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, 93 insertions, 0 deletions

diff --git a/worldalign/evaluate_prefix.py b/worldalign/evaluate_prefix.py
new file mode 100644
index 0000000..2d53a51
--- /dev/null
+++ b/worldalign/evaluate_prefix.py
@@ -0,0 +1,93 @@
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from .common import dtype_for_device, read_json, write_json
+from .evaluate import unigram_f1
+from .io import load_feature_pair, select_rows
+from .models import load_prefix
+
+
+def parse_args() -> argparse.Namespace:
+    p = argparse.ArgumentParser()
+    p.add_argument("--manifest", default="artifacts/manifest.json")
+    p.add_argument("--vision", default="artifacts/vision.pt")
+    p.add_argument("--text", default="artifacts/text.pt")
+    p.add_argument("--prefix", default="artifacts/prefix.pt")
+    p.add_argument("--split", choices=["val", "test"], default="val")
+    p.add_argument("--device", default="cuda:1")
+    p.add_argument("--samples", type=int, default=100)
+    p.add_argument("--max-new-tokens", type=int, default=32)
+    p.add_argument("--output", default="artifacts/prefix_evaluation.json")
+    return p.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+    manifest = read_json(args.manifest)
+    _, text, _, tlookup = load_feature_pair(args.vision, args.text)
+    rows = manifest[args.split][: args.samples]
+    semantic = select_rows(text["features"], tlookup, rows)
+    caption_lookup = {
+        int(row): caps for row, caps in zip(text["rows"], text["all_captions"])
+    }
+    prefix, prefix_state = load_prefix(args.prefix, args.device)
+    tokenizer = AutoTokenizer.from_pretrained(text["model"])
+    if tokenizer.pad_token_id is None:
+        tokenizer.pad_token = tokenizer.eos_token
+    dtype = dtype_for_device(args.device)
+    lm = AutoModelForCausalLM.from_pretrained(
+        text["model"], torch_dtype=dtype
+    ).to(args.device)
+    lm.eval()
+
+    generated = []
+    with torch.inference_mode():
+        for chunk in semantic.split(16):
+            embeds = prefix(chunk.to(args.device)).to(dtype)
+            attention = torch.ones(
+                embeds.shape[:2], dtype=torch.long, device=args.device
+            )
+            output = lm.generate(
+                inputs_embeds=embeds,
+                attention_mask=attention,
+                max_new_tokens=args.max_new_tokens,
+                do_sample=False,
+                eos_token_id=tokenizer.eos_token_id,
+                pad_token_id=tokenizer.pad_token_id,
+            )
+            generated.extend(tokenizer.batch_decode(output, skip_special_tokens=True))
+
+    records = []
+    for row, caption in zip(rows, generated):
+        references = caption_lookup[int(row)]
+        records.append(
+            {
+                "row": int(row),
+                "generated": caption,
+                "references": references,
+                "unigram_f1": unigram_f1(caption, references),
+            }
+        )
+    result = {
+        "split": args.split,
+        "samples": len(records),
+        "mean_best_reference_unigram_f1": sum(
+            r["unigram_f1"] for r in records
+        )
+        / max(len(records), 1),
+        "examples": records[:25],
+        "training": prefix_state["training"],
+    }
+    Path(args.output).parent.mkdir(parents=True, exist_ok=True)
+    write_json(args.output, result)
+    print(json.dumps(result, indent=2, ensure_ascii=False))
+
+
+if __name__ == "__main__":
+    main()