similarity 0.86)
• PICLe: Eliciting Diverse Behaviors from Large Language Models with Persona In-Context Learning (ICML 2024 · similarity 0.86)
• Persona Vectors: Monitoring and Controlling Character Traits in Language Models (NeurIPS 2026 · similarity 0.86)
• Bias Runs Deep: Implicit Reasoning Biases in Persona-Assigned LLMs (ICLR 2024 · similarity 0.85)
• PersonaVLM: Long-Term Personalized Multimodal LLMs (CVPR 2026 · similarity 0.85)
• Persona Generators: Generating Diverse Synthetic Personas for Arbitrary Contexts (NeurIPS 2026 · similarity 0.85)