research
The Transformer Had a Map but Lost Its Place
Causal probes traced TaxiGPT’s navigation failures to interference among learned location features.
Summary
Causal probes traced TaxiGPT’s navigation failures to interference among learned location features.
In a Manhattan navigation model, researchers found internal representations of streets, intersections, position and a goal compass. Interference among compressed intersection features disrupted localization, explaining failures that behavior alone had suggested were missing world knowledge.
Why it matters
Causal probes traced TaxiGPT’s navigation failures to interference among learned location features.
Limits and context
No additional limitation was separately recorded.
Key claims
Causal probes traced TaxiGPT’s navigation failures to interference among learned location features.
Evidence: source-2026-09-21-020
Sources
- arXiv preprint 2609.21748arXiv · primary research
Corrections
No corrections have been recorded for this story.