rogermt commited on
Commit
a57fa41
·
verified ·
1 Parent(s): 8cf6976

Update results: 70/400 (17.5%) — ITT:16 + Predicate:25 + DSL:29, zero regressions

Browse files
Files changed (1) hide show
  1. arc_results/RESULTS.md +24 -56
arc_results/RESULTS.md CHANGED
@@ -1,66 +1,34 @@
1
  # PEMF Solver — ARC-AGI Training Set Evaluation
2
 
3
- ## Results (v3 — ITT physics + DSL beam)
4
 
5
- | Metric | v1 | v2 | **v3** |
6
- |---|---|---|---|
7
- | **Tasks solved** | 31 (7.8%) | 40 (10.0%) | **47 (11.8%)** |
8
- | via ITT | — | — | **16** |
9
- | via DSL | 31 | 40 | **31** |
10
- | Transforms (DSL) | 19 | 33 | 33 |
11
- | ITT rule types | | | 7 |
12
- | Total time | 17s | 51s | **36s** |
13
- | Regressions | — | 0 | **0** |
14
 
15
- ## 7 Newly Solved Tasks (ITT-exclusive)
16
 
17
- | Task ID | ITT Rule Type | What it does |
18
  |---|---|---|
19
- | 0d3d703e | recolor | Direct color substitution |
20
- | 868de0fa | multi_region_fill | Frame size fill color mapping |
21
- | 8d5021e8 | tile | Tile with learned reflection pattern |
22
- | b1948b0a | recolor | Color remap |
23
- | c0f76784 | multi_region_fill | Frame size → fill color with fallback |
24
- | c8f0f002 | recolor | Color substitution |
25
- | d511f180 | recolor | Color remap |
26
 
27
- **None of these 7 tasks were solvable by the DSL beam search.** They require analyzing training pairs *together* to learn the rule — the ITT engine does this via `TransformationRule.learn()`.
28
 
29
- ## Architecture
30
-
31
- ### ITT-first, DSL-fallback
32
- 1. **ITT solver** tries first: learns rule from training pairs using σ-analysis, ρ_q boundary charge, and field invariants
33
- 2. If ITT achieves σ=0 on ALL training pairs → use it (16 tasks)
34
- 3. Otherwise → **DSL beam search** with 33 transforms + greedy stacker (31 tasks)
35
-
36
- ### ITT Physics Engine (`itt_engine.py`)
37
- - **PhiField**: dual-field Φ_q (int semantics) + Φ̃ (smooth float operators)
38
- - **ρ_q = |∇(∇²Φ̃)|**: boundary charge with physics-derived threshold (μ+1.5σ)
39
- - **SigmaResidue**: classifies transformations as fill/expansion/compression/recolor/erase
40
- - **Fan Signature**: 6-bit [Δ₁..Δ₆] task routing
41
- - **TransformationRule.learn()**: learns tile_pattern, size_to_color, frame_to_fill, color_map, shape_to_color from training pairs
42
- - **FieldInvariants**: enclosed mask, frame components, shape eigenspectrum, Fourier period detection
43
-
44
- ### ITT Rule Types
45
- | Rule | Description |
46
- |---|---|
47
- | tile | Tile with learned per-block reflection/rotation pattern |
48
- | self_tile | Kronecker: input nonzero mask as meta-layout |
49
- | fill_enclosed | Fill enclosed regions via ρ_q boundary detection |
50
- | multi_region_fill | Frame size → fill color with fallback chain |
51
- | periodic_extension | Fourier period detection + color remap + tile |
52
- | shape_indicator | Laplacian eigenspectrum → output color |
53
- | recolor | Direct color substitution learned from pairs |
54
-
55
- ## How to reproduce
56
- ```bash
57
- git clone https://github.com/fchollet/ARC-AGI.git /tmp/arc
58
- cp -r /tmp/arc/data/training arc_data/training
59
- python scripts/run_all_arc.py # DSL only (40 tasks)
60
- python scripts/run_all_arc_v3.py # ITT + DSL (47 tasks)
61
  ```
 
 
 
 
 
 
 
62
 
63
- ## References
64
- - Original ITT solver: [Sensei-Intent-Tensor/0.0_ARC_AGI](https://github.com/Sensei-Intent-Tensor/0.0_ARC_AGI)
65
- - ITT physics: [Sensei-Intent-Tensor/0.0._Executable_Physics](https://github.com/Sensei-Intent-Tensor/0.0._Executable_Physics)
66
- - Zenodo: [https://zenodo.org/records/18077258](https://zenodo.org/records/18077258)
 
1
  # PEMF Solver — ARC-AGI Training Set Evaluation
2
 
3
+ ## Results (v4 — ITT + Predicate + DSL)
4
 
5
+ | Metric | v1 | v2 | v3 | **v4** |
6
+ |---|---|---|---|---|
7
+ | **Tasks solved** | 31 (7.8%) | 40 (10.0%) | 47 (11.8%) | **70 (17.5%)** |
8
+ | via ITT | — | — | 16 | **16** |
9
+ | via Predicate | | | — | **25** |
10
+ | via DSL | 31 | 40 | 31 | **29** |
11
+ | Total time | 17s | 51s | 36s | **38s** |
12
+ | Regressions | | 0 | 0 | **0** |
 
13
 
14
+ ## Predicate Engine Breakdown (25 new solves)
15
 
16
+ | Rule Type | Tasks | Description |
17
  |---|---|---|
18
+ | neighborhood_rule | 20 | CA-style: (center_color, neighbor_signature) → output_color |
19
+ | global_enclosed_fill | 2 | Fill all bg regions not reachable from border |
20
+ | object predicate×action | 2 | E.g. "remove smallest object" |
21
+ | per_object_enclosed_fill | 1 | Fill each object's interior with its color |
 
 
 
22
 
23
+ ## Architecture: 3-Pass Pipeline
24
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
25
  ```
26
+ Task → ITT Physics → Predicate Enumeration → DSL Beam Search
27
+ (16 tasks) (25 tasks) (29 tasks)
28
+ ```
29
+
30
+ 1. **ITT** (PhiField + σ-analysis + Fan Signatures → rule learning)
31
+ 2. **Predicate** (enclosed fill → neighborhood rules → object predicate×action)
32
+ 3. **DSL** (33 transforms + dual-strategy beam + greedy stacker)
33
 
34
+ Each pass only runs if the previous one fails. Zero regression risk.