You have been told to add "let's think step by step" to hard prompts. For math and logic, that advice holds. For anything with a shape, it backfires.

If your prompt involves a grid, a map, a floor plan, a board state, or a layout, chain-of-thought is working against you. The model narrates its way into a wall. It invents positions that were never in the prompt. And the more you let it think out loud, the worse the answer gets.

That is not a hunch. It is now one of the better-documented failures in prompting, and the fix is both more accurate and cheaper than what you are probably doing.

The evidence is not subtle

A paper out of IIT Hyderabad and Microsoft Research India (arXiv 2604.16060, April 2026) tested seventeen models across thirteen spatial benchmarks. Chain-of-thought prompting consistently degraded performance on visual-spatial tasks. Not "sometimes." Consistently.

The strangest part: models trained with reinforcement learning to reason well often did better when that reasoning was suppressed. Turning the thinking off raised the score.

The mechanism is what makes this worth your attention. In an ablation the authors call "No-Image++," CoT-prompted models hallucinated visual details from textual priors even when the image was gone. The model's chain of words drifts toward what layouts usually look like in text, not what this layout actually is. Every extra reasoning step is another chance to overwrite the real position with a plausible-sounding fake one.

Then there is the maze study (arXiv 2604.10690), which is the clearest single demonstration I have seen. Researchers gave models grid-world mazes in two formats and measured the swing.

Gemini-2.5-Flash, with chain-of-thought, on 5x5 mazes: - Fed as a visual grid: 34% accuracy. - Fed as a tokenized adjacency list: 86% accuracy.

On 7x7 mazes, same model, same prompt style: 16% as a grid, 80% as an adjacency list.

Nothing changed except how the space was encoded. Same maze, same question, same model. A 2x to 5x accuracy swing from representation alone. (Worth stating precisely: those numbers are Gemini-2.5-Flash. The pattern held for Claude-Haiku-4.5 and DeepSeek-Chat too, but GPT-5-mini did not follow it. Keep the model attached to the claim.)

Symbols beat sentences, and cost less

This is not new. The technique that fixes it, Chain-of-Symbol prompting (arXiv 2305.10276), has been around since 2023. It swaps natural-language spatial descriptions for condensed relational symbols.

On the Brick World task, GPT-3.5-Turbo went from 31.8% to 92.6% accuracy. That is a 60.8-point gain. And it did it while using 65.8% fewer prompt tokens, dropping one test prompt from 407 tokens to 139.

Sit with that inversion for a second. Normally accuracy costs money. Here the better approach is the cheaper one. You are paying more, right now, for prose that produces worse answers.

So the rule is simple. For any spatial task, do not narrate the space. Encode it.

Warehouse layout: the before and after

Say you want a model to find the shortest path across a small warehouse floor.

Here is the version most people write. It reads naturally, and it fails.

The Prompt (prose-CoT, worse):

Let's think step by step. The forklift starts at the top-left corner.
To its right is the receiving zone. To the right of receiving is packing.
Below packing is the shipping dock. There's a wall between receiving
and the aisle directly below it. What is the shortest path from the
forklift to shipping?

Watch what happens: the model re-describes the room in its own words, and somewhere in that re-description "below packing" quietly becomes "below receiving," and the wall moves. The answer is confident and wrong.

Now encode it.

The Prompt (symbolic, better):

Grid layout. Cells given as name(row,col):
forklift(0,0) receiving(0,1) packing(0,2) shipping(1,2)

Walls (impassable between): (0,1)-(1,1)

Adjacency (passable moves):
forklift -> receiving
receiving -> packing
packing -> shipping

Question: shortest path from forklift to shipping.
Answer with the ordered list of cells only. Do not restate the layout.

Why This Works: Every position and every legal move is pinned as an explicit symbol, so there is nothing for the model to re-narrate and corrupt. The "do not restate the layout" line matters: it suppresses the free-form CoT that the maze paper showed drives the collapse.

Expected Output:

forklift(0,0) -> receiving(0,1) -> packing(0,2) -> shipping(1,2)

Same information, different container. The prose version invites drift. The symbolic version leaves no room for it.

Board state: feed the adjacency, not the picture

Game boards trip the same wire. A board is a grid, and people love to hand the model an ASCII picture of it and ask it to reason across the drawing.

Don't. Give it the relationships directly.

The Prompt:

Tic-tac-toe. Board as coordinates, (row,col), 0-indexed.

X: (0,0) (1,1)
O: (0,2) (2,0)
Empty: (0,1) (1,0) (1,2) (2,1) (2,2)

Winning lines are any 3 cells sharing a row, column, or diagonal.
It is X's turn. Name the single move that gives X two separate
winning threats. Answer with one coordinate and the two lines it opens.

Why This Works: The model never has to visually parse a drawn board, which is the exact step where spatial reasoning breaks. Positions are listed as symbols and the win condition is stated as a rule, so the work becomes lookup and comparison instead of image interpretation.

Expected Output:

Move: (2,2). This gives X the main diagonal (0,0)-(1,1)-(2,2) and the bottom-right column threat via (2,2) in column 2. X now has two lines needing one cell each, which O cannot block in a single turn.

One more rule: you build the encoding, not the model

There is a tempting shortcut here: ask the model to convert the layout into a grid itself, then reason over its own conversion. Skip it.

A UC Santa Cruz study (arXiv 2604.14641) found a read-write asymmetry. Models interpret structured spatial representations well, but struggle to produce them from a text description, and those construction errors then flow straight into wrong answers. The model is a decent reader of a grid and a bad drawer of one.

So supply the structured encoding yourself. Coordinates, adjacency lists, explicit walls and moves. Hand the model the map already drawn and ask only the question.

When to keep the thinking on

To be clear, chain-of-thought is not broken. It is the wrong tool for one specific shape of problem. Math proofs, multi-hop logic, code tracing: keep thinking step by step there, it earns its cost.

The line to watch for is spatial structure. The moment "who is next to what" or "how do I get from here to there" enters the task, switch containers. Encode the space as symbols or an adjacency list, tell the model not to restate it, and stop paying the reasoning tax for a negative return. This covers more of your real work than it sounds: seating charts, store planograms, floor maps, org charts, delivery routes, any layout where adjacency is the whole question.

You will get better answers and a smaller bill at the same time. That combination is rare enough to act on.

If your team keeps hitting walls with spatial or layout prompts and can't tell which tasks are quietly paying the reasoning tax, we run live prompt engineering training that works through exactly these tradeoffs. Connect with Kief Studio on Discord or schedule a session.

LLM maze reasoning

Chain-of-Thought Is Sabotaging Your Model's Sense of Direction: Prompt Grids as Symbols, Not Sentences