PersonaManifold Maps LLM Personas Through Curved Space

// 1h agoRESEARCH PAPER

PersonaManifold Maps LLM Personas Through Curved Space

PersonaManifold models LLM persona activations as curved Riemannian manifolds, using geodesic paths instead of straight-line interpolation for more coherent persona steering. Its accompanying BST benchmark evaluates persona similarity through behavioral responses.

// ANALYSIS

PersonaManifold makes persona steering feel less like vector arithmetic and more like navigation through a constrained behavioral landscape.

–Geodesic distances outperform Euclidean alternatives for predicting behavioral similarity
–Curvature-aware paths target the regions where linear steering produces the most distortion
–The BST benchmark grounds evaluation in situational behavior rather than self-reported traits
–The MIT-licensed implementation offers PCA-based geometry estimation, graph geodesics, steering vectors, curvature analysis, and triplet evaluation

// TAGS

personamanifoldllminterpretabilityevaluationopen-sourceresearch

DISCOVERED

1h ago

2026-09-30

PUBLISHED

1h ago

2026-09-30

RELEVANCE

10/ 10

AUTHOR

Discover AI

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

UPDATE16m ago

Ollama 0.35 Adds Local Decision Models

Ollama 0.35 adds the `/v1/systemone` API for local decision models that return typed choices, probabilities, and scores instead of prose. The feature supports ticket triage, model routing, policy checks, and content classification through models including Nimble and Tev1.

MODEL16m ago

Nimble lands in Ollama with typed decisions

Bespoke Labs’ Apache 2.0 Nimble is a 9B Qwen3.5-9B fine-tune for fast local classification, routing, moderation, and scoring. Ollama now exposes it through the /v1/systemone API, where it reports 75.7% accuracy across 13 public datasets—just behind Jev’s 76.0%.

TUTORIAL24m ago

Aura Makes DESIGN.md Design Memory

This tutorial shows how Aura uses reusable DESIGN.md files to carry typography, color, spacing, layout, motion, and interaction rules across AI-generated websites. Aura also supports URL-based design extraction, visual editing, and HTML export.