Research map / Large language models / Language, safety and interpretability
Do Language Models Track Entities Across State Changes?
Zilu Tang, Qiao Zhao, Gabriel Franco, Derry Tanti Wijaya, Aaron Mueller, Sebastian Schuster, Najoung Kim
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 |
|---|---|---|---|
| global (all objects) linear probe on last-token residual stream | CodeLlama-13B · box entity tracking | global probe non-trivial accuracy: global < 0.3 vs local near 0.9 | No clear effect |
| probe for prior states across layers | CodeLlama-13B · box entity tracking, 1-3 local operations | probe accuracy by layer | No clear effect |
| patch heads or subspaces from description circuit to PUT circuit (ablation) | CodeLlama-13B · PUT vs DESCRIPTION circuits | patching recovery | Improved |
| nullify remove tag at box ID token (ablation) | CodeLlama-13B · REMOVE intervention | target object success | No clear effect |
| nullify remove tag at object token (ablation) | CodeLlama-13B · REMOVE intervention, earlier layers | target object success | Improved |
| nullify remove tag at object token in failure cases (ablation) | CodeLlama-13B No-op Remove · No-op Remove | intervention success: 0.66 | Improved |
| nullify remove tag at object token in failure cases (ablation) | CodeLlama-13B re-introduction · Re-introducing removed object | intervention success: 0.24 | Mixed |
| scale model from 13B to 70B (ablation) | entity tracking diagnostics · Re-introducing removed object | degeneration rate: 0.62 -> 0.01 | Improved |
| add more operations in context (ablation) | CodeLlama-13B · box ID remove tag probe | probe accuracy | Hurt |
Closest papers by meaning
- (How) Do Language Models Track State? (ICML 2025 · similarity 0.90)
- Fine-Tuning Enhances Existing Mechanisms: A Case Study on Entity Tracking (ICLR 2024 · similarity 0.88)
- MET-Bench: Multimodal Entity Tracking for Evaluating the Limitations of Vision-Language and Reasoning Models (ICML 2026 · similarity 0.86)
- Persistent Semantic Entities in Tool-Augmented LLM Systems (ICML 2026 · similarity 0.85)
- The Expressive Capacity of State Space Models: A Formal Language Perspective (NeurIPS 2024 · similarity 0.84)
- The Illusion of State in State-Space Models (ICML 2024 · 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.
