infer.py: render the exact eval prompts, disable thinking, pass video metadata

- Sub-questions use the q1:/q2: numbering and the per-law questions from the PhyGround repo (evals/sub_questions.py); criterion text matches evals/physics_criteria.py. Rendered prompts are now identical to .
- Qwen3.5 thinking is turned off (enable_thinking=False); the judge answers with bare JSON.
- Pass video_metadata from qwen-vl-utils>=0.0.14 so frame timestamps are correct (previously defaulted to 24 fps) and fix the crash on transformers 5.x from the per-video fps list.
- README: require qwen-vl-utils>=0.0.14 and explain how infer.py relates to the paper's vLLM pipeline.

Context: https://github.com/NU-World-Model-Embodied-AI/PhyGround/issues/1

README.md

@@ -73,15 +73,16 @@ dimensions plus 2–4 laws associated with the prompt.
 pip install -U \
   "transformers>=5.2.0" \
   "peft>=0.19.1" accelerate pyyaml \
-  "qwen-vl-utils[decord]" \
+  "qwen-vl-utils[decord]>=0.0.14" \
   huggingface_hub
 
 hf download NU-World-Model-Embodied-AI/phyjudge-9B \
   --local-dir ./phyjudge-9B
 ```
 
-The card's prompt/parser path was checked with Transformers 5.2.0 and PEFT
-0.19.1. Loading the 9B base model in bf16 needs roughly 24 GB of GPU memory in the
+`infer.py` was checked end to end (video in, score out) with Transformers
+5.2.0 and qwen-vl-utils 0.0.14; the adapter was saved with PEFT 0.19.1.
+Loading the 9B base model in bf16 needs roughly 24 GB of GPU memory in the
 authors' setup. Actual memory use depends on software versions, video length,
 frame sampling, and resolution. Reduce `--fps` or `--max-pixels` if needed;
 doing so can also change the score.
@@ -110,7 +111,13 @@ python ./phyjudge-9B/infer.py \
 
 The script loads the base model recorded in `adapter_config.json`, attaches the
 LoRA adapter, samples the video at 2 FPS by default, and performs deterministic
-decoding. Output is a JSON object:
+decoding with Qwen3.5 thinking turned off. It renders exactly the same prompts
+as `python -m evals.vlm_eval --prompt_config subq+human.yaml` in the
+[evaluation code](https://github.com/NU-World-Model-Embodied-AI/PhyGround).
+The paper's numbers come from that repo's vLLM pipeline
+(`scripts/score_videos.sh`), which samples at temperature 1.0, so a single
+`infer.py` score can differ from it; use the GitHub pipeline to reproduce the
+benchmark. Output is a JSON object:
 
 ```json
 {

infer.py

@@ -32,6 +32,10 @@ from peft import PeftModel
 from transformers import AutoProcessor
 
 
+# The tables below mirror evals/prompts/__init__.py (GENERAL_SUB_QUESTIONS),
+# evals/physics_criteria.py (CRITERIA) and evals/sub_questions.py (SUB_QUESTIONS)
+# in https://github.com/NU-World-Model-Embodied-AI/PhyGround, so the rendered
+# prompts match `python -m evals.vlm_eval --prompt_config subq+human.yaml`.
 GENERAL_SUB_QUESTIONS: dict[str, list[str]] = {
     "SA": [
         "Are the main objects in the caption present in the video?",
@@ -58,12 +62,12 @@ PHYSICAL_CRITERIA: dict[str, str] = {
     "gravity": "Do unsupported objects fall downward? Do thrown objects follow a curved trajectory? Does poured liquid fall with gravity?",
     "inertia": "Do stationary objects remain still unless acted upon? Do moving objects maintain their motion unless stopped by friction, collision, or an obstacle?",
     "momentum": "After collision, push, or pull, is the direction of motion reasonable? Ignore speed magnitude.",
-    "impenetrability": "Do objects maintain impenetrability -- no passing through each other?",
+    "impenetrability": "Do objects maintain impenetrability — no passing through each other?",
     "collision": "After impact, is there reasonable bounce/shatter/deformation? Does response match impact force?",
     "material": "Does each material respond according to its properties? (glass shatters, rubber bounces, metal is rigid, cloth deforms softly, etc.)",
     "buoyancy": "Do dense objects sink? Do wood/plastic float?",
     "displacement": "When you add more liquid or put an object into it, does the liquid level rise in a realistic way? Does it overflow when full?",
-    "flow_dynamics": "Does the liquid's overall motion behave realistically over time -- flowing along surfaces, spreading, draining naturally?",
+    "flow_dynamics": "Does the liquid's overall motion behave realistically over time — flowing along surfaces, spreading, draining naturally?",
     "boundary_interaction": "When the liquid hits a boundary such as a rock face, container wall, or floor, does it respond realistically? Do local splash, rebound, or split patterns on impact look physically plausible?",
     "fluid_continuity": "Does the liquid avoid disappearing or appearing out of nowhere? Small splashes that briefly break apart are okay.",
     "reflection": "Does the reflection roughly match objects and colors in the scene, and avoid completely unrelated content?",
@@ -73,69 +77,62 @@ PHYSICAL_CRITERIA: dict[str, str] = {
 
 PHYSICAL_SUB_QUESTIONS: dict[str, list[str]] = {
     "gravity": [
-        "Do unsupported objects or liquids move downward over time?",
-        "Do thrown or falling objects follow a plausible gravity-driven path?",
-        "Does the video avoid objects floating or rising without support?",
+        "Do any unsupported objects float upward or hover in mid-air?",
+        "Does any heavy object drift or float down unrealistically slowly, as if it had no weight?",
+        "Does any object fall sideways or in an unnatural direction instead of downward?",
     ],
     "inertia": [
-        "Do stationary objects remain still unless a visible force acts on them?",
-        "Do moving objects continue plausibly until friction, collision, or an obstacle changes their motion?",
-        "Does the video avoid unexplained starts, stops, or direction changes?",
+        "Does any object spontaneously start moving without a visible cause?",
+        "Does any moving object suddenly stop or reverse direction with no visible contact or force?",
     ],
     "momentum": [
-        "After contact, push, pull, or collision, are motion directions plausible?",
-        "Does the reacting object move in a direction consistent with the interaction?",
-        "Does the video avoid impossible reversals or unrelated motion changes?",
+        "Does any object fly off in a direction completely unrelated to how it was hit or pushed?",
+        "Does the recoil direction contradict the direction of the applied force?",
     ],
     "impenetrability": [
-        "Do solid objects avoid passing through one another?",
-        "Do contacts and overlaps remain physically plausible?",
-        "Does the video avoid obvious clipping or penetration artifacts?",
+        "Does any object pass through another solid object as if it were not there?",
+        "Does any part of an object clip into a wall, floor, or another object's interior?",
     ],
     "collision": [
-        "Does impact cause a plausible bounce, break, deformation, or transfer of motion?",
-        "Is the response direction consistent with the collision?",
-        "Does the response avoid being much too weak, too strong, or unrelated to the impact?",
+        "Does any object remain completely unaffected after being clearly hit by another object?",
+        "Does the collision response look wildly too weak or too strong for the visible impact?",
+        "Does any object shatter or deform dramatically from a very light touch with no reasonable force?",
     ],
     "material": [
-        "Do objects respond consistently with their apparent material?",
-        "Are rigid, soft, brittle, elastic, or fluid-like objects animated appropriately?",
-        "Does the video avoid material behavior that contradicts the scene?",
+        "Does any rigid material (glass, metal, stone) bend or stretch like rubber?",
+        "Does any soft material (cloth, rubber, rope) behave as if it were completely rigid and unbending?",
     ],
     "buoyancy": [
-        "Do objects sink or float in a way consistent with apparent density?",
-        "Does the floating or sinking behavior stay stable over time?",
-        "Does the video avoid unsupported hovering or impossible underwater motion?",
+        "Does any heavy object (metal, stone) float on the liquid surface?",
+        "Does any light object (wood, cork, plastic) sink to the bottom?",
     ],
     "displacement": [
-        "Does liquid level rise when volume is added or an object enters it?",
-        "Does overflow happen only when the container is plausibly full?",
-        "Does the liquid volume remain visually plausible?",
+        "Does the liquid level remain completely unchanged when a large object is submerged?",
+        "Does a full container spill or overflow when more liquid or an object is added?",
+        "Does the liquid level behave in a clearly impossible way, such as dropping when volume is added?",
     ],
     "flow_dynamics": [
-        "Does liquid flow along surfaces, spread, or drain naturally?",
-        "Does the flow direction follow gravity and boundaries?",
-        "Does the video avoid abrupt stops, reversals, or unsupported uphill flow?",
+        "Does liquid flow uphill or against gravity without any force pushing it?",
+        "On a flat surface, does liquid spread outward and become thinner over time?",
+        "Does liquid suddenly stop flowing or freeze in place without an obvious reason?",
     ],
     "boundary_interaction": [
-        "Does liquid react plausibly when hitting a wall, floor, container, or obstacle?",
-        "Are splash, rebound, or split patterns locally plausible?",
-        "Does the liquid remain consistent after interacting with boundaries?",
+        "Does liquid ignore a solid boundary and continue moving as if nothing were there?",
+        "Does liquid striking a surface produce visible droplets, spray, or ripples?",
+        "Does liquid accumulate or pool on the wrong side of a barrier?",
     ],
     "fluid_continuity": [
-        "Does liquid avoid disappearing or appearing without cause?",
-        "Does the amount of liquid remain broadly consistent?",
-        "Are splashes and separations temporary and physically plausible?",
+        "Does a continuous pour or flow stay connected as a stream without sudden gaps?",
+        "Does liquid disappear into nothing (not counting brief splashes)?",
+        "Does liquid appear from nowhere with no visible source?",
     ],
     "reflection": [
-        "Does the reflection match nearby objects, colors, and motion?",
-        "Does the reflected content stay spatially consistent with the scene?",
-        "Does the video avoid unrelated or impossible reflection content?",
+        "Do reflections show completely unrelated content not present in the scene?",
+        "Does a reflection remain completely static while the scene clearly changes around it?",
     ],
     "shadow": [
-        "Are shadows consistent with the apparent light source direction?",
-        "Do shadows move with the objects that cast them?",
-        "Does the video avoid missing, detached, or contradictory shadows?",
+        "Do different shadows in the same scene point in contradictory directions?",
+        "Does any shadow remain fixed in place while its object clearly moves?",
     ],
 }
 
@@ -151,7 +148,7 @@ def load_yaml(path: Path) -> dict[str, Any]:
 
 
 def questions_block(questions: list[str]) -> str:
-    return "\n".join(f"{idx}. {question}" for idx, question in enumerate(questions, 1))
+    return "\n".join(f"q{idx}: {question}" for idx, question in enumerate(questions, 1))
 
 
 def build_prompt(
@@ -257,18 +254,21 @@ def prepare_inputs(
     fps: float,
     max_pixels: int,
 ) -> dict[str, Any]:
+    # Qwen3.5 thinks by default; the judge answers with bare JSON, so turn it off
+    # (same as `evals.vlm_eval --no-thinking`).
     text = processor.apply_chat_template(
         messages,
         tokenize=False,
         add_generation_prompt=True,
+        enable_thinking=False,
     )
 
     try:
         from qwen_vl_utils import process_vision_info
     except ImportError as exc:
         raise ImportError(
-            "qwen-vl-utils is required for local video inference. "
-            "Install it with: pip install qwen-vl-utils[decord]"
+            "qwen-vl-utils>=0.0.14 is required for local video inference. "
+            "Install it with: pip install 'qwen-vl-utils[decord]>=0.0.14'"
         ) from exc
 
     for msg in messages:
@@ -279,19 +279,26 @@ def prepare_inputs(
                     item.setdefault("fps", fps)
                     item.setdefault("max_pixels", max_pixels)
 
-    try:
-        image_inputs, video_inputs, video_kwargs = process_vision_info(
-            messages,
-            return_video_kwargs=True,
-        )
-    except TypeError:
-        image_inputs, video_inputs = process_vision_info(messages)
-        video_kwargs = {}
+    # Qwen3.5 uses the Qwen3-VL video processor: frames are resized to multiples
+    # of its patch size (16), and the processor needs per-video metadata to put
+    # the right frame timestamps in the prompt (without it, it assumes 24 fps).
+    patch_size = getattr(getattr(processor, "video_processor", None), "patch_size", 16)
+    image_inputs, video_inputs, video_kwargs = process_vision_info(
+        messages,
+        return_video_kwargs=True,
+        return_video_metadata=True,
+        image_patch_size=patch_size,
+    )
+    video_metadata = None
+    if video_inputs is not None:
+        video_inputs, video_metadata = (list(x) for x in zip(*video_inputs))
 
     inputs = processor(
         text=[text],
         images=image_inputs,
         videos=video_inputs,
+        video_metadata=video_metadata,
+        do_resize=False,
         padding=True,
         return_tensors="pt",
         **video_kwargs,