Research map / Architectures and efficiency / Attention and transformers
MetaOthello: A Controlled Study of Multiple World Models in Transformers
Aviral Chawla, Galen Hall, Juniper L Lovato
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 |
|---|---|---|---|
| Train on mixed-variant data instead of single game | 8-layer GPT · Othello variants | alpha score: about 0.5% dip | No clear effect |
| Use probe from one game to intervene on another (cross-probe) (ablation) | Classic-NoMidFlip and Classic-DelFlank models · Board-state intervention error | prediction error: correct 0.2 vs cross 0.2 (Classic, Classic-NoMidFlip) | No clear effect |
| Apply Procrustes rotation to Classic vs Iago probe weights (ablation) | Classic-Iago mixed model · Probe weight cosine similarity | cosine similarity: 0.03 -> 0.98 | Improved |
| Apply a single global orthogonal rotation to residual stream before predicting Iago moves (ablation) | Classic-Iago mixed model · Classic-to-Iago transfer | alpha score: -2.9 -> near-perfect | Improved |
| Add Procrustes alignment to Classic vs DelFlank probes (ablation) | Classic-DelFlank mixed model · Probe weight cosine similarity | cosine similarity: 0.67 -> >0.90 | Improved |
| Add game-ID steering vector at layer 5 (ablation) | Classic-NoMidFlip model · Ambiguous prefixes | NoMidFlip-consistent move share | Improved |
| Steer full disambiguation circuit (L4 MLP, L5 head 5, L5 MLP) (ablation) | Classic-NoMidFlip model · Held-out ambiguous prefixes | Classic share of unique-move mass: 0.48 -> 0.68 (Classic), 0.01 (NoMidFlip) | Improved |
| Steer runner-up layer-5 head as control (ablation) | Classic-NoMidFlip model · Held-out ambiguous prefixes | Classic share of unique-move mass: stays at baseline | No clear effect |
| Ablate disambiguation circuit (ablation) | Classic-NoMidFlip model · Held-out ambiguous prefixes | valid-move mass: 0.999 -> 0.964 | Hurt |
| Ablate disambiguation circuit (ablation) | Classic-NoMidFlip model · Held-out ambiguous prefixes | mass on wrong-game-only moves: 0.0001 -> 0.0115 | Hurt |
| Steer game identity at early layers 2-3 instead of layer 5 (ablation) | Classic-DelFlank model · Ambiguous sequences | normalized alpha toward DelFlank: delta alpha about 0.75 at low move numbers; layer 5 about 0 | Improved |
| Layer-2 steering toward DelFlank (ablation) | Classic-DelFlank model · Layer-5 DelFlank probe accuracy | probe accuracy (%): 67.1 -> 77.4 | Improved |
Closest papers by meaning
- Evaluating the World Model Implicit in a Generative Model (NeurIPS 2024 · similarity 0.86)
- Do Transformer World Models Give Better Policy Gradients? (ICML 2024 · similarity 0.85)
- How Far Can Transformers Reason? The Globality Barrier and Inductive Scratchpad (NeurIPS 2024 · similarity 0.85)
- JoMA: Demystifying Multilayer Transformers via Joint Dynamics of MLP and Attention (ICLR 2024 · similarity 0.85)
- Localizing Knowledge in Diffusion Transformers (NeurIPS 2025 · similarity 0.84)
- Transformers as Game Players: Provable In-context Game-playing Capabilities of Pre-trained Models (NeurIPS 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.
