Research map / Label-efficient and robust learning / Recognition and long-tail
Convergent World Representations and Divergent Tasks
Core Francisco Park
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 |
|---|---|---|---|
| train on two tasks instead of one | city-coordinate transformer · layer 5 CKA across models | CKA: 0.48 -> 0.84 | Improved |
| increase pretraining task count from 1 to 2 or 3 | city-coordinate transformer · layers 4-6 CKA, disjoint task sets | average CKA: saturating at 0.85 | Improved |
| add a second task to crossing | crossing task · single-task training | learnability | Improved |
| include distance task in two-task fine-tuning mixture (ablation) | 7-task pretrained model · Atlantis fine-tuning, cross-task generalization | deviation from best-teacher | Hurt |
| fine-tune on angle task alone | 7-task pretrained model · Atlantis integration | cross-task improvement | Improved |
| fine-tune on distance task alone | 7-task pretrained model · Atlantis integration | cross-task improvement | No clear effect |
| fine-tune with divergent (distance) task data (ablation) | 7-task pretrained model · Atlantis coordinate probe reconstruction | reconstruction error: nearly an order of magnitude larger | Hurt |
| include Atlantis cities from start of pretraining instead of fine-tuning (ablation) | 7-task model · Atlantis coordinate probe reconstruction | reconstruction error | Improved |
| use batch size 512 instead of 128 in fine-tuning (ablation) | fine-tuned models · fine-tuning | task performance | Hurt |
Closest papers by meaning
- Disentangling Representations through Multi-task Learning (ICLR 2025 · similarity 0.87)
- Position: The Platonic Representation Hypothesis (ICML 2024 · similarity 0.87)
- Bridging Critical Gaps in Convergent Learning: How Representational Alignment Evolves Across Layers, Training, and Distribution Shifts (NeurIPS 2025 · similarity 0.86)
- Do LLMs Build World Representations? Probing Through the Lens of State Abstraction (NeurIPS 2024 · similarity 0.85)
- Task structure and nonlinearity jointly determine learned representational geometry (ICLR 2024 · similarity 0.84)
- Parameter symmetries determine representational geometry in overparameterized nonlinear networks (NeurIPS 2026 · 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.
