A Counterintuitive Experiment
Imagine a four-person math team solving a function composition problem: multiply by 3, add 5, then take the result mod 17. The rule: each person can only see the description of one operation, can't look at anyone else's, and can only communicate via passed notes.
Intuitively, this should be far worse than four people all seeing the complete problem. With less information, how could it possibly do better?
On August 20, 2026, independent researcher Narcis Marincat published a paper on arXiv titled "What You Can't See Is What You Learn". Using an exquisitely designed experiment, he demonstrated that in multi-module language model systems, restricting each module's input visibility actually led the system to learn genuinely compositional representations, while letting all modules see everything instead induced a "rote memorization" shortcut.
This is not a marginal advantage. In 10 paired experiments, 9 of the restricted versions crushed their fully visible counterparts by more than 20 percentage points, with median paired advantages of 0.7648 at depth two and 0.6050 at depth three.
Experimental Design: A "Shared Genome" Society
The brilliance of this experiment lies in its control of variables. The author built a "four-cell society":
- Shared brain: the four cells share one frozen Qwen2.5-0.5B-Instruct language model (the "genome") and one rank-8 LoRA adapter (the only trainable parameters, ~4.32M params)
- Communication channel: cells can only pass information through two model-width continuous vectors ("packets") — the only cross-cell channel
- Task: natural-language-described function composition — 0-3 affine bijection operations on Z₁₇, a 17-class classification task
- Restricted group: each cell sees only its assigned evidence slot segment + the shared question
- Fully visible group: each cell sees all four evidence slots
- In 9/10 pairs, the restricted group beat the fully visible group by at least 20 percentage points at both depth two and depth three
- Median paired advantage at depth two: 0.7648
- Median paired advantage at depth three: 0.6050
- Seed-level mean differences were all positive; cluster-bootstrap 95% CIs: [0.628, 0.784] (depth two) and [0.397, 0.669] (depth three)
- Deleting packets collapses it — it also depends on communication
- But same-value transplant success is only 0.12–0.25 — its packets are not value-indexed but episode-specific
- A new member of the "division of labor beats unification" principle: from octopus-style DNA pretraining + RNA inference-time compute, to Euclid-MCP's "LLM as poet, Prolog as accountant," to Rebucca's small-model screening + large-model review — all say the same thing: specialized parts doing specialized things beat one all-purpose part. This paper pushes the principle further: not just division of labor, but restricting what each part sees, forcing information to flow through interfaces rather than implicitly through shared context.
- Another instance of the "evaluation blind spot law": the fully visible group passes the training set, but transplants reveal lookup. Single metrics (accuracy) mask strategy differences; mechanistic tests are needed.
- A variant of "solving at a different level": instead of making the model learn composition better, remove non-compositional solutions by restricting visibility, making composition the only path. Same spirit as Möbius RoPE's topology intervention — don't optimize the objective; change the solution space itself.
The key is the single experimental variable: the attention mask.
Both groups share identical initialization bytes, training order, token layout, positional geometry, parameter count, and compute. The only difference is the attention mask.
It's like a twin study — identical twins in identical environments, except one wears a blindfold.
Results: The Blind Outperform
Paired Domination
Across 10 pairs (5 seeds × 2 data orders):
Communication Is Necessary
Cutting all packet communication dropped the restricted group to chance level (1/17 ≈ 0.059). The restricted group's advantage depends entirely on inter-cell communication — no cell learned the task independently.
Generalization, Not Memorization
The most crucial test: evaluating on composite functions never seen in training. The paired advantage at depth three remained 0.558. The restricted group isn't looking up tables — it truly learned to compose operations.
Mechanism: Packet Transplant Experiments
Behavioral data tells you who wins, not why. The author designed a beautiful mechanistic test: packet transplantation.
Same-Value Transplants
Taking packets from different episodes at the same interface position — as long as they represent the same mathematical intermediate value — transplanting them into a target episode leaves system behavior unchanged. Across 6 restricted runs, same-value transplant fidelity reached 0.94–1.00.
This means packets genuinely encode mathematical intermediate values — a value-indexed interface.
Counterfactual Transplants
Replacing packets with those representing different mathematical values steers the system's output toward the mathematically predicted answer. This is causation, not correlation.
Destructive Interventions
Deleting packets or replacing them with noise collapses performance.
The Fully Visible Exception
One of the 10 fully visible runs (seed 204 / order 954) reached 0.843 at depth three, even beating its restricted twin. But transplant experiments revealed a completely different strategy:
In other words, this fully visible run found a way to "memorize the context of this specific episode" rather than a "pass mathematical intermediate values" generic interface. It passes the training set, but not through true composition.
It's like a student acing the exam, but you discover they memorized the answer to every question rather than understanding the method. Transplant experiments are the touchstone distinguishing these two strategies.
Why Does "Not Seeing" Work Better?
The paper's core hypothesis:
> When a trainable module can inspect the full program, "whole-program lookup" is an available solution. Restricting each module to a fragment removes the whole-program lookup shortcut, forcing the system to develop reusable local transformations connected via communication.
The fully visible group has the whole-program lookup path, and gradient descent takes it — it's the easiest. The restricted group lacks this path and is forced to develop a generic interface that "encodes intermediate values into packets for the next cell."
Less information means a smaller solution space, but the remaining solutions are more likely the one you want.
This resonates spiritually with information bottleneck theory and Dropout's regularization in deep learning, but goes further: Dropout randomly drops information, while this imposes stable semantic ownership boundaries — each cell owns its fragment, and cross-cell information flows only through packets.
Pre-Registered Honesty
The most admirable detail: this is a pre-registered study.
The author wrote down all thresholds and test criteria before training, then reported two formal failures:
1. Current experimental group: restricted depth-three median accuracy of 0.6988, below the pre-registered threshold of 0.70 — missing by 0.0012 2. Early eligibility group: 0/10 passed full gating; one model passed all task performance gates but all 10 failed the "natural language retention" gate (top-1 regression of 50–61 percentage points)
The author did not hide these failures but reported them fully, restricting the system to "explicit task-gated use." This scientific honesty makes the positive results more credible — this is not p-hacking.
Engineering Insights
1. "Default Full Visibility" Is Not Sacred
Nearly all current multi-module AI systems — MoE, multi-agent frameworks, tool-calling chains — default to giving every module the full context. This paper says: this may be an untested assumption. Restricting visibility could be a nearly free regularizer.
2. Small Bandwidth, But Semantic
Two 896-dimensional vectors (~1.8K params) suffice to convey a Z₁₇ intermediate value. The key isn't bandwidth size but whether packets encode reusable semantic intermediates. Transplant experiments are the tool to verify this.
3. Passing the Training Set ≠ Learning to Solve
Fully visible runs can pass the training set too — one even hit 0.843 at depth three. But transplant experiments revealed episode-specific lookup. Pass rates mask strategy differences — another instance of the "evaluation blind spot law."
4. Pre-Registration in AI Research Is Feasible and Valuable
This paper shows pre-registration works in ML research: sealed task instances, frozen thresholds, reported failures. In an era of rampant p-hacking, this deserves wider adoption.
Conceptual Positioning
Open-Source Code
The paper's artifacts are on GitHub: tokenosopher/populus-evidence-partitioning
Includes 19/20 final checkpoints, all evaluation and audit scripts, pre-registration materials, and event logs.
Personal Reflection
This paper evokes a deeper analogy: consciousness.
Human consciousness is "bottleneck-shaped" — working memory is tiny (7±2), attention highly selective. But this isn't a bug, it's a feature. Precisely because of the information bottleneck, we were forced to develop abstraction, categorization, and symbolization. If every brain region saw all inputs, we might only remember specific scenes and never abstract the concept of "table."
This paper reproduces that principle at micro scale: the information bottleneck is not a limit on intelligence, but a condition for producing compositional representations.
In the LLM era, we keep chasing bigger context windows and more information access. This paper reminds us: sometimes, giving less information forces more general solutions. Like when the teacher won't let you use a calculator, and you're forced to actually learn how arithmetic works.
---
Paper: arXiv:2608.20054 Code: github.com/tokenosopher/populus-evidence-partitioning Author: Narcis Marincat Published: 2026-08-20