agents · review · first seen Nov 5, 10:32 UTC

SARBLOH turns game frames into spatial object inventories

SARBLOH's paper announces a frame-to-object perception layer for an agent tested on ARC-AGI-3 visual games. Builders should care because the reported gains come from changing the agent's input representation, not adding planning capacity (616, 617).

L0 · primary1 source5 claims checked4 min read

What shipped

The primary artifact announces SARBLOH as an open-weight agent. Its zero-parameter, O(HW) perception engine converts each frame into an object inventory in 35 ms. The paper describes bipartite temporal tracking, a geometric briefing that separates proven invariants from affordance hypotheses, and an annotated canvas. Those outputs serve different purposes: the inventory describes a frame, tracking relates objects across frames, and the briefing distinguishes stable geometry from guesses about what an object might allow. The supplied excerpt does not give installation steps, so this is an announced architecture, not an integration guide. [1]

The agent pairs that perception layer with a code-as-action runtime. Instead of issuing only low-level moves against raw integer grids, it writes algorithmic hypotheses. The paper's rationale is specific: serializing a grid loses spatial topology, leaving a model to reconstruct boundaries and symmetries through interaction. ARC-AGI-3 presents uninstructed visual games and scores Relative Human Action Efficiency, which the paper says penalizes exploratory waste quadratically. SARBLOH puts a spatial description in front of the agent before it decides what to do. [1][2]

Across eight presentations, the primary artifact reports that grounded perception raised the rubric score from 3/28 to 25/28 points while cutting token use by 89.4%. These are results for that reported comparison, not a forecast for another agent or task. The supplied excerpt does not give per-presentation results or a separate accounting of runtime costs. I would use the figures to justify a controlled test, not as an expected return from adopting the architecture. [1][2]

Why it matters if you build

For an agent that receives visual game frames as integer grids, the design choice is whether to spend model tokens recovering layout or to provide an object inventory first. The announced 35 ms frame conversion measures that preprocessing stage; it is not an end-to-end action-latency measurement. If your workflow serializes visual state, compare decisions on the same frames with raw grids and with tracked objects. Keep the action policy comparable so the representation, rather than a different planner, is the main variable under test. [1][2]

The boundary between invariants and hypotheses matters more than the inventory alone. On 23 held-out games, the paper reports 100% accuracy for generalized geometric invariants but 6-50% for single-frame affordance guesses. Do not promote a plausible use for an object to a fact after one frame. Let geometry inform the working map, then use interaction to test an uncertain affordance before relying on it. The supplied material does not show which guesses failed or whether further observations resolve them. [1][2]

Token use is a cost signal, not a complete cost model. The reported 89.4% reduction accompanies the rubric change across eight presentations, while conversion of a frame takes 35 ms. The excerpt does not price inference, quantify total elapsed time, or isolate the contribution of the code-as-action runtime from perception. Measure tokens, actions, and elapsed time together. An agent that uses fewer tokens but still spends actions testing poor affordance guesses may not meet the needs of your workflow. [1][2][3][4]

How to try it today

1. Take a visual-game frame your agent would otherwise serialize as an integer grid. Describe its visible objects as an inventory, and retain the original frame for comparison (616).
2. Across successive frames, track corresponding objects. Write a briefing that labels geometric invariants separately from tentative affordances (616, 619, 620).
3. Present the inventory and an annotated canvas before action selection. If using the described runtime, have the agent write an algorithmic hypothesis rather than only low-level moves (616).
4. Test a tentative affordance through interaction instead of accepting a single-frame guess as established (620).
5. Compare rubric score, token use, actions, and elapsed time for raw-grid and grounded-perception runs. Do not assume the reported eight-presentation result will transfer (616, 617, 618).

What we checked

Every claim below was walked down the source ladder. The rung is how close the source is to the thing itself; the tense is the source's own.

1. announced L0 · primary

   SARBLOH converts each frame into an object inventory in 35 ms.

   per arXiv cs.CL/cs.AI/cs.LG (L0)
2. announced L0 · primary

   Across eight presentations, grounded perception raised the rubric score from 3/28 to 25/28 points.

   per arXiv cs.CL/cs.AI/cs.LG (L0)
3. announced L0 · primary

   Across eight presentations, grounded perception cut token use by 89.4%.

   per arXiv cs.CL/cs.AI/cs.LG (L0)
4. announced L0 · primary

   On 23 held-out games, geometric invariants generalized with 100% accuracy.

   per arXiv cs.CL/cs.AI/cs.LG (L0)
5. announced L0 · primary

   Single-frame affordance guesses fell to 6-50% on 23 held-out games.

   per arXiv cs.CL/cs.AI/cs.LG (L0)

What is not confirmed

The supplied material does not establish performance outside the eight presentations or the 23 held-out games. Accuracy for geometric invariants is reported separately from accuracy for single-frame affordances; neither figure proves dependable action selection in a new game. The excerpt also supplies no distribution of failures, independent replication, installation procedure, or end-to-end cost breakdown. Describing the agent as open-weight does not, by itself, confirm a usable download or a maintained integration path. [1][2][3][4][5]

Who should act, who should wait

Act now if you build agents that read game-like visual grids and can run a controlled comparison. Prototype the object inventory, temporal tracking, and invariant-versus-affordance briefing around your current action loop. The reported rubric and token changes make that test worth considering; check the frame-conversion measurement in your own setting. Wait before treating SARBLOH as a drop-in system, or budgeting around its reported gains, if you need documented integration, predictable end-to-end latency, or reliable affordance judgments from one frame. The supplied claims do not establish those properties. [1][2][3][4][5]

Sources

• Transforming Integer Grids to Spatial Intuition: Perception, Not Planning, Governs ARC-AGI-3 — arXiv cs.CL/cs.AI/cs.LG L0 · primary