mirror of
https://github.com/hanzoai/papers.git
synced 2026-10-22 17:20:17 +00:00
zooqueen
4048ba80b0
hanzo-context-memory.tex carries the measured LoCoMo retrieval result
(cosine 22.7 -> linked 36.2 multi-hop R@20 ALL, ablated), the memory
record and read path, the protocol from bench/brain/METHOD.md, and every
pending number as \input{tables/...} with the run it will come from.
hanzo-context-engine.tex applies the abstraction to repositories with a
deterministic LSP-gold benchmark and marks its results pending.
Both build with tectonic; twenty citations verified against arXiv.
10 lines
344 B
TeX
10 lines
344 B
TeX
% TODO: generated from runs/coverage-locomo-<extractor>/metrics.json
|
|
\begin{tabular}{lrrr}
|
|
\toprule
|
|
Fact layer & gold turns with $\geq 1$ fact & with the right entity & with a dated fact \\
|
|
\midrule
|
|
LoCoMo observations (dataset) & 79.9\% (multi-hop) & \ldots & \ldots \\
|
|
Hanzo extraction & \ldots & \ldots & \ldots \\
|
|
\bottomrule
|
|
\end{tabular}
|