Research map / Vision-language and multimodal / Human and embodied modeling
The Mechanistic Emergence of Symbol Grounding in Language Models
Shuyu Wu, Ziqiao Ma, Xiaoxi Luo, Yidong Huang, Josue Torres-Fonseca, Freda Shi, Joyce Chai
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 |
|---|---|---|---|
| zero out aggregate heads (vs random heads of equal number) (ablation) | 12-layer Transformer on CHILDES, 20k steps · CHILDES grounding (match condition) | surprisal: 4.77 -> 5.62 (control 4.76) | Hurt |
| zero out gather heads (vs random heads of equal number) (ablation) | 12-layer Transformer on CHILDES, 20k steps · CHILDES grounding (match condition) | surprisal: 4.69 vs control 4.79 | No clear effect |
| zero out aggregate heads (vs random heads) (ablation) | 12-layer Transformer on Visual Dialog, 300k steps, 70% threshold · Visual Dialog image-grounded | surprisal: 7.69 -> 8.96 (control 8.11) | Hurt |
| zero out aggregate heads (vs random heads) (ablation) | 12-layer Transformer on Visual Dialog, 300k steps, 90% threshold · Visual Dialog image-grounded | surprisal: 7.69 -> 8.40 (control 7.87) | Hurt |
| train Transformer or Mamba-2 (vs unidirectional LSTM) on grounded data (ablation) | 4-layer models on CHILDES · CHILDES grounding | match vs mismatch surprisal gap | Improved |
| grounding information gain increases beyond co-occurrence statistics | Transformer and Mamba-2 · CHILDES grounding | R2 with log co-occurrence | Mixed |
| replace CHILDES text environment with image patches (DINOv2) as environmental ground | 12-layer Transformer · Visual Dialog | grounding information gain | Mixed |
Closest papers by meaning
- The Grounding Gap: How LLMs Anchor the Meaning of Abstract Concepts Differently from Humans (NeurIPS 2026 · similarity 0.88)
- The Emergence of Abstract Thought in Large Language Models Beyond Any Language (NeurIPS 2025 · similarity 0.88)
- GroundingME: Exposing the Visual Grounding Gap in MLLMs through Multi-Dimensional Evaluation (CVPR 2026 · similarity 0.87)
- Ground-Compose-Reinforce: Grounding Language in Agentic Behaviours using Limited Data (NeurIPS 2025 · similarity 0.87)
- Bridging semantics and pragmatics in information-theoretic emergent communication (NeurIPS 2024 · similarity 0.86)
- The Hidden Life of Tokens: Reducing Hallucination of Large Vision-Language Models Via Visual Information Steering (ICML 2025 · 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.
