Research map / Large language models / Language, safety and interpretability
From Associations to Activations: Comparing Behavioral and Hidden-State Semantic Geometry in LLMs
Louis Schiekiera, Max Zimmer, Christophe Roux, Sebastian Pokutta, Fritz Günther
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 forced choice instead of free association behavior (ablation) | behavioral similarity · RSA with hidden states | Pearson r: .18 -> .43 | Improved |
| add FC similarity to baseline predictors (ablation) | ridge regression · held-out-words hidden-state similarity | test R2: +.022 | Improved |
| add FA similarity to baseline predictors (ablation) | ridge regression · held-out-words hidden-state similarity | test R2: +.002 | No clear effect |
| extract hidden states with task/meaning prompt instead of averaging over C4 sentences (ablation) | RSA with FC · RSA | mean Pearson r: .35 -> .46 | Improved |
| use Averaged extraction instead of meaning-focused prompts (ablation) | RSA with FC · layer of peak alignment | peak depth: about 30% -> about 72% depth | Mixed |
| use FC instead of FA by semantic category (ablation) | RSA · 20 of 21 WordNet supersenses | Pearson r: delta r about +.30 abstract, +.13 concrete | Improved |
| increase cue-word frequency (ablation) | FC alignment · RSA | Pearson r: .37 -> .47 | Improved |
| use cross-model consensus as predictor instead of behavior | hidden-state similarity prediction · RSA | Pearson r: .43 -> .74 | Improved |
| use FC behavior instead of FastText | behavioral reference · RSA | Pearson r: .20 -> .43 | Improved |
| increase neighborhood size k for FC (ablation) | nearest-neighbor overlap · NN@k | mean overlap: .20 -> .29 | Improved |
Closest papers by meaning
- Less is More: Local Intrinsic Dimensions of Contextual Language Models (NeurIPS 2025 · similarity 0.87)
- Unifying Attention Heads and Task Vectors via Hidden State Geometry in In-Context Learning (NeurIPS 2025 · similarity 0.87)
- Language Model Behavioral Phases are Consistent Across Architecture, Training Data, and Scale (NeurIPS 2025 · similarity 0.87)
- Invariant Features in Language Models: Geometric Characterization and Model Attribution (NeurIPS 2026 · similarity 0.87)
- The Labyrinth of Links: Navigating the Associative Maze of Multi-modal LLMs (ICLR 2025 · similarity 0.87)
- Decomposing how prompting steers behavior (NeurIPS 2026 · 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.
