The Encoder Saw Space Yet the Tongan Scholars Saw Time

When 5 researchers posted H-JEPA to arXiv on 5 October 2026, they gave a robot planner a hierarchy of world models. Wancong Zhang, Basile Terver, Michael Rabbat, Yann LeCun and Randall Balestriero sign the paper from NYU, Advanced Machine Intelligence, INRIA Paris and Brown University. Their recipe trains several action-conditioned JEPA world models jointly. Higher levels predict farther ahead, in their own latent space. On Visual AntMaze, a simulated quadruped in a maze, a 3-level hierarchy lifted planning success from 18% to 73% with less planner compute.

Level 1 of the AntMaze figure predicts 5 environment steps per prediction. The stride doubles at level 2 and doubles again at level 3, for 20 steps at the top. Planning starts with the level-3 planner, which optimises a path toward the goal. Its first predicted state becomes a subgoal for level 2. Level 2 repeats the step for level 1, the bottom level and the only level to output raw motor actions. Decoded images from the upper levels show the ant's position and the maze layout. The pose of its legs fades.

ʻŌkusitino Māhina presented a paper titled Tā, Vā and Faiva to a philosophy conference in Chico, California, in 2001. With Tēvita O. Kaʻili he had formulated the Tā-Vā theory of reality around 2000, a Moana account of tā (time) and vā (space) as inseparable. Across the Moana the pair also goes by kā and wā. An ASAO working session on the theory opened in Santa Cruz in 2009. Ping-Ann Addo joined it at the 2011 meeting in Honolulu. The 2017 special issue of Pacific Studies printed most of the 13 papers.

In Tongan drawing, a mata (eye, point) marks the intersection of 2 or more kohi (lines). A kohi is a serial row of mata and a manifestation of tā. Vā arises as a parallel summation of kohi. Beside the H-JEPA hierarchy, the Tongan units match its levels. A level-1 latent is a point in time. Level 2 strings those points at a stride of 2, a serial row or kohi in Tongan terms. Planners at the 3 levels optimise several candidate rows by gradient descent, a parallel set of futures.

Kaʻili and Māhina file tā under the verb and vā under the noun in Tongan grammar. Their examples are tala matangi (telling the conditions of winds) and tanu hala (making roads). Telling marks the winds as a space in time. Making marks the road. Tā marks time by tempo and beat, as in tānafa (the rhythmic beating of drums). Vā names the relational space between such markers, and its quality. There are 2 Tongan compounds for that quality, vāmamaʻo for distant physical space and vālelei for harmonious social space.

For the hardest test, the H-JEPA authors trained on DROID, a corpus of real teleoperated robot video. On DROID the scene changes between episodes, along with the lighting and the objects. Between-episode variance accounts for 75% of latent variance on DROID, against 28% on the OGBench Cube task. Against a changing background, the scene is the easy part to predict. Only the actions predict the arm. On the predictive loss alone, the DROID encoder retained the scene and discarded the moving arm, which dropped planning performance to 0. The authors call the result slow-feature collapse.

Kaʻili, Māhina and Addo had described this trade in 2017. Their target was the 2014 Ethos special issue Senses of Space, edited by Mawyer and Feinberg. By their account, its authors primarily saw only space. Giovanni Bennardo's paper on Tongan direction-giving, they wrote, treats directions as spatial knowledge alone. The DROID encoder made the trade in tāvāist terms. It retained vā in the form of the scene. It dropped tā in the form of the arm's action. Tāvāism attributes such errors to a mind separated from reality and failing to grasp connection and separation.

The authors repaired the collapse with inverse dynamics. They added a small head. It takes 2 consecutive state representations and regresses the action between them. Once the authors added the term, flat LeWM became a strong baseline. A second learned level then raised performance at a lower planner budget. The authors score DROID plans offline with Fréchet fidelity, with 0% for an arm left still and 100% for the expert path. Māhina and Kaʻili call tā the marker of vā. Vā, in turn, composes tā. The inverse-dynamics head penalises an encoder for dropping the marker between 2 states.

The paper lists goals stated in words among its broader directions, through upper-level latents aligned with language. Its authors are also proposing upper levels trained on variable-duration segments in place of fixed strides. Tāvāists treat a beat as a marker of time and the interval between beats as a vā.

On a course for a digital learning platform in lea faka-Tonga, a first-language reviewer stopped at a line of the lesson. The English source leaned on phrases with little currency in Tongan life. The reviewer asked for the full passage in Tongan, with added explanation. Huri's linguists then checked the terminology against the English source in 2 rounds. They locked the terms for the on-screen labels and the merge fields. A slip there breaks the build or mislabels a control. In the second round, the reviewer replaced ako with Fakataukei, a Tongan term already in use earlier in the sentence.

For an insurer's appeal notice in Tongan, a first-language reviewer rewrote the full sentence about the days between 2 dates. The first date was the day a member received a notice denying a claim. The second was the day the member lodged a request for a fair hearing. The reviewer's version counts those days in a vahaʻataimi. A standard Tongan dictionary glosses vahaʻataimi as a space or interval of time. It glosses vahaʻa as intervening space or time, with a cross-reference to vā. In Tā-Vā terms, the 2 dates are time-markers and the days between them form a vā.

On a broadcast script localised into Tongan, Samoan and Hawaiian, the reviewer asked for the times in brackets to be spoken in 24-hour form, in the target language. A global note then asked the linguist to listen to the recorded voice and confirm the recording against the script. The 24-hour format counts the day in fixed steps, the way H-JEPA levels count environment steps at a fixed stride.

The Tongan panel on Huri's engine comparison page sets 3 versions of a pain-rating question side by side, for patients in Nukuʻalofa. A generic engine and Huri's in-house translation model supply raw output. A native linguist signed off the third version. The instruments at stake on a training interface are the termbase entry and the do-not-translate list. An entry marked approved fixes a term for the whole course. The do-not-translate list protects merge fields and product names.

Huri reads 3 points of contact between H-JEPA and the Tā-Vā theory. The first point is collapse. On DROID the encoder kept vā and dropped tā until the inverse-dynamics head restored the action. Kaʻili and Māhina treat tā and vā as inseparable in reality. Their test for a learned level therefore covers the state and the action between 2 states. The second point is segmentation. The H-JEPA authors want variable-duration segments for the upper levels. Tāvāists cut time by event. In tānafa, the beat of the drum fixes the interval. On the Tongan appeal notice, 2 dates fix the vahaʻataimi.

The third point is language. A goal stated in Tongan puts tā in the verb and vā in the noun, the split of tala matangi. An upper level aligned on English goal phrases learns English idioms, the fault the Tongan course reviewer flagged in a fluent sentence. The H-JEPA authors name closed-loop control on a physical robot as their next test. On that robot, the arm's action between 2 camera frames is the tā the DROID encoder dropped.