PhysEvolve: Verifier-grounded Self-evolution for Physical Reasoning with Executable Worlds

Jinyuan Hu; Yilei Jiang; Chen Tang; Minghong Cai; Xinran Liu; Xiangyu Yue

cs.AIOct 26, 2026

PhysEvolve: Verifier-grounded Self-evolution for Physical Reasoning with Executable Worlds

Authors: Jinyuan Hu, Yilei Jiang, Chen Tang, Minghong Cai, Xinran Liu, Xiangyu Yue

Organizations: MMLab, The Chinese University of Hong Kong

Abstract

Physical reasoning underpins how embodied systems understand and act in the physical world, yet vision-language models (VLMs) remain unreliable on questions about dynamics and physical causality. Progress is critically constrained by the scarcity of grounded counterfactual training data, as observational videos capture only the factual trajectory that occurred. The self-evolving paradigm offers a route beyond fixed datasets by enabling models to construct their own training problems. Extending this paradigm to physics, however, requires an external source of truth that can both validate model-proposed environments and verify alternative physical outcomes under intervention. We introduce \physevolve{}, a verifier-grounded self-evolving framework that trains VLMs for causal physical reasoning from model-proposed executable worlds and simulation-authored tasks. At each round, a shared VLM proposes executable world programs and evidence schemas; a deterministic physics engine freezes each valid world, simulates matched baseline and body-elided counterfactual branches, and synthesizes discrete, answer-bound visual tasks. Exact verifier feedback from task solving updates both the shared policy via GRPO and an adaptive curriculum over physical intents. Starting from Qwen3-VL-8B-Instruct, \physevolve{} achieves targeted gains across PhysBench, CausalPhys, and ContPhy without human intervention data, improving PhysBench Throwing by $8.8$ points and ContPhy Counterfactual Removal and Goal-Pit by $7.4$ and $24.4$ points. These gains concentrate on interaction-sensitive and altered-world questions, demonstrating zero-shot transfer of causal principles to unseen complex dynamics (cloth, fluid, soft bodies) and real-world video.

Figures & tables

Figure 1: Naive self-play baseline for physical reasoning. A Proposer generates a world program; the executor simulates a single trajectory and derives Forward, Inverse, and Parameter tasks. Graded continuous proximity can reward a Solver that remains fundamentally incorrect, broadcasting distorted learnability signals back to the Proposer.

Figure 2: Overview of PhysEvolve. A shared VLM proposes an executable world program and evidence schema. A fixed deterministic executor validates and freezes the scene, simulates matched baseline and body-elided branches when required, and synthesizes an answer-bound item. Exact discrete Solver feedback updates the shared policy and the curriculum.

Panel of Table 1 — Table 1: Accuracy (%) across the three physical reasoning benchmarks. The upper block provides contextual reference models; bold text indicates the best performance within the strictly matched 8B comparison block.

Figure 4: PhysBench training dynamics and paired correctness analysis. Left: accuracy gains over Base across scheduled training checkpoints. Right: item-level paired analysis at the terminal checkpoint, separating questions answered correctly only by PhysEvolve from those answered correctly only by Base. Diamonds indicate net correct item gains; see Table 20 for complete counts.

Configuration	Branch Sim.	CF/Inv	Groups	Rollouts	Ovr.	Interv.	Ovr.	CF-Rm.	Goal-Pit
	Training Path		Realized Exposure		CausalPhys		ContPhy
PhysEvolve	On	On	2,167	8,668	54.50	45.03	39.54	51.64	44.44
w/o branch-enabled path	Off	Off	1,990	7,960	51.42	41.71	37.28	43.44	21.11
Difference	–	–	+177	+708	+3.07	+3.32	+2.26	+8.20	+23.33

Table 2: Ablation of the complete branch-enabled training path. Evaluating endpoint accuracy (%) under identical protocols when branch execution and dependent Counterfactual/Inverse families are deactivated.

	Training updates		CausalPhys		ContPhy
Configuration	Proposer	Solver	ACC	Parse rate	ACC	Parse rate
PhysEvolve	$\checkmark$	$\checkmark$	54.50	97.84	39.54	100.00
w/o Proposer-role updates	–	$\checkmark$	15.01	26.07	17.85	99.90
$\Delta$	–	–	$+39.48$	$+71.77$	$+21.69$	$+0.10$

Table 6

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Capability	Calls	Effect
Geometry	add_box , add_sphere , add_cylinder , add_capsule , add_plane	Register fixed or free rigid bodies and visual/material attributes.
Initial state	set_velocity , set_rotation	Set linear/angular velocity or a normalized orientation quaternion.
Constraints	add_hinge , add_slide , add_spring	Add one-DOF joints or a damped spatial spring between existing bodies.
Environment	set_gravity , set_duration , add_camera	Set the execution horizon, gravity vector, and look-at camera.
Finalization	freeze	Add default ground/camera if absent and snapshot a SceneDef .

Appendix

Table 5: Complete PhysDSL callable surface. Arguments are validated when the call is made; freeze() returns the immutable executor input.

Field	Enforced contract
Names	Pattern [a-z_][a-z0-9_]* ; one namespace across bodies, joints, springs, and cameras; ground/world/sky reserved ( ground is allowed for the plane).
Body centers and count	$\|x\|,\|y\|<1.5$ and $0<z<2.5$ m; a plane may use $z=0$ . At least one body after finalization and at most 20 dynamic bodies; fixed scene furniture does not consume the dynamic-body cap.
Geometric size	Each non-plane radius/half-size/half-height in $[0.01,1.0]$ m; each plane half-size is at least $0.01$ m and has no API upper cap.
Material	Mass $[0.01,100]$ kg, sliding friction $[0,2]$ , restitution $[0,1]$ , and every RGBA channel $[0,1]$ .
Initial state	Each linear and angular velocity component has magnitude at most $20$ ; orientation is a nonzero quaternion and is normalized.
Joint	Nonzero axis; optional limits satisfy $\ell<u$ ; parent and child differ; only the parent may be world ; one explicit joint per child.

Appendix

Table 6: PhysDSL construction-time bounds. Intervals are closed unless an inequality is written explicitly.

World class	Required role fields	Served active families
Throwing	balls , three or more fixed platforms , two or more removable walls , region_kind , and three destination regions derived from platform gaps	final destination; current-run judgment; counterfactual destination; inverse intervention
Collision	dynamic target ; trigger ; exactly four dynamic movers ; relation_pair={a,b,expected}	final state; final relation; motion attribution
Motion	dynamic motion_target ; optional speed_pair={a,b}	motion-state change, including forward, speed-profile, and pairwise surface forms

Appendix

Table 7: The three SCENE_MANIFEST contracts enabled in the reported run. Braces denote a body-role mapping, not model-predicted state.

Panel of Figure 5 — Figure 5: Throwing reference world and executable interventions. The frozen baseline and three altered inputs share the same camera and initial conditions for retained bodies. Each altered input removes the designated wall set and the non-query balls. The executor runs these inputs separately to obtain the region readouts used for question and answer construction.

Evidence key	Elided bodies	Readout
$\tau_{0}(b)$	none	baseline settled region
$\tau_{A}(b)$	wall A and all non-query balls	altered settled region
$\tau_{B}(b)$	wall B and all non-query balls	altered settled region
$\tau_{AB}(b)$	walls A+B and all non-query balls	altered settled region

Appendix

Table 8: Throwing evidence objects. Each altered input is executed with the same seed and horizon as the frozen baseline.

Panel of Figure 6 — Figure 6: Collision reference world. The initialization and later frames connect the entity roles to execution. The lower panels bracket contact between the orange ball and blue plate; times are measured from simulation start.

Panel of Figure 7 — Figure 7: Motion reference world. Initialization and two execution states show the ramp-to-platform motion and resulting contacts. All frames share the same camera; times are measured from simulation start.

Family	Representative item and option semantics	Evidence, binding, and suppression
throwing_final_destination	“What is the final destination of the [referent]?” Regions use {left pit, right pit, center pit, none}; an unsettled layout uses {left, right, neither, none}.	Baseline only. A region requires final speed $\leq 0.30$ m/s, a non-ejected status, exactly one derived gap interval, and final height no more than $0.30$ m above the platform tops. A settled region binds that region; otherwise neither is bound rather than guessed.
throwing_current_run_judgment	“Will the [referent] eventually drop into the left pit?” Options are {Yes, No, Can not answer, Uncertain}; a second surface asks whether it reaches a different destination region than a named one.	Baseline only. Yes iff the settled-region readout equals the named target; otherwise No , including an unsettled/ejected readout. Target choice alternates between matching and nonmatching regions across ball indices.
counterfactual_destination	“If we removed the [wall] and other balls, which pit would the [referent] drop into?” Options are {left, right, center, none}.	Matched baseline and one elision branch. Admit only when the branch is available, settles in one region, and differs from the baseline region. The branch region is the answer; failed, unsettled, or unchanged branches suppress the item.
inverse_intervention	“What can we do to make the [referent] drop into the [target] pit?” Options are {remove A + other balls, remove B + other balls, remove both + other balls, none}.	Baseline plus all three A/B/A+B branches. In a first fixed-order pass, the first non-baseline target reached by exactly one branch binds that intervention. Only if no target has a unique achiever, a second pass binds none to the first non-baseline target reached by no branch. A multiply achieved target is skipped; missing branches suppress.

Appendix

Table 9: Exact contracts for the four active throwing families. The option descriptions give canonical semantics; an eligible surface layout may permute slots and substitute compatible distractors, but must preserve a unique binding for the admitted answer semantic.

Family	Representative item and option semantics	Evidence, binding, and suppression
collision_final_state	“Which phrase best describes the final pose of the [target]?” Options are {standing upright, leaning, lying horizontally, floating}.	Terminal status and speed. Ejected with $z>0.35$ binds floating; fallen binds lying; moving with speed $\geq 0.10$ binds leaning; stable with speed $\leq 0.20$ binds standing. Other/borderline combinations suppress the item.
collision_final_relation	“Which phrase best describes the interaction between A and B during the run?” Options are {ends in contact, never touches, passes near without touching, ends vertically separated without contact}; a binary surface uses {Yes, No, Sometimes, Uncertain}.	Contact is trajectory-level: any recorded pairwise contact dominates and binds the contact semantic (canonically rendered as “ends in contact”). Otherwise vertical separation requires final $\|\Delta z\|\geq 0.30$ and lateral gap $<0.30$ m; pass-by requires initial or final distance $\leq 0.20$ m; the remainder is no-contact. The computed class must equal the manifest declaration or the family is suppressed.
motion_attribution	“Which object experienced a change in motion as a result of the [trigger]?” The four options are the manifest’s four mover bodies.	Candidates must be distinct, dynamic, and different from the trigger. Eligibility requires trigger contact and velocity-change magnitude $\lVert\mathbf{v}_{f}-\mathbf{v}_{0}\rVert_{2}>0.30$ m/s; the argmax is bound only if its lead over a second eligible candidate is at least $0.30$ m/s. The absence of an eligible candidate, or a lead below this margin, suppresses the item.
motion_state_change	Three surfaces: motion change {stop, slow but moving, speed up, same}; speed profile {accelerate, decelerate, accelerate then decelerate, constant}; pairwise semantics {final tie, final A faster, final B faster, non-tied order reversal}. The canonical middle options use “consistently faster” wording.	Require target initial speed $\geq 0.10$ m/s. Stop uses final speed $<0.10$ ; speed-change/profile boundaries use an absolute $0.30$ m/s gap; pairwise order uses $0.10$ m/s. Crossover is bound only when both endpoint orders are non-tied and reverse; otherwise the final order is bound. The pairwise surface requires both declared bodies.

Appendix

Table 10: Exact contracts for the other four active families. Canonical options are shown; eligible layouts may change slot order and compatible distractors while preserving the admitted answer semantic.

Execution	Baseline	Remove A	Remove B	Remove A+B
Stored semantic	$r_{0}$	$r_{A}$	$r_{B}$	$r_{AB}$
Required for counterfactual(A)	yes	yes	no	no
Required for inverse	yes	yes	yes	yes

Appendix

Table 11: Branch-result contract used for counterfactual and inverse items.

Component	Reported configuration
Failure scale $\lambda_{\mathrm{fail}}$	$1.5$
Instance-level tent coefficient $\lambda_{\mathrm{learn}}^{\mathrm{eff}}$	held at $0$ (validity/diversity dominate instance scoring)
Intrinsic weight $\lambda_{\mathrm{int}}$	$0.1$ ; structural diversity only
Repair factors	$0.90$ minor; $0.80$ schema; $0.75$ geometry
Near-copy control	similarity $\geq 0.95$ capped at $0.1$ before repair and family scaling
Exact/within-group duplicate proposal	pre-family reward override $-0.8$

Appendix

Table 12: Reward and sampling controls in the reported 100-step run.

Component	Configuration
Base and horizon	Qwen3-VL-8B-Instruct; 100 steps; checkpoints at 25/50/75/100
Run multiplicity	One training run per reported configuration; no training-seed sweep
Compute	4 GPUs; 32 Ray CPUs; 4 executor workers
Proposer sampling	$N=4$ intent groups per step; $G=4$ programs per intent; at least one throwing, collision, and motion intent per step
Solver retention and rollout	At most 5 items per scene and 32 groups per step; 4 answers per group; rollout chunk size 8
Visual variants	One additional color permutation per accepted baseline world; one executor camera; 32 rendered frames at $480\!\times\!480$

Appendix

Table 13: Resolved configuration for the reported PhysEvolve run.

Benchmark	Items	Visual input	Generation	Scoring
PhysBench	9,802	8 uniform frames	greedy; 2,048 tokens	official prompts, normalization, and scorer
CausalPhys	3,057	5 uniform frames	CoT; greedy; 512 tokens	final A–D answer; unparsed output is incorrect
ContPhy	1,950	10 uniform frames	greedy; 64 tokens	normalized MC or normalized open-response exact match

Appendix

Table 14: Evaluation protocols for the three reported benchmarks.

View ( $n$ )	Base	Naive	$\Delta$ vs. Base	PhysEvolve	$\Delta$ vs. Base
Overall (9,802)	52.92	52.39	$-0.53$	53.47	$+0.55$
Dynamics (2,932)	44.37	44.44	$+0.07$	45.29	$+0.92$
Throwing (410)	46.10	45.85	$-0.24$	54.88	$+8.78$
Collision (650)	38.62	38.77	$+0.15$	40.77	$+2.15$
Motion (1,433)	54.71	55.48	$+0.77$	56.45	$+1.74$
Prediction (1,430)	42.17	42.87	$+0.70$	43.71	$+1.54$

Appendix

Table 15: PhysBench accuracy and paired differences for matched models.

Taxonomy level	View ( $n$ )	Base	PhysEvolve	$\Delta$
Task	Dynamics (2,932)	44.37	45.29	+0.92
	Relationships (2,088)	61.97	63.46	+1.48
	Scene (2,240)	45.13	45.62	+0.49
	Property (2,542)	62.20	61.61	$-0.59$
Dynamics subtype	Throwing (410)	46.10	54.88	+8.78
	Collision (650)	38.62	40.77	+2.15

Appendix

Table 16: Complete top-level PhysBench task and ability partitions for the matched Base and PhysEvolve models. The dynamics-subtype rows are nested inside Dynamics and therefore do not form a complete partition.

	PhysBench								CausalPhys			ContPhy
Model	Ovr.	Dyn.	Throw.	Coll.	Motion	Pred.	Judg.	Reason.	Ovr.	Antic.	Interv.	Ovr.	CF-Rm.	Goal-Pit
InternVL2.5-8B	43.8	42.1	40.0	38.9	27.7	43.0	32.7	27.4	50.2	51.5	42.4	31.4	48.4	23.3
Llama-3.2-11B-Vision	38.7	43.2	36.6	35.5	27.2	44.1	30.8	29.6	44.2	44.1	40.1	20.7	0.0	20.0
Pixtral-12B	46.9	44.6	41.2	39.2	35.2	46.9	31.0	35.1	50.2	53.0	44.5	21.5	40.2	13.3
Qwen2.5-VL-7B	49.4	44.3	44.4	36.5	50.6	44.1	33.1	36.5	52.6	56.5	45.7	25.9	11.5	24.4
Qwen3-VL-32B	60.9	48.7	47.1	46.0	83.5	48.0	44.3	51.4	58.9	61.9	52.9	41.0	45.9	26.7

Appendix

Table 17: Full accuracy (%) table for open-weight context and matched 8B models.

Benchmark view ( $n$ )	Base	PhysEvolve	$\Delta$
CausalPhys Overall (3,057)	53.94	54.50	$+0.56$
Anticipation (896)	55.92	58.37	$+2.46$
Goal orientation (637)	50.71	49.61	$-1.10$
Intervention (573)	42.76	45.03	$+2.27$
Perception (951)	60.99	59.83	$-1.16$
ContPhy Overall (1,950)	37.08	39.54	$+2.46$

Appendix

Table 18: Matched cross-benchmark accuracy and complete category/domain breakdowns.

ContPhy
Task	$n$	Base	Naive	w/o branch	PhysEvolve
Counterfactual Removal	122	44.26	43.44	43.44	51.64
Goal-Pit	90	20.00	25.56	21.11	44.44
CausalPhys / Anticipation
Collision Prediction	300	53.67	50.67	54.67	51.67
Fluid Flow	200	47.50	60.50	35.50	58.50

Appendix

Table 19: Complete accuracies for the task-level transfer comparison (%). CausalPhys controls use the matched video-packing protocol.

Task	$n$	Both correct	Ours only	Base only	Neither	$\Delta$ (pp)
Overall	9802	4487	754	700	3861	+0.55
Throwing	410	156	69	33	152	+8.78
Judgment	767	220	87	48	412	+5.08
Motion	1433	730	79	54	570	+1.74

Appendix

Table 20: Paired terminal-checkpoint correctness on PhysBench. Each row partitions the same evaluation items by Base and PhysEvolve correctness.

Panel of Figure 8 — Figure 8: Matched ContPhy examples for Counterfactual Removal and Goal-Pit. Each example displays three of the ten evaluation input frames, the original question and options, the extracted model answers, and the reference answer. The examples illustrate reasoning about removal and a requested destination in held-out soft-body scenes. The Goal-Pit reference answer is “None of the above.”

Reasoning for VLMs

See all 58 papers

Aug 5, 2026

PhysMind: From Video to Executable Worlds for Training-Free Physical Reasoning

Reliable physical reasoning from video requires understanding how objects move, interact, and respond to interventions. Existing vision-language models (VLMs) often struggle to interpret these dynamics and reason reliably about future and counterfactual outcomes. We introduce PhysMind, a training-free agentic framework that constructs one reusable, question-agnostic executable world per video. PhysMind recovers a temporally consistent dynamic scene through object segmentation, mesh reconstruction, and 6D pose tracking, then fits analytic continuous-time dynamics and latent physical parameters without unrolling a time-stepped simulator. Given a question, it inspects, continues, or edits the world and answers from the resulting trajectories and interactions. Relative to direct chain-of-thought (CoT) reasoning with the same VLM, PhysMind improves accuracy by 38.23 points on CLEVRER and 8.08 points on Physion++. On counterfactual questions, it exceeds the strongest evaluated VLM baseline, GPT-5.5, by 19.25 points.

Chen Yang, Shenxiang Zeng, Haoyang Zhao +6

1Tsinghua University · 2The University of Hong Kong

Counterfactual Reasoning Physical Reasoning in Video Embodied World Models Temporal Reasoning in VLMs

Ask PDF

Jun 4, 2026

Causal Scaffolding for Physical Reasoning: A Benchmark for Causally-Informed Physical World Understanding in VLMs

Understanding and reasoning about the physical world is the foundation of intelligent behavior, yet state-of-the-art vision-language models (VLMs) still fail at causal physical reasoning, often producing plausible but incorrect answers. To address this gap, we introduce CausalPhys, a benchmark of over 3,000 carefully curated video- and image-based questions spanning four domains: Perception, Anticipation, Intervention, and Goal Orientation. Each question is paired with an expert-annotated causal graph capturing object-attribute-event dependencies, enabling interpretable and fine-grained evaluation of causal understanding. Building on this, we formulate a causal-graph-grounded metric that quantitatively measures how well a model's chain-of-thought reasoning aligns with the correct causal relations, moving beyond answer-only accuracy and enabling systematic diagnosis of VLMs' causal reasoning failures. Using this metric, we conduct a comprehensive analysis of leading VLMs, revealing systematic gaps in capturing causal dependencies and underscoring the need for causality-aware learning. To address these limitations, we further propose Causal Rationale-informed Fine-Tuning (CRFT), which explicitly aligns VLM reasoning with causal structures. Extensive experiments demonstrate that CRFT substantially enhances both reasoning accuracy and interpretability across multiple model backbones. By unifying dataset curation, causal evaluation, and causality-informed learning, CausalPhys establishes a strong foundation for advancing modern VLMs toward causally grounded physical reasoning.

Tianyi Tang, Zhuoyi Lin, Zeyu Feng +4

CFAR, IHPC, Agency for Science, Technology and Research (A*STAR) Singapore · I2R, Agency for Science, Technology and Research (A*STAR) Singapore · Nanyang Technological University Singapore

VLM Evaluation VLM Reasoning Physical Reasoning in Video

Ask PDF

Jul 11, 2026

PhysMRV: Physical Memory Retrieval and Verification for Physics Plausibility Reasoning

Video-language models (VLMs) have achieved remarkable performance on video understanding and visual question answering, yet they remain unreliable in reasoning about physical plausibility, where understanding object interactions, causal dynamics, and fundamental physical principles is essential. This limitation is particularly evident on challenging physical reasoning benchmarks, revealing a persistent gap in physical commonsense reasoning. To address this challenge, we propose PhysMRV, a training-free physical memory and verification framework for physical plausibility reasoning. Unlike retrieval-augmented VLMs that retrieve semantically similar videos as additional context, PhysMRV transforms training videos into a Hierarchical Memory Bank of structured physical knowledge comprising three complementary levels: scene descriptions capturing visual context, physical-event graphs modeling object interactions and causal structure, and physics-rule summaries distilling reusable physical principles and cues. During inference, PhysMRV retrieves physically relevant memories and leverages their structured physical evidence to guide a frozen VLM in verifying physical plausibility, requiring neither fine-tuning nor parameter updates. We evaluate PhysMRV on three challenging physical reasoning benchmarks, ImplausiBench, IntPhys2, and GRASP Level 2, across multiple state-of-the-art VLMs. Experimental results demonstrate consistent improvements over direct prompting across diverse VLMs and evaluation benchmarks, showing that structured physical memories provide an effective and scalable means of enhancing physical plausibility reasoning without additional training.

Wenyuan Wang, Lianyu Hu, Hao Wang +1

1Nanyang Technological University · 2Rutgers University · University of Illinois Urbana-Champaign

Memory-Augmented VLMs Physical Reasoning VLM Reasoning Physical Reasoning in Video

Ask PDF

Abstract

Reasoning for VLMs

Explore similar work

PhysMind: From Video to Executable Worlds for Training-Free Physical Reasoning

Causal Scaffolding for Physical Reasoning: A Benchmark for Causally-Informed Physical World Understanding in VLMs

PhysMRV: Physical Memory Retrieval and Verification for Physics Plausibility Reasoning