Research map / Large language models / Retrieval and knowledge
Mixing Mechanisms: How Language Models Retrieve Bound Entities In-Context
Yoav Gur-Arieh, Mor Geva, Atticus Geiger
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 |
|---|---|---|---|
| model positional mechanism as Gaussian instead of one-hot (ablation) | causal model M · Music task, n=20 | Jensen-Shannon similarity: 0.85 -> 0.95 | Improved |
| use full three-mechanism model instead of positional-only prevailing view | entity retrieval model · Music task, n=20 | Jensen-Shannon similarity: 0.44 -> 0.95 | Improved |
| ablate lexical term from M when target is first in group (te=1) (ablation) | causal model M · Music task, te=1 | Jensen-Shannon similarity: nearly no effect | No clear effect |
| ablate reflexive term from M when target is last in group (te=3) (ablation) | causal model M · Music task, te=3 | Jensen-Shannon similarity: nearly no effect | No clear effect |
| block attention from last token to query entity token in layers 11-16 (ablation) | gemma-2-2b-it · Binding retrieval accuracy | accuracy: 98% -> 37% | Hurt |
| also block attention to the token following the query token (ablation) | gemma-2-2b-it · Binding retrieval accuracy | accuracy: 37% -> 8% | Hurt |
| block last position to query token and its following token's attention (ablation) | gemma-2-2b-it · Binding retrieval accuracy | accuracy: accuracy drops to 6% | Hurt |
| block query token attention to matching entity token (reflexive signal) (ablation) | gemma-2-2b-it · Reflexive, te=1 | accuracy: reducing accuracy to 6% | Hurt |
| block query token attention to answer entity token (lexical signal) (ablation) | gemma-2-2b-it · Lexical, te=2 | accuracy: reduces accuracy only to 86% | No clear effect |
| block query token attention to all entity tokens (lexical) (ablation) | gemma-2-2b-it · Lexical, te=2 | accuracy: accuracy at 90% | No clear effect |
| add filler-text padding between entity groups (up to 500 tokens) | gemma-2-2b-it · Boxes task | mechanism effect share | Mixed |
| increase number of entity groups n | positional mechanism · Middle entity groups, n in [3,20] | positional efficacy | Hurt |
Closest papers by meaning
- How do Language Models Bind Entities in Context? (ICLR 2024 · similarity 0.90)
- Monitoring Latent World States in Language Models with Propositional Probes (ICLR 2025 · similarity 0.87)
- KScope: A Framework for Characterizing the Knowledge Status of Language Models (NeurIPS 2025 · similarity 0.87)
- L$^2$M: Mutual Information Scaling Law for Long-Context Language Modeling (NeurIPS 2025 · similarity 0.87)
- Understanding the Skill Gap in Recurrent Language Models: The Role of the Gather-and-Aggregate Mechanism (ICML 2025 · similarity 0.87)
- Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts (ICLR 2024 · similarity 0.87)
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.
