A compact model learned to walk a knowledge graph one edge at a time and answer with a checkable evidence chain, then transferred to three graphs it had never seen
AspireAI community · arXiv
*Most LLM-plus-knowledge-graph systems preselect a slice of the graph before reasoning starts, which means the model is told what is relevant instead of finding out - and in specialised engineering domains where the answer lives in proprietary literature, frontier models tend to fall back on generic answers.* Prospector instead learns a traversal policy: it calls a small set of tools one at a time, builds its context progressively in light of what it has already found, and answers with a chain of evidence that can be checked edge by edge against what its tools actually retrieved. On a knowledge graph built from the semiconductor-manufacturing literature it approaches the strongest off-the-shelf models given the same graph and tools, and far exceeds off-the-shelf models answering closed-book or from a preselected slice. Trained only on the semiconductor graph, it transfers to three unseen graphs - bio-inspired materials, PFAS and materials properties - and its latent state turns out to be indicative of its next move on both seen and unseen graphs, so modulating that state can steer the policy between behaviour modes. It is served on a single GPU, answers faster and at a fraction of the cost of a hosted service, and keeps proprietary data on the owner's own hardware. The limitation is that the abstract reports relative standings rather than absolute accuracy numbers, so the practical gain over a strong frontier baseline still needs quantifying.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.