Sample-efficient fluid intelligence requires agents that rapidly form working models of unfamiliar interactive environments. ARC-AGI-3 tests this across uninstructed visual games scored by Relative Human Action Efficiency, which penalizes exploratory waste quadratically. When frontier models read raw integer grids, serialization destroys spatial topology, and they spend hundreds of actions rediscovering basic boundaries and symmetries. We introduce SARBLOH, an open-weight agent built on the premise that spatial perception, not planning capacity, governs interactive reasoning on ARC-AGI-3. A zero-parameter O(HW) perception engine turns each frame, in 35 ms, into an object inventory with bipartite temporal tracking, a geometric briefing that separates proven invariants from affordance hypotheses, and an annotated canvas. Paired with a code-as-action runtime, the agent writes algorithmic hypotheses instead of low-level moves. Across eight presentations, grounded perception raises the rubric score from 3/28 to 25/28 points while cutting token use by 89.4%. On 23 held-out games, geometric invariants generalize with 100% accuracy, while single-frame affordance guesses fall to 6-50%, so active verification probes remain necessary. On the public set, Run A (briefing off) solved 143 of 183 levels in 8.2 hours at 59.4% single-pass efficiency; Run B (briefing on) reached 63.5%. Within the nine-hour budget on one Kaggle RTX PRO 6000 GPU, SARBLOH scored 16.53 on the hidden Kaggle evaluation, 13.7x the first Milestone winner.
Figures & tables
Figure 1: Action efficiency across 25 public ARC-AGI-3 environments. Sarbloh (open-weight Qwen3.8-Flash-Next on one GPU) leads five frontier API baselines on 21 of 25 games (72.5% peak mean vs. 59.4% single-pass mean; 100% on 13 games). Because the strict 9-hour Kaggle ceiling precludes exhausting all 25 games in one pass, the curve reports the best score per game across runs, including isolated re-evaluations on complex environments ( bp35 ).
Hidden Kaggle set (offline, 1 GPU, 9 h)
Public set (frontier APIs, unbounded)
Sarbloh this work
Duck first Milestone winner
Reki second Milestone winner
Prime Agent open harness
Arcgentica Symbolica
AVO NVIDIA
Provider Adapters GPT-5.5
Provider Adapters Gemini 3.8 Flash
Wall-clock limit
9 h
9 h
9 h
unbounded
unbounded
unbounded
unbounded
unbounded
Network access
none
none
none
yes
yes
yes
yes
yes
Protocol
peak / single pass
–
–
self-reported
self-reported
self-reported
self-reported
self-reported
Open weights
✓
✓
✓
–
–
–
–
–
Backbone
Qwen3.8-Flash-Next
Qwen 3.6 27B
Gemma-4-31B
Claude Opus 5
Claude Opus 4.6
Opus 5 / GPT-5.6
GPT-5.5
Gemini 3.8 Flash
Table 1: ARC-AGI-3 systems across two distinct evaluation regimes. The left group executes offline within the official Kaggle evaluation sandbox on a single GPU under strict hardware and nine-hour wall-clock limits without network access. The right group evaluates proprietary frontier models via unconstrained commercial APIs without standardized verification. The official hidden-set score (16.53) represents the cumulative competition metric ( E , Equation ( 3 )), which aggregates triangular level weights across multiple undisclosed games under the nine-hour ceiling, whereas Figure 1 reports the unweighted per-game efficiency percentage ( Sl , Equation ( 2 )) on public environments. † Self-reported by authors on public games without independent certification.
Figure 2: Controlled ablation of input observation modalities. Action expenditure required to resolve the opening level of ls20 across four representation tiers using an identical frontier reasoning model (Gemini 3.1 Pro, greedy decoding, single exploratory trial n=1 on 12 October 2026). Serializing the environment as raw integer tokens causes budget exhaustion, whereas rendered visual rasters and programmatic geometric relations achieve super-human efficiency ( 6 actions versus human baseline ≈ 10).
Figure 3: One real turn, end to end ( ls20 , level 1, Run B, briefing on, Section 5.2 ; transcript and frames as logged). (a) The map the engine drew for the opening frame. (b) An excerpt of the briefing sent with it: measured facts first, then guesses, each with a one-action test. (c) The agent’s first action in the level, written as Python at turn 7 after six turns spent only reading observe() : one key press, the test that the briefing proposed for you . (d) What the kernel printed: the player (orange and blue, 25 cells) moved one tile of 5 cells up and the HUD lost two cells. (e) The map after the move. One action confirmed two guesses, obj 10 = you and obj 15 = moves left .
Figure 4: Three of the nine lab presentations on the opening frame of ls20 , with rubric score (of 28) and text cost. (a) The Legacy Encoder, text only (excerpt of 183 lines, grid rows cut at 44 columns). (b) The Role Ledger: a role map and, below it, each guess with its evidence and a one-action test. (c) The Fused Briefing, the presentation we kept: a schematic map with object tags, shape threads and tethers, then measured facts and guesses as separate tiers. The other six presentations are in Figure 12 .
Figure 5: One real frame (the opening frame of ls20 ) through the perception engine. (a) The rendered 64×64 frame. (b) View 1: the tracked objects with their ids; one-cell parts and the three HUD copies are outlined but not tagged one by one. (c) View 2: four of the measured relations from (e), drawn as straight lines between the objects they relate (a line is not a path). (d) View 3: the map the model receives: floor in light tones, walls hatched, the HUD moved to a side panel, objects tagged id: ROLE? , look-alike shapes joined by a dashed thread. (e) The briefing for the same frame, word for word, sent as text with the map. The model receives (d) and (e); (b) and (c) are drawn here from the engine’s object table and briefing to show what each view contributes.
Role
Evidence (scored)
One-action test
You
Compact, multi-coloured, unique, on the lattice
Press one arrow: it moves one tile
Goal
Hollow box; shape repeated in the HUD
Bring the matching thing into it
Moves left
Long bar at the screen edge
Any action: it shrinks one unit
Lives
Identical copies in the HUD
Watch it on a failure or reset
Inventory
HUD icon matching a world object
Watch it after the modifier
Modifier
Small symbol, lopsided colours
Step on it: the inventory changes
Table 2: Role guesses. Each is printed with its confidence and a test that settles it in one action. On held-out games these guesses were right about half the time or less (Table 4 ), which is why they are never stated as facts.
Figure 6: Share of each request’s prompt tokens served from the prefix cache (moving mean over 25 requests), for the first 960 requests of two Qwen3.8-Flash-Next runs on SGLang. Before (15 games): the picture is sent as an unstored last message and the GPU slot changes hands every turn, so each request rewrites the end of the one before. After (Run A, 25 games): every request is the previous request plus its reply, and a game keeps its slot for 20 requests. Labels on the right give the whole run.
Presentation
ls20
r11l
Total
Int.
Tokens
/14
/14
/28
1–5
Legacy Encoder (baseline)
2
1
3
1
3500
X1 Exploded Inventory
2
1
3
2
180
X2 Surprise Map
3
1
4
2
270
X3 Visual Routines
7
8
15
4
240
X4 Look-alike Threads
4
2
6
3
140
Table 3: Perception Lab evaluation across candidate presentations. Rubric scores on the opening frames of ls20 and r11l , intuition rating (Int., 1–5), and text token footprint per frame. The Fused Briefing (X8) achieves 25/28 points with an 89.4% reduction in token cost relative to the Legacy Encoder baseline.
Figure 7: Perception Lab rubric score versus token cost. The Fused Briefing establishes an empirical Pareto frontier, scoring 25/28 against 3/28 for the Legacy Encoder while reducing prompt tokens by nearly an order of magnitude.
Check on 23 Held-Out Official Games
Empirical Generalization
Screenshot to grid, exact cells
23/23 games (100%)
Agent identity (X7), keyboard games
6/14 games (43%)
Moves-left gauge found (X7)
10/20 games (50%)
Click handles that respond (X7)
7/29 handles (24%)
Keyboard foresight fully right (X5)
1/16 games (6%)
Table 4: Empirical validation of the Measured–Guessed Rule across 23 held-out environments. Deterministic coordinate recovery and geometric relations generalize unconditionally ( 100% ), whereas latent affordance roles and predictive forward models from a single frame achieve between 6% and 50% accuracy, confirming that operational affordances must be framed as testable hypotheses.
Figure 8: Action efficiency and score error budget. (a) Distribution of agent-to-human action ratios ( al/hl ) across 143 cleared levels (Run A, briefing off), where hatched bars indicate performance surpassing human baselines. (b) Decomposition of the 40.6 total points lost from a perfect score, illustrating that 72% of lost credit stems from uncleared levels stalled by perceptual misinduction rather than action overhead on resolved levels.
Figure 9: Evaluation trajectory across completed submissions on the official held-out Kaggle benchmark. Submissions are ordered by intervention, not by date: each sits in the stage of the last intervention it carried, and every stage has the same width. The solid curve tracks the best score so far, hollow markers indicate regressions and failed experiments, and each label gives the stage’s best score and its gain over the stages before it. The last stage’s submission also changed the level memory and the context drain (Section 5.3 ). The official hidden-set score (16.53, produced with the full perception engine active) is evaluated on the competition-wide RHAE metric across undisclosed games, whereas Figure 1 reports per-game action efficiency percentages on public environments.
Figure 10: Cleared levels over wall-clock execution time on 25 public environments under identical hardware. Indigo: Sarbloh (single pass, Run A, briefing off). Solid grey: distinct levels cleared by the first Milestone winner (Duck harness) across four parallel passes; dotted grey: total levels summed across passes. Sarbloh clears 143 of 183 levels in 8.2 hours, surpassing the baseline’s 129 distinct levels.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Symbol
Meaning
Environment and score (Section 2 )
gt
grid observed at step t , in {0..15}H×W
H,W≤64
grid height and width
A
action set, Equation ( 1 )
n ; l ; k
levels in a game; level index; levels cleared
hl ; al
human baseline actions; agent actions on level l
Appendix
Table 5: Notation.
Figure 11: System topology of Sarbloh . The host-side perception engine extracts invariant relations and affordance hypotheses from each frame, passing views to an append-only context. The reasoning model generates Python search routines in a persistent kernel, committing multi-step paths via await play() . Twenty-five games share ten GPU execution slots under the utility-driven priority scheduler.
Parameter
Value
Games / GPU slots
25 / 10
Scheduler quantum
20 model calls
Action cap per game
2,000 actions
Actions per play() path
at most 100
Reply cap per request / requests per turn
12k tokens / 12
Thinking without a move before the turn yields
16,384 tokens
Appendix
Table 6: Settings of Run A (briefing off) and Run B (briefing on).
Figure 12: Visual presentations evaluated in the Perception Lab on ls20 . (a) Exploded Inventory separates unique entity morphologies. (b) Surprise Map highlights unpredicted local textures. (c) Visual Routines computes lattice BFS step counts to candidate targets. (d) Look-alike Threads connects shapes matching under rotation, reflection, or integer scaling. (e) Foresight Strip simulates counterfactual forward states. (f) Blueprint redraws terrain and lattices as a clean architectural schematic. (g) Role Ledger annotates candidate affordance hypotheses. (h) Fused Briefing combines the blueprint map with measured invariants and testable role hypotheses.
Figure 13: The perception pipeline on the opening frame of ka59 (arrow keys and clicks, 3×3 tile grid). (c) Rings 4 and 11 are identical; ring 4 encloses an empty 3×3 that exactly fits square 7 (and square 6); squares 6 and 7 have the same shape in different colours. Panels as in Figure 5 ; the model receives (d) and (e).
Figure 14: The perception pipeline on the opening frame of m0r0 (arrow keys and clicks). The engine finds four objects and measures two facts: squares 3 and 4 are identical and share rows 49–53. It makes no guesses. Panels as in Figure 5 .
Figure 15: The perception pipeline on the opening frame of r11l (clicks, no tile grid). (b) The dotted ring (4–15) and the two connecting lines (18–24, 28–35) are made of one-cell parts and are tagged once each. (c) Disc 25 sits at the midpoint of handles 16 and 36, and fits the empty 5×5 inside ring 4. Panels as in Figure 5 .
Figure 16: The perception pipeline on the opening frame of sb26 (clicks only). (c) Squares 20–23 are the grey squares 13–16 drawn 2 × bigger, and ring 12 sits directly above each of them. (e) The briefing reaches its line limit; the last MEASURED line says eight more facts can be read from observe().briefing . Panels as in Figure 5 .
Figure 17: The briefing line budget on the first frame of tr87 , both boxes built by the agent’s own Briefing code (object descriptions shortened; r / c = shared rows/columns). The mark at the left of each line gives its kind: enclosure, alignment, look-alike, guess. (a) One cap of 30 lines over all facts in the engine’s order: 23 of the 30 lines are alignments, and the look-alike pair and both guesses are cut. (b) The policy the agent runs, which shares the lines between fact kinds and keeps up to a third for guesses: it shows the look-alike pair and the two guesses and counts the 40 facts left out (“ +40 more facts ”).
Run A: 25 games, one GPU, 8.2 h
Value
Prompt tokens served from the prefix cache
98.4% of 1.78 B
Decode speed per request
76 tok/s
Busy time: decode / prefill
50.1% / 1.3%
Model calls / server failures
21,224 / 0
Appendix
Table 7: Systems telemetry of Run A (briefing off).
25 public games, one GPU
Run A
Run B
Logged (2026)
23 Oct
27 Oct
Configuration
Briefing and map
off
on
Extra guardrails, stuck review
–
✓
Context drain target
55k
70k
Wall clock
8.2 h
5.4 h
Appendix
Table 8: The two full runs nearest to a perception ablation. They differ in more than the briefing, so the comparison is suggestive only. Level counts are from each run’s own log of clear times.