What does vision buy
in a mirror maze?

Compare the same maps

The public text study contains 32 measured cases across four model and harness combinations. We match each case to a completed visual trial with the same map, initial pose, and goal. The text arm reached 23 goals. The visual arm reached 13.

Text agents received the grid rules once and then ordered action outcomes. Visual agents received first-person images and execution feedback. Both protocols allow action sequences and continue executing valid actions after a collision. Text cases use 200 Easy or 400 Medium planning rounds.

The comparison measures the observed systems under two protocols. DeepSeek uses the 1017 source in text and the alternate V4.1 Flash route in vision. Gemini's native AGY client versions and prompt reminder schedules also differ. Four visual Sol Medium successes retain their declared historical caps of 500, 500, 200, and 200. All four reached the goal within 400 calls. These differences limit a causal interpretation of image input alone.

Change the recorded budget

Restricting physical wall hits or executed actions changes the ordering. This calculation censors existing traces at the selected budgets. Fresh trials under those budgets could choose different actions.

One planning round can contain many primitive actions. The collision filter counts physical wall hits. Invalid sequences are recorded separately in the text data.

Look at the trajectories behind the totals

GPT-6.1 Sol with Codex passed all eight cases in both protocols. On Easy Hexagonal MIRROR-042 it used 315 actions and hit no walls with vision. With text it took 8,846 actions and hit 3,823 walls. On Easy Square MIRROR-042, Gemini 3.8 Flash with AGY finished in 9 text rounds with 268 wall hits, compared with 116 visual rounds and 40 wall hits.

The text feedback gives a direct answer about each attempted move, and a long batch can probe several directions before the next planning call. Visual feedback can help the agent avoid some of those probes, but reflections can also make passages ambiguous. Longer image histories and differences in the tested clients and model routes are additional explanations to separate experimentally.

Replay a text trajectory

The map is a reviewer view. During these episodes, the agent received action outcomes in text. Step through each recorded action or planning round. A red edge shows the wall that blocked a forward action.

Eight case comparisons

Each clip shows the four public model and harness combinations on one map. The paired rows use the same planning-round clock. The maps are reviewer diagrams. The counters distinguish executed actions, wall hits, and coverage.

Blind exploration changes the question

A fixed contour scan reaches all eight original goals within the recorded caps from the canonical initial headings using only successful and blocked movement outcomes. The original passage graphs are trees. The simulation below varies the initial heading and opens selected internal edges while holding the cells and goal fixed. A depth-first controller with relative position memory remains effective after adding cycles.

Collision-limited navigation and a same-pose wall decision measure different skills from goal reaching. The next controlled comparison should keep the model snapshot, client, feedback, and action budget fixed while changing mirror and matte rendering.

Reproduce the protocol

Each executed episode retains the declared round cap, action count, wall hits, coverage, and request overhead. Native provider request counts and host planning rounds use different boundaries.

The public repository contains the text-only evaluation commands, example configurations, and video recipes.