Research map / Graph learning / Graph algorithms and clustering
Learning to Navigate Wikipedia by Taking Random Walks
Manzil Zaheer, Kenneth Marino, Will Sussman Grathwohl, John Schultz, Wendy Shang, Sheila Babayan, Arun Ahuja, Ishita Dasgupta, Christine Kaeser-Chen, Rob Fergus
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 random forward trajectories instead of reverse trajectories (ablation) | RFBC · 200k graph forward navigation 20-step | success rate (%): 0.2 -> 67.5 | Improved |
| train on shortest-path trajectories instead of random forward (ablation) | RFBC · 200k graph forward navigation 20-step | success rate (%): 67.5 -> 84.6 | Improved |
| use random features instead of RoBERTa embeddings (ablation) | RFBC · 200k graph navigation 5-step | success rate (%): 85.3 -> 17.4 | Hurt |
| replace behavioral cloning with RL (IMPALA) (ablation) | RFBC · 200k graph navigation 20-step | success rate (%): 67.5 -> 41.6 | Hurt |
| replace behavioral cloning with RL (ablation) | RFBC · 200k graph sentence search 10-step | success rate (%): 47.6 -> 1.5 | Hurt |
| fine-tune RFBC policy with RL (ablation) | RFBC · 200k graph navigation 20-step | success rate (%): 67.5 -> 68.1 | No clear effect |
| train embeddings on navigation instead of fixed RoBERTa (ablation) | feed-forward policy · 38M graph sentence search 20-step | success rate (%): 77.5 -> 90.2 | Improved |
| train embeddings on navigation instead of fixed RoBERTa (ablation) | feed-forward policy · 38M graph navigation 20-step | success rate (%): 71.5 -> 89.8 | Improved |
| use 4-layer Transformer policy instead of feed-forward (ablation) | learned embeddings · 38M graph navigation 20-step | success rate (%): 89.8 -> 92.2 | Improved |
| use 4-layer Transformer policy instead of feed-forward (ablation) | learned embeddings · 38M graph sentence search 20-step | success rate (%): 90.2 -> 79.7 | Hurt |
| scale training from 200k graph to full 38M graph | RFBC · sentence search 20-step | success rate (%): 34.6 -> 90.2 | Improved |
| add RFBC navigation to BM25 + TF-IDF (ablation) | BM25 + TF-IDF · FEVER dev | F1@5: 0.36 -> 0.46 | Improved |
Closest papers by meaning
- Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking (ICLR 2026 · similarity 0.85)
- Go-Browse: Training Web Agents with Structured Exploration (ICLR 2026 · similarity 0.83)
- WebLINX: Real-World Website Navigation with Multi-Turn Dialogue (ICML 2024 · similarity 0.83)
- DeepDiver: Adaptive Web-Search Intensity Scaling via Reinforcement Learning (NeurIPS 2025 · similarity 0.83)
- WebVLN: Vision-and-Language Navigation on Websites (AAAI 2024 · similarity 0.82)
- WebNavigator: Global Web Navigation via Interaction Graph Retrieval (NeurIPS 2026 · similarity 0.82)
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.
