Current research2025-2026 · Present (2015 – Oct 2026)

Whole Herculaneum Scroll Read

In June 2026 the Vesuvius Challenge team announced it had virtually unwrapped and read PHerc. 1667 end to end, the first Herculaneum papyrus read in full without opening (preprint, not yet peer reviewed).

Open in the interactive tree →

Synchrotron X-ray scans are split into papyrus layers by software, and machine-learning models find the faint carbon ink, so the scroll never has to be unrolled. PHerc. 1667, kept in Naples, gave about 1.4 to 1.5 metres of continuous Greek in 20 to 22 columns, depending on the source; it was judged unreadable when part of it was opened in the 1980s. The text looks like a Stoic treatise on ethics from the second century BC, author unknown.

As of October 2026

On 25 June 2026 in Naples the team announced PHerc. 1667 read in full, PHerc. 139 identified as Philodemus's On Gods, Book 8, and the bottom 70% of PHerc. 172 unwrapped; PHerc. Paris 4 was rescanned at 2.4 micrometres and about 140 columns unwrapped. The reading is described in a 27 June arXiv preprint that is not yet peer reviewed, and the project says it has released the data, surfaces, transcriptions and code. A new $1M Grand Prize (deadline 25 June 2027) asks for another full scroll with at most eight hours of human input, and the team says its methods still need humans in the loop. scrollprize.org lists 45 scrolls and fragments scanned, with ink found in 9.

Open steps

Cut human input to hours per scroll High AI leverageAutomate tracing and flattening of papyrus layers so a full scroll needs no more than eight hours of human input, the rule for the 2027 Grand Prize.
Catch letters a model invents Medium AI leverageBuild checks that separate real ink from hallucinated letters, for example by comparing independent models and scans of the same scroll.
Scan more scrolls at high detail Low AI leverageSynchrotron scans give the best detail but reach up to 300 TB per scroll; more scrolls need beamtime, storage and permission from the Naples library.
Identify and edit the new texts Medium AI leverageTurn unwrapped images into edited Greek and name the works, as for Philodemus's On Gods, Book 8 (PHerc. 139), which showed the work ran to at least eight books.

Where AI could help

High AI leverage. Layer tracing and ink detection are image problems that machine learning already handles in part; people still guide and check each step.

Trace fused papyrus layers through terabyte scan volumes
Detect faint ink in 3D scan data
Flag letters a model may have invented
Align and compare text across columns and scrolls

Shown so far

A 27 June 2026 arXiv preprint by 27 authors, not yet peer reviewed, claims the first complete digital unrolling and reading of a sealed Herculaneum papyrus, using micro-CT, improved computational unrolling and machine learning. source
On 15 July 2026 the Vesuvius Challenge said its methods still need humans in the loop and that full automation is required to read the whole collection in a reasonable time. source

Prerequisites

Unlocks

Reading All Herculaneum Scrollsopen