PhysScope

PhysScope: Agentic Multimodal Physics Reasoning with Grounded Visual Memory

1 Sun Yat-sen University2 Dalian University of Technology† Corresponding author
Scope inspectorMechanics
Two masses connected by a rope over a fixed pulley In state B, block P is closer to the pulley and its rope makes a 53-degree angle with the horizontal. The dashed rope shows the other state. Q Q 53° P 30° P O BA
Velocity constraintvQ = vP cos 53°
Switch states to see which rope angle applies.
6/6

Backbones with higher solving scores on PhysAlign

+9.34 pp

Over Canvas-CoT with Gemini on PhysAlign

+150

Additional correct answers on PhysUniBench with Gemini

When does a premise apply?

A model can read both angles in a physics diagram correctly and still substitute the wrong one into an equation.

In the pulley problem above, block P starts at A. When it reaches B, the rope makes a steeper angle with the rod. Reusing the initial 30° angle gives the wrong speed for Q, even if every subsequent algebraic step is correct.

PhysScope records which objects, states, and conditions a premise refers to. The 30° reading remains available for questions about A. A calculation at B uses the 53° reading associated with B.

View the source diagram and its state annotation
The original problem drawing of a pulley, spring, inclined plane, and blocks P and Q.
Source diagram
The paper’s annotation at B, showing the velocity directions, a 53-degree active angle and the earlier 30-degree angle.
Annotated state B

How PhysScope works

PhysScope gives a vision-language model a working memory it can inspect and revise at inference time. The model weights stay fixed.

The PhysScope framework connects a two-call reasoning process to scoped premises, derived views, and a source-checked revision loop.
The harness connects the original diagram, versioned premises, derived views, and the model calls that receive them. Paper, Fig. 2
Record the scope
On the first call, the model reads the problem and proposes physical facts. The harness records each premise with its source, state, and conditions of use.
Show the relevant state
For the second call, the harness delivers the evidence, checks, and supported diagrams drawn from those same records. Unresolved quantities remain visible as unknowns.
Recheck what depends on a revision
Correcting a source reading requires a fresh inspection. If the correction is accepted, the harness marks dependent equations and views as stale. An eligible issue can trigger one follow-up call.

Two ways to lose the scope

Recorded PhysScope and Canvas-CoT trajectories for a sphere hitting a step and a gas following a two-leg thermodynamic process.
Recorded reasoning trajectories from PhysAlign and PhysUniBench, with explanatory redraws of intermediate diagrams. Paper, Fig. 3

A change of pivot

A rolling sphere hits a step. The pre-impact rolling relation applies on the floor. To work out whether the sphere climbs, the solver needs its angular velocity just after impact about the corner.

A change of endpoints

A gas expands from A to B and follows an adiabatic path from B to C. A and C have equal temperatures. Their internal-energy difference is zero even though the energy changes along the first leg.

Experimental results

All six backbones earn more solving credit with PhysScope on the 986-problem PhysAlign evaluation. The size of the gain varies by model.

PhysAlign

SolveAcc (%)
Direct solving and PhysScope on PhysAlign, with percentage-point gains. Paper Table 1.
BackboneDirectPhysScopeGain (pp)
Gemini-3.8-flash21.3130.70+9.39
GPT-5.6-sol71.0888.03+16.95
DeepSeek-V4.1-Flash-101769.8590.94+21.09
Qwen3.5-9B12.1223.39+11.27
Qwen3.5-27B18.3924.54+6.15
InternVL3.5-14B12.7114.54+1.83

SolveAcc averages normalized problem credit, including partial credit on open-ended problems. With Gemini-3.8-flash, Canvas-CoT scores 21.36%, compared with 30.70% for PhysScope. Tables 1–2

Transfer to PhysUniBench

On 1,024 external problems, Gemini-3.8-flash answers 384 correctly with direct solving and 534 with PhysScope. This evaluation scores the final answer as correct or incorrect.

Direct solving37.50%
PhysScope52.15%

What changes when scope is removed?

On a fixed subset of 512 PhysAlign problems, removing scope management leaves content recognition at 93.75%, while SolveAcc falls from 41.15% to 24.31%. Separately, removing dependency invalidation lowers SolveAcc to 35.71%. Table 4

Cite PhysScope

BibTeX
@misc{liu2026physscope,
  title = {{PhysScope}: Agentic Multimodal Physics Reasoning with Grounded Visual Memory},
  author = {Haoyang Liu and Kecheng Liang and Xinyu Li and Weixing Chen and Yang Liu},
  year = {2026},
  eprint = {2611.23010},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  doi = {10.48550/arXiv.2611.23010},
  url = {https://arxiv.org/abs/2611.23010}
}