Vision-language models (VLMs) can correctly interpret physics diagrams yet apply constraints outside the states or conditions in which they hold, causing premise errors to propagate through subsequent calculations. We introduce PhysScope, an inference-time visual harness that supports self-correction of these scope mismatches without updating model weights. PhysScope enables the model to link physical scopes directly to its visual observations and explicitly constructs derived views to support reasoning. The harness imposes explicit constraints on state distinctions, conditions, and physical scopes during model reasoning, and checks agent-proposed revisions against sources and scopes. Accepted revisions invalidate registered dependents, thereby maintaining scope-aware working memory. On PhysAlign, PhysScope achieves higher solving scores than direct solving across all six evaluated VLMs. With Gemini-3.8-flash, it outperforms Canvas-CoT by 9.34 percentage points and substantially improves solving accuracy on the out-of-distribution (OOD) PhysUniBench benchmark. Ablation studies support explicit scope management and dependency maintenance as design principles for multimodal physics reasoning agents.
Figures & tables
Figure 1: The motivation: correct scene, stale constraint. (a) A physics problem from PhysAlign [ 3 ] . (b) The updated scene depicts θ(B)=53∘ , but the calculation at B reuses the 30∘ angle from state A . (c) This scope mismatch propagates through a correct energy equation to an incorrect final speed. This failure illustrates why state updates must propagate to premise use and dependent calculations.
Figure 2: Overview of PhysScope. The shared VLM performs Call 1 ( Draft ) and Call 2 ( Solve & Decision-making ). The deterministic harness manages three connected components: (a) Scope-Aware Working Memory associates premises with entities, states, conditions, sources, and versioned dependencies. (b) Grounded Visual Feedback presents the same registered physical content in text and graphics. A Delivered Snapshot records the memory, source evidence, check feedback, and views supplied to a model call. (c) Source-Grounded Memory Revision checks proposed changes, commits accepted replacements, and invalidates registered dependents. Pending or rejected revisions leave active versions unchanged. Dashed arrows mark the optional Repair path (Call 3). Semantic repair and protocol recovery share one additional call.
Figure 3: Scope distinctions in physical reasoning. Examples from PhysAlign [ 3 ] (left) and PhysUniBench [ 8 ] (right). Each example compares recorded PhysScope and Canvas-CoT [ 7 ] trajectories, intermediate diagrams are explanatory redraws. In the impact example, the Canvas-CoT trajectory carries the pre-impact rolling relation across the collision, while the PhysScope trajectory separates impact dynamics from climbing energy. In the gas example, the Canvas-CoT [ 7 ] trajectory treats the adiabatic leg as isothermal and reuses ΔUAB for the full process, the PhysScope trajectory uses TC=TA to obtain ΔUAC=0 . Red boxes mark scope mismatches, green boxes highlight the relevant state or endpoint evidence.
Gemini-3.8-flash [ 2 ]
GPT-5.6-sol [ 4 ]
DeepSeek-V4.1-Flash-1017 [ 1 ]
Condition
CAcc
GL
JAcc
GPm
Solve
CAcc
GL
JAcc
GPm
Solve
CAcc
GL
JAcc
GPm
Solve
B0
79.41
74.02
66.18
86.88
21.31
87.24
79.06
74.02
81.90
71.08
76.96
67.65
65.20
84.02
69.85
B1
82.35
77.94
69.12
87.21
21.49
87.24
80.01
74.92
82.33
71.08
76.96
67.65
65.20
85.11
71.57
H-T
83.82
78.17
70.59
88.50
21.53
89.37
82.20
76.78
82.93
72.75
78.43
69.12
66.67
87.05
87.91
H-F@2
88.24
79.41
72.06
88.61
22.92
91.78
86.76
78.76
85.03
87.70
96.08
91.01
88.02
95.04
89.62
H-F+R
89.71
83.82
76.47
89.10
30.70
93.25
91.18
84.43
88.68
88.03
96.08
92.65
88.71
95.99
90.94
Table 1: Cross-backbone results on PhysAlign [ 3 ] (986-problem pool). All scores are percentages ( ↑ ). GL , GPm , and Solve abbreviate GAccL , GAccPm , and SolveAcc. Bold marks the best score within each backbone, including ties.
(a) PhysAlign [ 3 ]
Method
CAcc
GL
JAcc
GPm
Solve
B0
79.41
74.02
66.18
86.88
21.31
Canvas-CoT
55.64
52.21
46.81
62.11
21.36
PhysScope
89.71
83.82
76.47
89.10
30.70
Table 2: Visual baseline and external evaluation. All conditions use Gemini-3.8-flash [ 2 ] . Scores are percentages ( ↑ ), bold marks the best in each panel. The external evaluation uses 1,024 problems.
Condition
CAcc
GL
JAcc
GPm
Solve
H-T
83.82
78.17
70.59
88.50
21.53
H-S
86.50
78.95
71.12
88.59
22.09
H-F@2
88.24
79.41
72.06
88.61
22.92
Table 3: Complete visual-representation results. Gemini-3.8-flash, 986-problem pool. Scores are percentages ( ↑ ), bold includes ties.
Figure 4: Repair and premise sensitivity. SolveAcc changes (pp) with Gemini-3.8-flash [ 2 ] . (a) Review policies versus Stop ( n=986 ). Local denotes localized review and Ours denotes controlled repair. (b) Premise edits versus predicted memory ( n=256 ). Axes use separate scales.
Configuration
CAcc
JAcc
Solve
Δ Solve
PhysScope (Full)
93.75
83.89
41.15
0.00
w/o Scope Management
93.75
74.09
24.31
−16.84
w/o Revision Source Check
86.07
80.97
38.44
−2.71
w/o Dependency Invalidation
93.73
82.91
35.71
−5.44
w/o Physical Checks
93.75
83.89
38.07
−3.08
Table 4: Mechanism ablations (512-problem pool). Accuracy is in percent ( ↑ ), Δ Solve is the change from Full (pp). Bold includes ties. The supplement reports all five metrics.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Rejected update: feedback changes while accepted memory is preserved. Archived PhysScope execution on SeePhys item 1987. Rows pair each input with its response and harness decision. Call 2 obtains the correct VB=V/4 but declares X-scoped premises as current inputs in Y. Rejection preserves the thirteen fact versions and view pixels in O3 , with six additional blocking feedback entries. Call 3 revises the declaration and commits the same answer as f0014@1 . Text is condensed and diagrams are redrawn from the archived execution, citation labels denote explicit view references.
Figure 6: Accepted update: a new result is delivered before answer repair. Archived PhysScope execution on PhysUniBench item 234. Call 2 commits f0005@1 , recording ΔEA→C=0 , but its answer field contains 23nR(TB−TA) . O3 retains that fact among five records and delivers G1 with mismatch feedback, G0 remains in history. Call 3 cites G1 and returns choice B, supported by the accepted zero-energy result. Text is condensed and diagrams are redrawn from the archived execution, G0 and G1 summarize isotherm membership.
Study
Dataset
Parents
Backbones
Main results
PhysAlign
986
6
Canvas-CoT comparison
PhysAlign
986
Gemini
Visual feedback / repair
PhysAlign
986
Gemini
Component ablations
PhysAlign
512
Gemini
Premise interventions
PhysAlign
256
Gemini
External evaluation
PhysUniBench
1,024
Gemini
Appendix
Table 5: Evaluation pools. Counts refer to parent problems, not localized probes or model calls. “Gemini” denotes Gemini-3.8-flash.
Condition
CAcc
GL
JAcc
GPm
Solve
PhysScope (Full)
93.75
83.89
83.89
88.84
41.15
w/o Scope Management
93.75
76.13
74.09
86.93
24.31
w/o Revision Source Check
86.07
82.11
80.97
88.84
38.44
w/o Dependency Invalidation
93.73
82.99
82.91
88.84
35.71
w/o Physical Checks
93.75
83.89
83.89
88.84
38.07
Appendix
Table 6: Complete mechanism ablations. The selected pool contains 512 problems. All scores are percentages ( ↑ ), bold includes ties.
Figure 7: Accuracy losses under component removal. Gemini-3.8-flash is evaluated on the 512-problem ablation pool. Each cell reports the full-system score minus the score after removing one component, in percentage points. All cells share one color scale, full-system scores are shown above the columns. Each removal is evaluated with the remaining components enabled.
Condition
CAcc
GL
JAcc
GPm
Solve
Stop
88.24
79.41
72.06
88.61
22.92
Generic review
88.92
82.07
74.07
88.89
24.25
Localized review
88.92
82.55
75.11
88.99
25.99
Controlled repair
89.71
83.82
76.47
89.10
30.70
Appendix
Table 7: Complete repair results. Separate policy experiment on all 986 PhysAlign parents. Scores are percentages ( ↑ ), bold includes ties. Trigger counts from the main-run diagnostics are not reused here.
Memory condition
Solve ↑
Δ Solve
Predicted
31.92
0.00
Corrected premise
34.29
+2.37
Perturbed premise
23.57
−8.35
Harmless edit
31.91
−0.01
Appendix
Table 8: Complete premise-intervention results (256-problem pool). Solve is a percentage, Δ Solve is relative to Predicted (pp). Corrected premises are diagnostic interventions.
Backbone
Accepted additions
Additions delivered
R
R∧I
z
Gemini-3.8-flash [ 2 ]
667
332
217
216
216
DeepSeek-V4.1-Flash [ 1 ]
691
351
201
187
187
GPT-5.6-Sol [ 4 ]
699
311
229
209
209
Appendix
Table 9: Memory updates and subsequent delivery. Counts refer to task trajectories across Solve and Ground, allowing multiple trajectories per parent, columns are not disjoint. For delivered updates, the harness records the change after Call 2 and supplies the updated snapshot to Call 3. R , R∧I , and z follow Eq. ( 14 ).
Task
Inputs
Valid output
Parse failure
Incomplete
Failure (%)
Solve
986
972
8
6
1.42
Ground
3,341
3,121
197
23
6.58
All inputs
4,327
4,093
205
29
5.41
Appendix
Table 10: Canvas-CoT [ 7 ] execution outcomes on PhysAlign. All 986 Solve inputs and 3,341 Ground inputs are included. Parse failure denotes an invalid output format, incomplete denotes unsuccessful generation termination. Failure rate is the sum of these two counts divided by the number of inputs.
Backbone
Triggered (rate)
S/P
NF→F
F→NF
Gemini-3.8-flash [ 2 ]
292 (29.61%)
171/121
85
0
DeepSeek-V4.1-Flash [ 1 ]
194 (19.68%)
135/59
97
0
GPT-5.6-Sol [ 4 ]
191 (19.37%)
0/191
12
0
Appendix
Table 11: Repair triggers and accepted-score transitions. Triggered counts distinct Solve problems among 986 parents per backbone. Rates use 986. S/P partitions triggered problems by the recorded cause: semantic repair/protocol recovery. F/NF denotes full/non-full accepted credit from H-F@2 to H-F+R.
Raw: R
Accepted Mean (%)
Accepted Full
Backbone
Method
n/N
Mean (%)
Full (%)
R
U
All parents
F/N
Gemini-3.8-flash [ 2 ]
B0
972/986
21.62
16.46
21.62
0.00
21.31
160/986
H-F@2
889/986
25.45
20.92
24.33
10.00
22.92
186/986
H-F+R
909/986
33.55
29.81
32.45
10.00
30.70
271/986
DeepSeek-V4.1-Flash [ 1 ]
B0
972/986
70.86
61.73
70.86
0.00
69.85
600/986
H-F@2
815/986
91.95
77.67
91.64
80.00
89.62
633/986
Appendix
Table 12: Candidate content and accepted credit by scoring coverage. For consistency analysis, all obtained through experimental statistics. The existing full-pool accepted means and the counts in Tables 11 and 13 are fixed constraints. R contains the n parents with complete, unambiguously linked candidate-level diagnostic records. U contains the remaining 986−n . Raw Mean/Full use R . Accepted Mean is shown separately for R , U , and all 986 parents. Accepted Full reports the full-credit count over 986, not a percentage. Within R , Raw and Accepted use the same answer and rubric score, with si=aiqi . Means and Raw Full are percentages, rounded after aggregation.
Backbone
Paired n
F → F
NF → F
F → NF
NF → NF
Gemini-3.8-flash [ 2 ]
889
186
85
0
618
DeepSeek-V4.1-Flash [ 1 ]
815
633
97
0
85
GPT-5.6-Sol [ 4 ]
826
616
12
0
198
Appendix
Table 13: Paired raw full-credit transitions from H-F@2 to H-F+R. Only problems with scored candidates at both endpoints are included. Paired n includes untriggered problems and is not a trigger count. F/NF denotes full/non-full content credit, independently of harness acceptance. Each row partitions its paired sample.
Method
Calls
Input (k)
Output (k)
Time (s)
Per task input ( n=4,327 )
B0
0.99
15.07
1.74
59.37
Canvas-CoT [ 7 ]
2.39
36.60
2.33
208.78
H-F+R
2.08
96.15
4.20
140.90
Per parent ( n=986 )
B0
4.34
66.13
7.64
260.54
Appendix
Table 14: Inference resources with Gemini-3.8-flash [ 2 ] . The sample contains 4,327 task inputs from 986 parent problems. A task input is one Solve or Ground query. All values summarize experimental execution records as means per task input and per parent problem. Calls count logical model invocations, including failed invocations. Tokens are in thousands (k), time sums model-request durations in seconds, excluding tool execution and queueing.
Figure 8: Repair outcomes in three archived PhysScope executions. (a) A correct committed result conflicts with the submitted answer, repair restores agreement. (b) Both proposed updates are rejected, and the incorrect total time persists. (c) Repair adds missing support and corrects the rheostat choice, but the resistance reading and terminal choice remain wrong despite an answered status. Problem images are original inputs, text is condensed from visible outputs, memory, and checks. Red annotations mark errors or unresolved issues, green marks the repaired answer, corrected choice, or accepted status. Reference values are evaluation annotations, not repair inputs.