Tag
This paper introduces Spatial Memory Agent (SMA), a runtime framework that improves frozen vision-language models' spatial reasoning through verifier-guided reflection and reusable memory without parameter updates or external tools, achieving strong results across five benchmarks and four base VLMs.
This paper presents a frozen 12B model that uses a persistent memory of verified solutions to achieve 100% accuracy on solved problem families with zero generation tokens, and demonstrates a 6M-token movable window on a single 46GB GPU.
A frozen 12B parameter model augmented with a persistent memory of verified solutions achieves 100% accuracy on over 180 instances across nine problem families with zero generation tokens, bit-exact deterministic outputs, outperforming frontier models on those solved tasks while using negligible compute.
The article discusses a Hugging Face experiment where an automated loop rewrites only the code (harness) around a frozen model, raising its benchmark score from 0% to near Sonnet 4.6 at lower cost, demonstrating that many benchmark failures stem from the harness, not the model itself.
The paper introduces δ-mem, a lightweight online memory mechanism that augments frozen LLMs with a compact associative memory state updated by delta-rule learning, achieving significant improvements on memory-heavy benchmarks without fine-tuning or context extension.