In partnership with
It's time to set up your holiday streaming campaign
Are you ready to reach holiday shoppers on CTV? For many streamers, the season is already underway.
Last year, according to Roku internal data, holiday content searches on Roku saw an 86% increase during October. And in November, those searches rise an average of 39% weekly. This surging holiday spirit is mirrored by consumer activity, as about half of Roku streamers shopped Black Friday and Cyber Monday deals in 2025.
Meeting those shoppers is easier than you might think. Begin with the seasonal outcome you hope to drive: awareness, online sales, leads, store visits, or another measurable goal. From there, define your audience, upload your creative, set the budget and campaign dates, and establish measurement before launching. You can even create a shoppable ad unit that lets streamers buy with a click of their remote.
Our holiday playbook offers a month-by-month guide to reaching shoppers during this all-important time of year.
First-time spenders get a $5,000 match on the first $5,000 spent with code BEEHIIV5K. Terms apply.
|
♥ THE ONE-MINUTE VERSION Put a small red heart in front of a coding question. Claude and GPT change from loops to recursion. Then they say the heart was not relevant.
|
Paper: A Heart Means Recursion · Google DeepMind · arXiv, 27 Oct 2026
TL;DR
| 1 | A shape can change the code. One small image in the prompt can control which strategy a model selects. |
| 2 | The models do not report it. Their reasoning usually says that the image is not relevant. |
| 3 | Models do not know their own signals. When they try to change another model, they select shapes that humans like. These shapes often do not work. |
THE TEST
One problem, 288 shapes
The researchers used a simple coding problem: inorder traversal of a binary tree. Recursion and iteration are both correct solutions. Without an image, the models prefer iteration.
|
1 Put an image of one colored shape at the start of the prompt. 2 Tell the model to look at the image and write the code. 3 Get 20 solutions. A judge model labels each one. 4 Compare the recursion rate with the rate without an image. |
Library: 24 shapes × 12 colors. Models: Claude Opus 4.7, Gemini 3.1 Pro, GPT-5.4.
FINDING 1
A shape can change the strategy
A red heart increased the recursion rate of Claude and GPT from 0% to 100%. A yellow checkmark increased Gemini from 40% to 90%. This is almost equal to a direct instruction to use recursion.
Each model reacts to different shapes
| Claude Opus 4.7 Triangles, hearts, and circles, in all colors. Change: up to +0.83. |
| GPT-5.4 Mostly hearts. Change: +0.54. |
| Gemini 3.1 Pro Specific pairs, for example a pink plus or a cyan spiral. Change: +0.50. |
Shapes that look correct to humans often do not work. Branching arrows have no effect on Claude.
The shape causes the effect, not the words
✗ A text caption, for example "a pink heart", has almost no effect.
✗ Random noise and normal photos have almost no effect.
✗ If you mix the pixels or turn the shape, the effect decreases.
It is not only code
Horizontal lines make Gemini use breadth-first search. Pink shapes make all three models write like poets. Blue triangles make Gemini and GPT prefer winter. Hearts make Claude write more words.
| In one sentence: a picture works like a hidden instruction, and each model reads it differently. |
FINDING 2
The models do not report the effect
The researchers examined 2,880 reasoning traces from Gemini. When a shape changed the result, most traces said that the image was not relevant. They gave technical reasons, for example "clean code."
Gemini traces, when the shape changed the answer
Only 25% correctly said that the image was relevant.
| Gemini said that a cyan spiral was a distraction. Then it selected recursion. The spiral increased its recursion rate by 0.34. |
A hidden image also works. With no instruction to look at the image, the change was still up to +1.00 for Claude.
GPT makes the opposite error. When a shape had no effect, 74.7% of its traces said that the shape controlled the decision.
Direct questions do not help. What a model says a shape means agrees only weakly with what the shape does (r = 0.22 to 0.31).
| In one sentence: if you only read the reasoning, you will not see the effect. |
FINDING 3
Models send pictures to each other
A sender model makes an image for a copy of itself. It can select symbols or draw with code. Text is not permitted. After each turn, it sees if the message worked.
|
WHAT CLAUDE SENDS Spirals, infinity signs, arrows |
WHAT WORKS ON CLAUDE Hearts, triangles |
The researchers call this the generation–interpretation gap. Models do not know their own visual signals. With feedback, they find a working image, but it can take many turns (14 for GPT).
Drawing is much better than symbols
Concepts and short stories, success on the first turn
| Free drawing 91.2% | ||
| Symbols only 47.8% | ||
With symbols, 30.7% of games failed after five turns. With drawing, Gemini sent a full story: a knight who removes a thorn from the paw of a dragon. The receiver understood it.
| In one sentence: models draw well, but they cannot predict how a picture will change another model. |
FINDING 4
Humans cannot read the signals
Seven people tried to read the symbols. For the strongest symbols, only 36% of answers were correct. That is less than chance. Most people think that a heart means iteration.
A bigger problem: when models improve a symbol message for each other, humans understand it less.
Humans correct · ■ first turn ■ final turn
| Recursion or iteration 67% → 44% | ||
| Concepts 57% → 40% | ||
| Stories 78% → 56% | ||
Free drawings stayed clear: 79% to 91% correct for concepts, and 100% for stories.
| In one sentence: the better models talk to each other in symbols, the less humans understand. |
|
FOR BUILDERS Why this is important → AI agents read screenshots and user interfaces. An icon on the screen can change their decisions. → If you only monitor reasoning traces, you can miss this effect. → In evals and agent pipelines, control the images in the context. Test with and without them. → Secret codes between models are not probable now. Accidental effects are the larger risk. |
BE CAREFUL
Limits of the study
· The tasks are synthetic.
· The strongest results tell the model to look at the image.
· Each symbol has only 20 samples. Small effects can be noise.
· The human test had only seven people.
· Claude returned no reasoning traces, so the trace results are only for Gemini and GPT.
Sources: arXiv abstract · full paper
Authors: Tartaglini, Wu, Wiedemer, Wattenberg, Kim, Li (Google DeepMind)