PhysiXFails: Evaluating Physical Anomalies

Updated 5 July 2026

PhysiXFails is a domain defining and benchmarking physically implausible events in dynamic visuals and physics-engine simulations.
It categorizes failures like gravity breaches, impossible trajectories, temporal discontinuities, and material inconsistencies across diverse applications.
The field leverages fine-tuned multimodal detectors and physics-informed generation techniques to enhance video realism and simulation reliability.

PhysiXFails denotes observable violations of expected physical behavior in dynamic visual content and simulated systems. In recent work, the term is used both generically—for failures such as improbable mid-air rotations, implausible landings, levitation, interpenetration, temporal discontinuities, and object disappearance—and specifically as the name of a benchmark for runtime videos from physics-engine based software. Across text-to-video generation, gameplay analysis, multimodal physics reasoning, and black-box testing of physics engines, the core issue is whether motions, interactions, and state transitions remain consistent with forces, constraints, material properties, and temporal causality (Shao et al., 19 May 2025, Wang et al., 1 Dec 2025, Li et al., 29 Jul 2025).

1. Formal definition and problem setting

A physically plausible video is defined as one in which motions, interactions, and state transitions of objects over time are consistent with core physical laws under the depicted conditions. In the formulation used by PhyDetEx, trajectories $r(t)$ , velocities $v(t)$ , and accelerations $a(t)$ should be compatible with forces and constraints present in the scene, including Newton’s second law $F = m a$ , basic kinematics, momentum and impulse, conservation laws, collisions, friction, drag, buoyancy, and fluid continuity (Wang et al., 1 Dec 2025).

This definition is deliberately stricter than generic perceptual realism. Physion-Eval treats a physical realism failure as any visually observable motion, interaction, state change, or causal event in a generated video that violates real-world physical constraints, and explicitly distinguishes physics-critical plausibility from general perceptual quality or production polish. A video may therefore look visually polished while still containing a physically implausible contact event, force response, or causal chain (Zhang et al., 20 Mar 2026).

In black-box testing of physics-engine based software, the same problem is framed as a detection task over observable outputs rather than engine internals. Given observed frames $O = (o_1, \ldots, o_T)$ , a detector $D(O) \rightarrow \{\text{true}, \text{false}\} \times B'$ determines whether any physics failure is present and, if so, which categories $B' \subseteq B$ apply. This formulation is central to the PhysiXFails benchmark for runtime failure hunting in PE-based systems (Li et al., 29 Jul 2025).

2. Taxonomic structure of physics failures

Recent benchmarks converge on a recurring set of failure families: gravity and support violations, impossible trajectories, collision and contact errors, temporal continuity failures, object permanence failures, material and state inconsistencies, optical inconsistencies, and anatomical or biomechanical breakdowns. The taxonomies differ in granularity, but they largely isolate the same observable breakdowns from different application viewpoints (Wang et al., 1 Dec 2025, Cao et al., 2024, Zhang et al., 20 Mar 2026, Li et al., 29 Jul 2025).

Source	Taxonomic structure	Representative failures
PhyDetEx	Gravity violations; impossible trajectories; interpenetration and rigid-body constraints; conservation law violations; fluid and cloth dynamics inconsistencies; temporal continuity/causality; object permanence; contact and support errors	Hovering bodies, discontinuous paths, solids passing through solids
PhysGame	4 domains and 12 physical commonsense categories	Gravity, elasticity, friction, velocity, acceleration, reflection, refraction, absorption, color, rigidity, object shape, body gesture
Physion-Eval	8 glitch categories	Object permanence violation, temporal coherence breakdown, material/state inconsistency, contact/interaction failure, causal sequence violation, force and motion inconsistency, geometric/collision violation
PhysiXFails benchmark	17 categories across 10 broader physical principles	Weightlessness, anti-gravity, clipping-through, spontaneous rapid spinning, sudden appearance/disappearance, biomechanical failures, buoyancy violations, optical physics violations

The PhyDetEx taxonomy is physics-law-centric. It includes unsupported bodies hovering, constant-height motion with $a_y \approx 0$ despite no upward force, non-conservative collisions, motion through fluids with no drag, and temporal jumps that violate $\Delta p = \int F\,dt$ . PhysGame, by contrast, emphasizes intuitive physical commonsense in gameplay videos, organizing glitches into mechanics, kinematics, optics, and material properties, with typical examples such as instant speed jumps, mirror lag, impossible refraction, and knees bending backwards. Physion-Eval further compresses the space into a compact annotation taxonomy suited to expert temporal localization, while the PhysiXFails runtime benchmark expands the category set to include spontaneous motion, delayed gravity effects, fluid dynamics failures, thermodynamic anomalies, and structural/material mechanics failures (Wang et al., 1 Dec 2025, Cao et al., 2024, Zhang et al., 20 Mar 2026, Li et al., 29 Jul 2025).

This suggests that “PhysiXFails” is not a single benchmark-specific error label but a broader family of physically implausible events whose exact operationalization depends on the evaluation task: QA over gameplay clips, binary plausibility detection, temporal expert annotation, or black-box runtime monitoring.

3. Benchmarks, datasets, and evaluation corpora

The current literature operationalizes PhysiXFails through several distinct resources. Some isolate implausibility by paired construction, some rely on expert human reasoning, and some test whether multimodal systems can solve visually grounded physics problems rather than merely identify glitches (Wang et al., 1 Dec 2025, Shen et al., 21 May 2025, Cao et al., 2024, Zhang et al., 20 Mar 2026, Yuan et al., 13 Oct 2025, Li et al., 29 Jul 2025).

Resource	Task	Scale
PID	Plausibility detection and explanation in T2V	Train split 2,588 paired videos; test split 500 videos
PhyX	Visually grounded physics reasoning	Approximately 3,000 multimodal questions
PhysGame	Physical commonsense violations in gameplay videos	880 videos
Physion-Eval	Expert diagnosis of physical realism in generated video	12,718 generated videos paired with 2,486 real-world source videos; 10,990 expert reasoning traces
LikePhys benchmark	Pairwise valid-invalid intuitive-physics evaluation	12 scenarios spanning over four physics domains; 10 variations per scenario
PhysiXFails benchmark	Runtime failure detection in PE-based software	1,000 curated runtime video clips

The PID dataset uses contrastive supervision: each real-world plausible video is paired with an implausible counterpart generated from a caption rewritten only in its physical component, preserving shared scene context. PhysGame instead builds one four-way multiple-choice QA item per gameplay video and explicitly filters out question-only solvable items. Physion-Eval pairs each generated clip with a corresponding real-world source video and records temporally localized expert glitch annotations with severity and natural-language explanations. LikePhys controls appearance by rendering valid-invalid video pairs in Blender with identical camera, lighting, textures, and geometry, varying only the targeted physics violation. PhyX moves to a different evaluation axis by testing whether models can solve de-redundantized university-level physics questions grounded in diagrams and naturalistic visuals (Wang et al., 1 Dec 2025, Cao et al., 2024, Zhang et al., 20 Mar 2026, Yuan et al., 13 Oct 2025, Shen et al., 21 May 2025).

These resources also differ in what they measure. PID reports Acc. Impl., Acc. Plaus., F1, and a 0–5 Reasoning Score. PhysGame uses multiple-choice accuracy. Physion-Eval reports failure rate, glitch density, severity, and Youden’s $J$ . LikePhys introduces Plausibility Preference Error (PPE), where lower is better. PhyX reports accuracy under open-ended and multiple-choice protocols across Full-Text, Text-DeRedundancy, and Text-Minimal variants (Wang et al., 1 Dec 2025, Cao et al., 2024, Zhang et al., 20 Mar 2026, Yuan et al., 13 Oct 2025, Shen et al., 21 May 2025).

4. Detection, explanation, and quantitative diagnosis

Two methodological trends dominate current detection work: fine-tuned multimodal detectors that output both a plausibility judgment and a textual rationale, and training-free or weakly adapted critics that score plausibility from internal model signals. PhyDetEx is representative of the first trend. It fine-tunes Qwen2.5-VL 7B with LoRA of rank 8 for 3 epochs at learning rate $v(t)$ 0 with cosine scheduler, using a unified language modeling objective in which the first token is the plausibility label and the remainder is the explanation. On Impossible Videos, PhyDetEx reports Acc. Impl. $v(t)$ 1, Acc. Plaus. $v(t)$ 2, F1 $v(t)$ 3, and Reasoning $v(t)$ 4; on the PID test split it reports Acc. Impl. $v(t)$ 5, Acc. Plaus. $v(t)$ 6, F1 $v(t)$ 7, and Reasoning $v(t)$ 8 (Wang et al., 1 Dec 2025).

LikePhys represents the second trend. It evaluates intuitive physics in video diffusion models without additional training by using the denoising objective as an ELBO-based likelihood surrogate. For a video $v(t)$ 9, it aggregates per-timestep noise-prediction errors into a surrogate likelihood and compares valid-invalid pairs through Plausibility Preference Error,

$a(t)$ 0

LikePhys reports overall Kendall’s $a(t)$ 1 with human preference, outperforming VideoPhy at $a(t)$ 2, VideoPhy2 at $a(t)$ 3, and Qwen2.5-VL at $a(t)$ 4. In its model benchmark, Hunyuan T2V attains average PPE $a(t)$ 5, Wan2.1–T2V–14B and CogVideoX1.5–5B both attain $a(t)$ 6, and AnimateDiff attains $a(t)$ 7, with lower values indicating better intuitive-physics preference (Yuan et al., 13 Oct 2025).

Human reasoning remains substantially stronger than automated critics when the task requires precise temporal localization and causal diagnosis. Physion-Eval reports that 83.3% of exocentric and 93.5% of egocentric generated videos contain at least one human-identifiable physical glitch, while state-of-the-art multimodal LLM critics fall well below ordinary viewers: the best $a(t)$ 8 is approximately $a(t)$ 9 in exocentric and $F = m a$ 0 in egocentric settings, and Gemini 3.0 Pro fails to identify glitches in 74.4% of exocentric and 90.1% of egocentric videos that humans find obviously unrealistic (Zhang et al., 20 Mar 2026).

In runtime PE-based software analysis, prompt-engineered LMMs currently outperform generic video anomaly detectors for fine-grained failure identification. On the PhysiXFails benchmark, DEVIL-Gemini-VD reports Accuracy $F = m a$ 1, Precision $F = m a$ 2, Recall $F = m a$ 3, F1 $F = m a$ 4, and AUC $F = m a$ 5 for violation detection, while Gemini-Custom-Freeform-VI reports Accuracy $F = m a$ 6, Precision $F = m a$ 7, Recall $F = m a$ 8, F1 $F = m a$ 9, and AUC $O = (o_1, \ldots, o_T)$ 0 for violation identification. The same study also finds that multi-violation clips amplify detectability for physics-aware methods, whereas generic video quality evaluators remain ineffective (Li et al., 29 Jul 2025).

5. Reducing PhysiXFails during generation

The most explicit mitigation strategy in the supplied literature is FinePhys, which targets fine-grained human action generation by treating the skeleton as a first-class guidance modality and explicitly encoding Lagrangian rigid-body dynamics inside the generation loop. FinePhys first detects online 2D keypoints, producing $O = (o_1, \ldots, o_T)$ 1 for $O = (o_1, \ldots, o_T)$ 2 joints, then lifts them to a data-driven 3D sequence through in-context learning,

$O = (o_1, \ldots, o_T)$ 3

Because purely data-driven lifting can violate physical laws under large deformations and rapid dynamics, PhysNet re-estimates motion with Euler–Lagrange dynamics,

$O = (o_1, \ldots, o_T)$ 4

and refines states with bidirectional temporal updates before decoding physically predicted poses $O = (o_1, \ldots, o_T)$ 5. The two 3D estimates are then fused by simple averaging,

$O = (o_1, \ldots, o_T)$ 6

re-projected to 2D, converted into Gaussian multi-scale heatmaps, and injected through LoRA adapters into a Stable Diffusion v1.5 plus AnimateDiff backbone (Shao et al., 19 May 2025).

This pipeline is designed to reduce physics-related failures that current video generators show on fine-grained gymnastics actions, including improbable mid-air rotations, implausible landings, limb distortions, and joint-limit violations. FinePhys is evaluated on three FineGym subsets—FX-JUMP, FX-TURN, and FX-SALTO—covering approximately 350 videos across 35 classes. Against the strongest reported baseline, AnimateDiff T+I, FinePhys improves user study scores from 3.20 to 4.13 on text alignment, from 3.20 to 3.86 on domain consistency, and from 3.17 to 4.03 on smooth stability; CLIP-SIM* domain from 0.769 to 0.826; CLIP-SIM* smooth from 0.793 to 0.833; PickScore to 19.941; and FVD from 529.38 to 484.49. The paper attributes these gains to physically refined skeletal trajectories, second-order bidirectional updates, and fused 3D guidance that constrain both spatial layouts and temporal kinematics (Shao et al., 19 May 2025).

The failure analysis is equally important. FinePhys reports that baseline methods exhibit levitation and static characters, severe limb distortions and body proportion errors, implausible rotations, and joint-limit violations, especially in classes such as “split leap with 1 turn” and multi-twist SALTOs. FinePhys reduces these effects, but extremely complex aerial SALTO combinations remain intractable, complete failure of the online pose estimator can force reliance on priors and produce overly static outputs, and contact dynamics such as feet-floor timing are not explicitly modeled (Shao et al., 19 May 2025).

6. Capability limits, adjacent failure analyses, and open directions

PhysiXFails is also a diagnosis of current model capability limits rather than only a catalog of visible glitches. PhyX shows that even strong multimodal models struggle with physics-grounded reasoning when text is de-redundantized and equations must be grounded in the image: on open-ended Text-DeRedundancy, GPT-4o scores 32.5%, Claude 3.7 Sonnet 42.2%, and GPT-o4-mini 45.8%, whereas human physics undergraduates and graduates score 75.6%, 77.8%, and 78.9%. In a sampled GPT-4o error set, Visual Reasoning Errors account for 39.6%, Lack of Knowledge for 38.5%, Text Reasoning Errors for 13.5%, and Calculation Error for 8.3%. Multiple-choice narrows the gap sharply, indicating that option cues permit shallower strategies than genuine visual grounding (Shen et al., 21 May 2025).

The same pattern appears in software and simulation settings. The runtime PhysiXFails study reports that 87.5% of surveyed practitioners regard physics failures as subtle and hard to notice, 75% cite unpredictable runtime behavior as a major challenge, and 81.3% want real-time detection during simulations. In foundation-model physics simulation, the failure language reappears around discretization artifacts, long-horizon error accumulation, negative transfer, and pretraining interference: the PhysiX simulation model explicitly notes that tokenization introduces quantization errors and that generalizing to novel physical processes requires fine-tuning (Li et al., 29 Jul 2025, Nguyen et al., 21 Jun 2025).

The proposed research directions are correspondingly structural. PhyDetEx calls for explicit physics priors or differentiable physics engines, stronger motion and pose tracking, temporal consistency modules, and expanded datasets with richer fluid and cloth cases. Physion-Eval emphasizes temporally grounded physics critics, better egocentric modeling, and physically constrained generation with stronger contact and conservation handling. FinePhys identifies contact modeling, explicit torque or gravity estimation, harder joint-limit enforcement, and energetic or center-of-mass regularization as natural extensions. Taken together, these directions treat PhysiXFails not as isolated anomalies but as evidence that current systems still lack robust mechanisms for causality, contact, conservation, support, and temporally coherent state evolution (Wang et al., 1 Dec 2025, Zhang et al., 20 Mar 2026, Shao et al., 19 May 2025).