Caesar keeps a graph of its web exploration: pages are nodes and followed links become edges. Its note-taking prompt includes insights from nearby pages, allowing each source to be read in the context of earlier material.
An exploratory comparison found that notes from well-connected pages scored 24.43 out of 30, compared with 21.88 for notes from pages with one neighbour. The comparison suggests graph position may help select useful material, but it does not isolate a benefit from the graph context itself.
Selecting hub and leaf notes
The tested prompt draws on pages up to three hops away, with at most 24 neighbour entries. A separate traversal-history block supplies up to eight previous pages regardless of graph position. Even a leaf can inherit context through the page that led to it: that edge exists before its notes are written.
Five completed explorations, one per research challenge, supplied two groups apiece. Hubs were the ten most-connected pages visited more than once. They had 5 to 36 neighbours and 3 to 45 visits. Leaves were the ten least-connected pages visited exactly once; each had one neighbour.
Each group’s ten notes were concatenated into a file, producing five hub files and five leaf files. These contained roughly 5,000–5,800 words each; the leaf file was longer in three challenges. The evaluated artifacts were raw notes, not research answers generated from those notes. There was no further writing step.
Three LLM judges scored every file five times on New, Useful and Surprising, each from 1 to 10. That yields 75 scorings per group, but only five distinct files per group.
| Dimension | Hub notes | Leaf notes |
|---|---|---|
| New | 8.35 | 7.32 |
| Useful | 7.80 | 7.21 |
| Surprising | 8.28 | 7.35 |
| Total | 24.43 | 21.88 |
Hub notes led on pooled scores for all three dimensions and on total scores in all five challenges. Usefulness had two exceptions: leaves led 7.87 to 7.67 on open-ended synthesis and tied hubs at 7.67 on cross-domain synthesis.

Effect sizes depend on the unit
At the challenge level, hub means were 25.07, 23.93, 25.07, 24.40 and 23.67; leaf means were 20.87, 21.33, 22.27, 21.47 and 23.47. The lowest hub mean exceeds the highest leaf mean, yielding Cliff’s delta = 1.00 under the paper’s challenge-mean convention.
At the individual-scoring level, the distributions overlap substantially. Of all hub–leaf score pairings, about 73% favour hubs, 19% favour leaves and 8% tie, yielding delta = 0.54 (reported values and plotting code). Hub totals range from 18 to 29; leaf totals from 15 to 28. As in the human-evaluation discussion, a delta computed over means does not describe every individual score.
The margin also varies by task. Four challenges show leads of 2.6–4.2 points; open-ended synthesis differs by only 0.20 points. The analysis script’s Mann–Whitney U test reports p = 8.3 × 10⁻⁹ over individual scorings, but repeated judgments of the same texts do not provide 150 independent text samples.
A second set of explorations used the older one-hop prompt without traversal history and differed in other configuration settings. Two judging passes gave 23.29 versus 21.76, then 23.49 versus 22.17, using 45 and 75 scorings per group, respectively. Both favoured hubs overall, but leaves won counterfactual reasoning. Challenge-level deltas were 0.92 and 0.52; individual-scoring deltas were 0.33 and 0.30. The direction recurred, while the five-for-five result did not. Configuration differences prevent treating this as a controlled prompt ablation. The additional replication results use unpublished run records.
Four unresolved confounders
Visible labels. Both files appeared in the same prompt under names including
answer_cat_hubs.txt and answer_cat_leaves.txt. Judges could infer the
condition and bring expectations about connectivity to the comparison. No
blinded rerun is available; label effects cannot be ruled out as an explanation
for the full gap.
Unreadable sources. Sixteen of the 50 leaf entries concern bot checks, redirects or access gates, compared with one of the 50 hub entries. Nine occur on constrained synthesis, the challenge with the largest gap, and six on counterfactual reasoning. Across the other three challenges, which contain only one such leaf entry, the mean gap is 1.98 rather than 2.55 points. The selection rule encourages this imbalance: blocked pages offer few links and little reason to return.
Unequal revision. Hubs were revisited; leaves were not. Caesar rewrites notes on a return visit with previous notes in context. Additional processing, in one case across 45 visits, is inseparable from position in this design.
Different source quality. Surveys and canonical references may attract more connections and contain better material. The comparison cannot distinguish a useful location in the graph from a useful source.
Connectivity remains a potentially inexpensive selection signal when an agent has more notes than a synthesis prompt can hold. Testing whether it adds value requires a controlled comparison of the same pages with and without neighbour context, matched access and revision opportunities, and blinded scoring. The current result is an association between the selected note groups under the reported judging conditions.
Source: Caesar paper, revised version.