World Alignment: unpaired cross-modal correspondence by relational identifiability

Method: scene states are sets of part states; relation fields are built
within each modality and are invariant to how each side labels its own
features; the cross-modal bridge is a coupling searched under an energy
that is a closed-form functional of one matrix; solving is spectral
initialisation followed by exact local refinement.

Evidence: in a procedurally generated closed world, blind recovery of a
hidden image-caption correspondence reaches 95.3% at 256 scenes against
0.39% chance, and the recovered pairs transfer to 200 held-out scenes at
93.0% exact retrieval with random-pair and shuffled-image controls at or
near chance. Cross-modal value correspondence is derived from disjoint
corpora rather than declared. On Visual Genome the field correlation
reaches 0.656 against the 0.9 that polynomial recovery needs, with the
deficit attributed away from segmentation and discretisation.

Protocol: no image-text pair enters any objective, optimiser,
initialisation, or model selection; hidden pairs score orderings only.

Co-Authored-By: Claude <noreply@anthropic.com>

1 files changed, 108 insertions, 0 deletions

diff --git a/worldalign/extract_text_orbits.py b/worldalign/extract_text_orbits.py
new file mode 100644
index 0000000..43b9020
--- /dev/null
+++ b/worldalign/extract_text_orbits.py
@@ -0,0 +1,108 @@
+from __future__ import annotations
+
+import argparse
+from pathlib import Path
+
+from datasets import load_dataset
+import torch
+from tqdm import tqdm
+from transformers import AutoModel, AutoTokenizer
+
+from .common import batch_indices, dtype_for_device, read_json
+from .extract_text import mean_pool
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--manifest", default="artifacts/manifest.json")
+    parser.add_argument(
+        "--output", default="artifacts/text_orbits_qwen0p5b.pt"
+    )
+    parser.add_argument("--model", default="Qwen/Qwen2.5-0.5B")
+    parser.add_argument("--device", default="cuda:3")
+    parser.add_argument("--batch-size", type=int, default=128)
+    parser.add_argument("--max-length", type=int, default=64)
+    parser.add_argument("--layer", type=int, default=-1)
+    parser.add_argument(
+        "--row-groups",
+        default="text_only_train,val,test",
+        help="Comma-separated manifest row lists to encode.",
+    )
+    parser.add_argument("--limit", type=int)
+    return parser.parse_args()
+
+
+@torch.inference_mode()
+def main() -> None:
+    args = parse_args()
+    manifest = read_json(args.manifest)
+    groups = [group.strip() for group in args.row_groups.split(",")]
+    rows = list(
+        dict.fromkeys(
+            int(row)
+            for group in groups
+            for row in manifest[group]
+        )
+    )
+    if args.limit:
+        rows = rows[: args.limit]
+    dataset = load_dataset(
+        manifest["dataset"], split=manifest["dataset_split"]
+    ).remove_columns("image")
+    captions = [dataset[row]["caption"] for row in rows]
+    views = len(captions[0])
+    if any(len(items) != views for items in captions):
+        raise ValueError("Every text orbit must have the same view count")
+    flat_text = [text for items in captions for text in items]
+
+    tokenizer = AutoTokenizer.from_pretrained(args.model)
+    if tokenizer.pad_token_id is None:
+        tokenizer.pad_token = tokenizer.eos_token
+    tokenizer.padding_side = "right"
+    dtype = dtype_for_device(args.device)
+    model = AutoModel.from_pretrained(
+        args.model, torch_dtype=dtype
+    ).to(args.device)
+    model.eval()
+
+    output: list[torch.Tensor] = []
+    for indices in tqdm(
+        batch_indices(len(flat_text), args.batch_size),
+        desc="Qwen text orbits",
+    ):
+        tokens = tokenizer(
+            [flat_text[index] for index in indices],
+            padding=True,
+            truncation=True,
+            max_length=args.max_length,
+            return_tensors="pt",
+        )
+        tokens = {key: value.to(args.device) for key, value in tokens.items()}
+        result = model(
+            **tokens, output_hidden_states=True, return_dict=True
+        )
+        output.append(
+            mean_pool(
+                result.hidden_states[args.layer],
+                tokens["attention_mask"],
+            )
+            .float()
+            .cpu()
+        )
+    features = torch.cat(output).reshape(len(rows), views, -1)
+    state = {
+        "model": args.model,
+        "layer": args.layer,
+        "rows": rows,
+        "features": features,
+        "captions": captions,
+        "views_per_orbit": views,
+        "row_groups": groups,
+    }
+    Path(args.output).parent.mkdir(parents=True, exist_ok=True)
+    torch.save(state, args.output)
+    print(f"Wrote {args.output}: {tuple(features.shape)}")
+
+
+if __name__ == "__main__":
+    main()