Visual reasoning stops rewriting programs each run: stored executable skills answer applicable inputs directly

A visual reasoning agent no longer has to implement or invoke a procedure on every run; it can adopt an already learned reusable visual program and return that program's answer on applicable inputs. Earlier approaches to learning reusable skills still required the agent to implement or invoke a procedure and produce the final answer, and errors or hallucinations at those steps kept acquired experience from translating into reliable single-run performance. Across four models, four benchmarks and 69 task families, VisualCoding exceeds the strongest evaluated baseline in mean accuracy by 4.45 percentage points with Gemini-3.1-Pro and by 14.78 points when the frozen Pro-acquired artifacts are reused with Flash-Lite.

The approach works in two stages: visual primitives and additional code are composed into family-level solvers, and a retained solver then returns its answer directly where the input is in scope. Execution-guided program acquisition iteratively tests and revises those solvers on labeled training instances, while scope-aware direct execution runs a retained solver and returns its valid in-scope answer. Ablations support execution-guided acquisition and show that direct execution outperforms textual or callable reuse of the same programs.

This report draws on summary-level material covering an evaluation of four models, four benchmarks and 69 task families, together with ablations of execution-guided acquisition and direct execution. It is aimed at researchers and engineers working on visual reasoning agents and reusable skills. Code is available at https://github.com/jiaq-liu/VisualCoding.

The summary states only that direct execution returns its valid in-scope answer, and does not say how out-of-scope inputs are handled. Ablations support execution-guided acquisition and show direct execution outperforms textual or callable reuse, but per-benchmark results are not given.