Research map / Large language models / Language, safety and interpretability
Neural Synchrony Between Socially Interacting Language Models
Zhining Zhang, Wentao Zhu, Chi Han, Yizhou Wang, Heng Ji
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 |
|---|---|---|---|
| keep genuine social engagement (versus a passive reader agent) between the two LLMs (ablation) | SyncR2 · Sotopia model pairs | SyncR2: significantly reduced compared to the experimental group | Improved |
| pair representations at the same time step rather than at a lag k (ablation) | SyncR2 · Sotopia model pairs | SyncR2: synchrony collapses when representations are paired across different lags | Improved |
| clip negative R2 to zero (ablation) | SyncR2 · Sotopia model pairs | SyncR2 stability: keeping negative R2 values substantially reduces SyncR2 and markedly inflates standard errors | Improved |
| average top-k layer matches instead of best match (larger k), Mistral family (ablation) | SyncR2 · Sotopia correlation with social performance | correlation: increasing k strengthens the correlation | Improved |
| average top-k layer matches (larger k), Llama and cross-family (ablation) | SyncR2 · Sotopia correlation with social performance | correlation: slightly weaken this correlation; r > 0.7, p<0.05 | Hurt |
| use new sampling with no shared persona pairs across train/test (ablation) | SyncR2 · Sotopia, three model pairs | SyncR2: 0.46 -> 0.47; 0.55 -> 0.58; 0.75 -> 0.74 | No clear effect |
| replace affine map with two-layer nonlinear map, Mistral-v0.2 pair (ablation) | SyncR2 · Sotopia, Mistral-v0.2 & Mistral-v0.2 | SyncR2: 0.46 -> 0.47 | No clear effect |
| replace affine map with two-layer nonlinear map, Mistral-Llama pair (ablation) | SyncR2 · Sotopia, Mistral-v0.2 & Llama-3-8B | SyncR2: 0.55 -> 0.60 | Improved |
| replace affine map with two-layer nonlinear map, Llama-3 pair (ablation) | SyncR2 · Sotopia, Llama-3-8B & Llama-3-8B | SyncR2: 0.75 -> 0.59 | Hurt |
| control for IFEval and MuSR abilities via partial correlation | SyncR2 vs social performance · Sotopia | partial correlation: Llama IFEval 0.27; Mistral 0.81 and 0.92; Cross 0.71 and 0.89 | No clear effect |
| closer agent relationships (strangers to romantic partners) | SyncR2 · Sotopia relationship subsets | SyncR2: distribution shifts upward as social closeness increases | Improved |
| increase training sample size beyond 6500 | SyncR2 · Sotopia model pairs | SyncR2: little improvement beyond 6500 samples | No clear effect |
Closest papers by meaning
- Divergences between Language Models and Human Brains (NeurIPS 2024 · similarity 0.88)
- Language Models Represent Beliefs of Self and Others (ICML 2024 · similarity 0.87)
- Scaling and context steer LLMs along the same computational path as the human brain (NeurIPS 2025 · similarity 0.86)
- Position: LLM Social Simulations Are a Promising Research Method (ICML 2025 · similarity 0.86)
- Exploring Prosocial Irrationality for LLM Agents: A Social Cognition View (ICLR 2025 · similarity 0.85)
- Using Artificial Populations to Study Psychological Phenomena in Neural Models (AAAI 2024 · similarity 0.85)
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.
