research
Memory Becomes a Path Instead of a Search
A persistent local traversal head reached 0.95 mean return after 0.59 million training steps in a five-token MiniGrid test.
Summary
A persistent local traversal head reached 0.95 mean return after 0.59 million training steps in a five-token MiniGrid test.
Long-context systems repeatedly search large stores from scratch, while recurrent compression must decide early what to keep. This proof of principle separates a memory's information value from its address and lets an agent traverse a persistent local neighborhood. On MiniGrid Memory with a five-token context, the local traversal head reached 0.95 mean return after 0.59 million training steps, compared with 0.94 million for Gaussian absolute addressing and more than 2 million for categorical addressing. The experiment used fixed geometry and did not test the paper's full learned-geometry or repeated-retrieval proposal.
Why it matters
A persistent local traversal head reached 0.95 mean return after 0.59 million training steps in a five-token MiniGrid test.
Limits and context
- The experiment used fixed geometry and did not test the paper's full learned-geometry or repeated-retrieval proposal.
Key claims
A persistent local traversal head reached 0.95 mean return after 0.59 million training steps in a five-token MiniGrid test.
Qualification: The experiment used fixed geometry and did not test the paper's full learned-geometry or repeated-retrieval proposal.
Evidence: source-2026-10-31-004
Sources
- arXiv preprint 2610.40545, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.