Research map / Label-efficient and robust learning / Recognition and long-tail
The Information Geometry of Softmax: Probing and Steering
Kiho Park, Todd Nief, Yo Joong Choe, Victor Veitch
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 |
|---|---|---|---|
| steer in the dual space (dual steering) instead of adding the probe in the primal space (Euclidean steering) (ablation) | Gemma-3-4B steering, English to French concept · C4 contexts, off-target preservation vs target concept probability | off-target distribution preservation (counterfactual mass, KL, rank difference) | Improved |
| steer in the dual space instead of the primal space (ablation) | MetaCLIP-2 on synthetic objects, yellow to green · synthetic object images | off-target distribution preservation | Improved |
| steer in the dual space instead of the primal space (ablation) | MetaCLIP-2 on COCO, carrot to broccoli · COCO images | off-target distribution preservation | Improved |
| use dual mean difference instead of primal mean difference as probe direction for dual steering (ablation) | dual steering · LLM and CLIP steering | off-target preservation | No clear effect |
| Euclidean steering with primal mean difference probe (ablation) | Euclidean steering · multiple concepts, KL of off-target distribution | off-target KL divergence | Mixed |
| add regularization to the covariance in the Newton step | dual steering implementation · concentrated softmax contexts | feasibility of steering step | Improved |
| violate probe invariance assumption (probe entangled with off-target concepts) | dual steering · finite-data probes | off-target preservation vs Euclidean | No clear effect |
Closest papers by meaning
- The Linear Representation Hypothesis and the Geometry of Large Language Models (ICML 2024 · similarity 0.85)
- When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search (NeurIPS 2026 · similarity 0.85)
- Displacement Geometry Captures Platonic Shared Reality Across Models and Modalities (NeurIPS 2026 · similarity 0.84)
- Position: The Platonic Representation Hypothesis (ICML 2024 · similarity 0.84)
- The Sharp Directions Are Against You: Curvature Analysis of Activation Steering (NeurIPS 2026 · similarity 0.84)
- Task Vector Geometry Underlies Dual Modes of Task Inference in Transformers (NeurIPS 2026 · similarity 0.84)
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.
