A Counterintuitive Experiment
Imagine a four-person math team solving a function-composition problem: multiply by 3, add 5, then mod 17. The rule: each person can only see the description of one operation, cannot look at anyone else's work, and can only communicate via notes.
Intuitively, this should perform far worse than four people who can all see the complete problem. How could less information lead to better results?
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 demonstrates that in multi-module language model systems, restricting each module's input visibility actually forces the system to learn genuinely compositional representations, while letting all modules see everything induces rote-memorization shortcuts.
This is not a marginal advantage. In 10 paired experiments, 9 of the "restricted" versions crushed their "fully visible" twins by over 20 percentage points, with median paired advantages of 0.7648 at depth 2 and 0.6050 at depth 3.
Experimental Design: A "Shared Genome" Society
The elegance lies in the control of variables. The author built a "four-cell society":
- Shared brain: all 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)
- Communication channel: cells exchange information only through two model-width continuous vectors ("packets") — the sole cross-cell pathway
- Task: function composition described in natural language — 0-3 affine bijection operations over Z17, a 17-class classification task
- Restricted group: each cell sees only its assigned segment of evidence slots + 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 2 and depth 3
- Median paired advantage at depth 2: 0.7648
- Median paired advantage at depth 3: 0.6050
- Initialization-level mean differences were all positive, with cluster-bootstrap 95% CIs of [0.628, 0.784] (depth 2) and [0.397, 0.669] (depth 3)
- Deleting packets collapses it — it too relies on communication
- But same-value transplant success is only 0.12-0.25 — its packets are not value-indexed but episode-specific
The key is the only manipulated 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 Ones Win
Paired Domination
Across 10 pairs (5 initializations × 2 data orders):
Communication Is Necessary
Cutting all packet communication dropped the restricted group to random level (1/17 ≈ 0.059). The restricted group's advantage depends entirely on inter-cell communication — no single cell learned the task alone.
Generalization, Not Memorization
The crucial test: on composite functions never seen in training, the depth-3 paired advantage remained 0.558. The restricted group isn't doing table lookup — it genuinely learned to compose operations.
Mechanism: Packet Transplant Experiments
Behavioral data only tells you *who* won, not *why*. The author designed an elegant 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 — and transplanting them into a target episode leaves system behavior unchanged. In 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
Swapping a packet for one representing a different mathematical value steers the system's output toward the mathematically predicted answer. This is causality, not correlation.
Destructive Interventions
Deleting packets or replacing them with noise collapses performance.
The Fully Visible Exception
One of the 10 fully visible runs (initialization 204 / order 954) reached 0.843 at depth 3, even beating its restricted twin. But transplant experiments revealed a completely different strategy:
It's like a student acing an exam — until you discover they memorized every answer rather than understanding the method. The transplant experiment is the litmus test separating these two strategies.
Why "Not Seeing" Works Better
The paper's core hypothesis:
> When a trainable module can inspect the whole program, "whole-program lookup" is an available solution. Restricting each module to a fragment removes that shortcut, forcing the system to develop reusable local transformations connected via communication.
The fully visible group has the whole-program-lookup route, so gradient descent takes it — it's the simplest. The restricted group has no such route and is forced to develop a generic interface that encodes intermediate values into packets.
Less information means a smaller solution space — but the remaining solutions are more likely to be the ones you want.
This resonates with information bottleneck theory and Dropout's regularization, but goes further: Dropout randomly drops information, whereas this is a stable semantic ownership boundary — 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. All thresholds and criteria were written down before training. The author then reported two formal failures:
1. Current experimental cohort: restricted depth-3 median accuracy 0.6988, below the pre-registered threshold of 0.70 — missing by 0.0012 2. Early eligibility cohort: 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)
Rather than hiding these failures, the author reports them fully and limits 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-call chains — default to giving every module the full context. This paper says: this may be an unexamined assumption. Restricting visibility could be a nearly free regularization mechanism.
2. Communication Bandwidth Should Be Small but Semantic
Two 896-dimensional vectors (~1.8K parameters) suffice to carry an intermediate value over Z17. What matters is not bandwidth but whether packets encode reusable semantic intermediate values. Transplant experiments are the tool for verifying this.
3. Passing the Training Set ≠ Learning the Method
The fully visible group also passes the training set — one even hit 0.843 at depth 3. But transplant experiments revealed an episode-specific lookup strategy. Pass rates mask strategy differences — another instance of the "evaluation blind spot" problem.
4. Pre-Registration Is Feasible and Valuable in AI Research
This paper demonstrates the feasibility of pre-registration in ML research: sealed task instances, frozen thresholds, reported failures. In an era of rampant p-hacking, this practice deserves wider adoption.
Conceptual Positioning
A new member of the "division of labor beats unification" principle: from octopus-style DNA pretraining + RNA inference-time computation, to Euclid-MCP's "LLM as poet, Prolog as accountant," to small-model triage + large-model review — all say the same thing: specialized parts doing specialized jobs beat one omnipotent part. This paper pushes the principle to a more radical position: not just division of labor, but restricting each part's input visibility, forcing information through interfaces rather than implicitly through shared context.
Another instance of evaluation blind spots: the fully visible group passes the training set, but transplant experiments reveal lookup. A single metric (accuracy) masks strategy differences; mechanistic tests are needed to distinguish them.
A "solve at a different level" variant: rather than making the model learn composition better, remove non-compositional solutions by restricting visibility, making composition the only path. Same spirit as Möbius RoPE's topological intervention — don't optimize the objective function; change the solution space itself.
Open Source
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 "bottlenecked" — working memory is tiny (7±2), attention highly selective. But this is not a bug; it's a feature. Precisely because of the information bottleneck, we are forced to develop abstraction, categorization, symbolization. If every brain region saw all inputs, we might only memorize specific scenes and never abstract the concept of "table."
This paper reproduces that logic at micro scale: the information bottleneck is not a limit on intelligence, but the condition for compositional representations.
In the LLM era we keep chasing larger context windows and more information access. This paper reminds us: sometimes, giving less information forces more general solutions. Like exams that ban calculators — you're forced to learn the principles of mental arithmetic.
---
Paper: arXiv:2608.20054 Code: github.com/tokenosopher/populus-evidence-partitioning Author: Narcis Marincat Published: 2026-08-20