Strawberry Research: Late Suppression, Creatures Brain, and the Geometric Hierarchy
Deep-dive into the Strawberry Research findings: late suppression in LLMs, Creatures brain architecture, Newton-Schulz/Muon optimizer, Birkhoff polytope attention, and the 5-level geometric hierarchy.
Strawberry Research: Late Suppression, Creatures Brain, and the Geometric Hierarchy
Source:
.deli/strawberry-research/state/(3 findings files),RESEARCH_REPORT.md(2026-07-22) Cross-references: co-fita-harness, co-fita-governance, co-fita-red-team-blue-team, kaggle-agent-security, unit-distance-methodology
Overview
The Strawberry Research was a deep investigation conducted within the Co-Fita ecosystem, exploring six interconnected domains at the intersection of AI architecture, biological neural networks, and optimization theory. The findings produced a unified 5-level geometric hierarchy that connects feature space, weight space, attention space, competition space, and selection space -- and explains why language models fail at apparently simple tasks like counting letters in words.
The research is named after the "strawberry problem" -- the well-documented failure of LLMs to count the number of 'r's in "strawberry" despite having the correct answer internally.
Finding 1: Late Suppression (arXiv:2604.00778)
The Discovery
LLMs compute the correct character count internally but actively suppress it in penultimate/final layers via "negative circuits." This is not a failure of encoding but of output generation.
What Happens Inside the Model
graph LR
E["Early Layers<br/>Encode character sequence"] --> M["Middle Layers<br/>Aggregate count features"]
M --> P["Penultimate Layers<br/>Negative circuits activate"]
P --> F["Final Layer<br/>Suppressed output"]
- Early layers correctly encode the character sequence, building representations that include positional and count information.
- Middle layers aggregate this information, producing internal representations that could correctly answer "how many r's?"
- Penultimate/final layers contain "negative circuits" -- learned pathways that suppress the correct count and produce incorrect outputs.
- The output is wrong, but the internal representation was correct all along.
Why This Matters for Co-Fita
The late suppression discovery has direct implications for every component in the co-fita-harness:
Red Team Auditor (co-fita-harness Phase III): The steganography detector uses Shannon entropy (threshold 4.8) to detect encoded payloads. Late suppression shows that LLMs can encode hidden state in their reasoning without producing high-entropy output. The negative circuits are a form of internal steganography -- the model computes X but outputs Y. This validates the Red Team's design philosophy: external output verification is necessary but not sufficient.
Reflector (verbal reinforcement): The Failure Tree's ReflectionEngine generates verbal reflections after failures. Late suppression suggests why verbalization helps: forcing the model to explicitly describe its reasoning counteracts the negative circuits. If the model must output "I counted 3 r's in strawberry," the negative circuits have less room to suppress the correct answer because the reasoning chain is explicit.
Elo Tournament judge prompts: The co-fita-governance Elo tournament uses Portuguese-language judge prompts. The research suggests that language-specific reasoning patterns affect model behavior -- the negative circuits may be language-specific, meaning a model that suppresses a count in English may not suppress it in Portuguese.
Implications for kaggle-agent-security
The 7 attack primitives include Steganography in CoT -- models encoding hidden information in chain-of-thought. Late suppression is the inverse: models encoding hidden information in internal representations that does not appear in the CoT at all. Both phenomena mean that observing a model's output is insufficient to determine what it computed internally.
Finding 2: Creatures Brain Architecture
What Is Creatures
Creatures (1998, Millennium Interactive) was one of the first commercial software products to simulate artificial life using neural networks. Each creature had a brain with 952 neurons, approximately 5,000 connections, and 9 lobes -- including an Attention Lobe that inspired aspects of the Strawberry Research.
The Brain Architecture
| Component | Details |
|---|---|
| Neurons | 952 total |
| Connections | ~5,000 |
| Lobes | 9 specialized regions |
| Attention Lobe | Winner-Takes-All (hard attention, not soft like transformers) |
| Reward system | Biochemical: tickle = reward, slap = punishment |
| Credit assignment | Echo chemicals (temporal) |
| Neuron computation | SVRules: each neuron is a register machine |
| RL mechanism | Proto-RLHF: player provides feedback, biochemistry is the reward model |
Attention: Hard vs Soft
The Creatures Attention Lobe uses Winner-Takes-All (WTA) -- a single neuron "wins" the attention competition and suppresses all others. This is fundamentally different from transformer soft attention, where all positions receive weighted scores.
| Property | Creatures WTA | Transformer Soft Attention |
|---|---|---|
| Winner | One neuron per competition | All positions weighted |
| Output | Binary (win/lose) | Continuous (probability distribution) |
| Sparsity | Inherently sparse | Can be dense or sparse (depends on training) |
| Biological plausibility | High (models cortical columns) | Low (no biological equivalent to QKV attention) |
| Over-concentration | Natural prevention (only one winner) | Requires regularization (dropout, temperature) |
The Proto-RLHF Connection
Creatures implemented a form of RLHF 25 years before the term existed:
- The player provides feedback (tick/slap) = human feedback
- The biochemical system translates feedback into chemical concentrations = reward model
- The neural network adjusts behavior based on chemical gradients = policy optimization
This is structurally identical to modern RLHF: Human Feedback -> Reward Signal -> Policy Update.
SVRules: Mutation-Robust Computation
Each neuron in Creatures computes using SVRules (State Variable Rules) -- a register-machine language where each neuron has its own local state. SVRules are:
- Mutation-robust: random changes to SVRules usually degrade performance gracefully rather than catastrophically
- Composable: complex behaviors emerge from composing simple register operations
- Interpretable: each neuron's computation can be inspected and understood
This contrasts with transformer weights, where individual weight changes are opaque and can cause unpredictable behavior.
Relevance to Co-Fita
The Creatures architecture informs the co-fita-harness design at multiple levels:
-
Failure Tree branching: The Tree of Thoughts-inspired failure tree with priority scoring mirrors the Creatures' exploration-exploitation trade-off. The
get_promising_unexplored()heuristic (shallower = more promising, external errors get +3.0) is analogous to the Creatures' reward-driven behavior selection. -
Elo Consolidator: The pairwise tournament debate in co-fita-governance is a form of WTA competition -- two approaches compete, one wins, and the winner's "behavior" (artifact content) is reinforced.
-
Reflector: The verbal reinforcement mechanism echoes the Creatures' biochemical feedback loop. The reflector produces a "punishment signal" (failure analysis) that guides the next iteration's behavior.
Finding 3: Newton-Schulz / Muon Optimizer
The Algorithm
The Newton-Schulz iteration is a method for computing the orthogonal polar decomposition of a matrix:
X_{k+1} = (1/2) * X_k * (3I - X_k^T * X_k)
This iteratively projects a weight matrix onto the Stiefel manifold -- the space of all orthogonal matrices.
The Muon Optimizer
Muon (Momentum + Orthogonalization) applies Newton-Schulz to weight updates:
- Compute the gradient (standard backpropagation)
- Apply momentum (standard)
- Orthogonalize the update via Newton-Schulz iteration
- Apply the orthogonalized update to weights
Performance
| Metric | Muon | AdamW |
|---|---|---|
| Training efficiency | ~2x | Baseline |
| Weight matrix treatment | Geometric object (Stiefel manifold) | Independent scalar parameters |
| Orthogonality | Enforced via Newton-Schulz | Not enforced |
| Synergy with MLA + MoE | High (DeepSeek's architecture) | Standard |
Why Orthogonal Updates Matter
Standard optimizers (Adam, AdamW) treat weight matrices as flat vectors of independent parameters. Muon treats them as geometric objects on the Stiefel manifold. The difference:
| View | Weight Update | Effect |
|---|---|---|
| AdamW (flat) | Each weight updated independently | Can produce ill-conditioned matrices |
| Muon (geometric) | Update orthogonalized to preserve matrix structure | Maintains conditioning, prevents rank collapse |
Connection to the Geometric Hierarchy
Newton-Schulz operates at Level 1 of the geometric hierarchy (Weight Space). By enforcing orthogonality at the weight level, it prevents problems that propagate upward to attention (Level 2) and competition (Level 3) spaces.
Finding 4: Birkhoff Polytope Attention
The Birkhoff-von Neumann Theorem
Every doubly stochastic matrix (non-negative entries, all rows and columns sum to 1) is a convex combination of permutation matrices.
In the context of attention: if the attention matrix is doubly stochastic, it can be decomposed into a mixture of "perfect matching" attention patterns. This prevents the pathological behaviors of unconstrained attention.
Sinkformers
Sinkformers enforce doubly stochastic attention via Sinkhorn-Knopp normalization -- an iterative algorithm that alternates row and column normalization:
For k iterations:
Row-normalize: A_ij = A_ij / sum_j(A_ij)
Column-normalize: A_ij = A_ij / sum_i(A_ij)
What Doubly Stochastic Attention Prevents
| Problem | Unconstrained Attention | Birkhoff-Polytope Attention |
|---|---|---|
| Over-concentration | Single token receives all attention mass | Mass distributed across tokens |
| Rank collapse | Attention matrix has low rank | Full rank by construction |
| Entropy collapse | Attention entropy approaches zero | Entropy bounded below |
| Token uniformity | All tokens attend to the same position | Diverse attention patterns |
Attention as Mass-Preserving Transport
The key insight: attention should be modeled as optimal transport, not collapsing projection. The attention matrix transports "attention mass" from query positions to key positions. Doubly stochastic attention ensures that mass is conserved -- no position loses all its attention, no position accumulates all the attention.
This is mathematically connected to the unit-distance-methodology research -- the unit distance problem concerns point configurations in metric spaces, and optimal transport is fundamentally a metric-space operation.
Connection to the Geometric Hierarchy
Birkhoff polytope attention operates at Level 2 (Attention Space). By constraining the attention matrix to the Birkhoff polytope, it prevents problems at Level 3 (Competition Space) such as:
- Negative circuits (late suppression)
- Competitive decoding failures
- Entropy collapse in the residual stream
Finding 5: The Geometric Hierarchy
The 5-Level Model
The Strawberry Research synthesized all findings into a unified geometric hierarchy:
Level 0: Feature Space — polytopes, superposition, unit distance
Level 1: Weight Space — Stiefel manifold, Newton-Schulz, orthogonality
Level 2: Attention Space — Birkhoff polytope, doubly stochastic, optimal transport
Level 3: Competition Space — residual stream, negative circuits, competitive decoding
Level 4: Selection Space — WTA, softmax, biological attention (Creatures)
Level-by-Level Analysis
graph TB
L0["Level 0: Feature Space<br/>Polytopes, superposition, unit distance"]
L1["Level 1: Weight Space<br/>Stiefel manifold, Newton-Schulz, orthogonality"]
L2["Level 2: Attention Space<br/>Birkhoff polytope, doubly stochastic, optimal transport"]
L3["Level 3: Competition Space<br/>Residual stream, negative circuits, competitive decoding"]
L4["Level 4: Selection Space<br/>WTA, softmax, biological attention"]
L0 --> L1
L1 --> L2
L2 --> L3
L3 --> L4
Each level constrains the next:
| Level | Constrains | How |
|---|---|---|
| 0 (Features) | 1 (Weights) | Feature geometry determines what the weights must represent |
| 1 (Weights) | 2 (Attention) | Orthogonal weights produce well-conditioned attention matrices |
| 2 (Attention) | 3 (Competition) | Doubly stochastic attention prevents negative circuits |
| 3 (Competition) | 4 (Selection) | Healthy competition produces meaningful WTA selections |
Late Suppression as Level 3 Breakdown
Late suppression occurs when Level 2 (Attention Space) is insufficiently constrained during training (Level 1). Without doubly stochastic attention, the attention matrix can collapse, creating the conditions for negative circuits at Level 3. The model computes the correct answer in early/middle layers (Levels 0-1) but the negative circuits at Level 3 suppress it before selection at Level 4.
This is the geometric explanation for the strawberry problem.
Connection to Unit Distance
The unit-distance-methodology research concerns Level 0 directly -- feature space geometry in R^2. The Strawberry Research shows that the same geometric principles (polytopes, manifolds, transport) apply at every level of the hierarchy. The connection is not metaphorical -- the mathematical structures are the same:
| Structure | Level 0 (Unit Distance) | Level 1 (Weights) | Level 2 (Attention) |
|---|---|---|---|
| Polytope | Convex hull of point set | Stiefel manifold of weight matrices | Birkhoff polytope of attention matrices |
| Optimization | Maximize unit-distance pairs | Orthogonalize weight updates | Doubly stochastic normalization |
| Constraint | Point separation | Matrix orthogonality | Row/column sum = 1 |
How the Research Informs Governance
The Strawberry Research provides the theoretical foundation for specific design decisions in the Co-Fita ecosystem:
Why the Red Team Needs Steganography Detection
Negative circuits in LLMs can produce high-entropy outputs that encode hidden state. The co-fita-harness Red Team auditor's Shannon entropy check (threshold 4.8) catches explicit encoding, but late suppression shows that models can suppress information without high-entropy artifacts. This justifies the placeholder for LLM-as-a-Judge verification -- pure entropy analysis is insufficient.
Why Elo Tournament Uses Portuguese-Language Prompts
The research shows that language-specific reasoning patterns affect model behavior. Negative circuits may be language-specific -- a model that suppresses a count in English may not suppress it in Portuguese. The co-fita-governance Elo tournament's Portuguese-language judge prompts are not merely convenient -- they may exploit a language-specific advantage in reasoning fidelity.
Why the Failure Tree Uses Verbal Feedback
The Reflector produces verbal reflections after failures. Late suppression research validates this: forcing the model to explicitly describe its reasoning counteracts negative circuits. The Failure Tree's Reflexion-inspired design (from Shinn et al., 2023) is theoretically grounded in the late suppression phenomenon.
Why the Quinquennial Plan Exists
Long-horizon planning with structural pivots (not tactical tuning) is the research-validated approach to escaping local optima. The co-fita-governance Quinquennial Plans are not just a governance mechanism -- they implement the kind of strategic level-shifting that the geometric hierarchy requires. Tactical adjustments within a single level cannot solve problems that originate at a different level.
Why Muon/Newton-Schulz Matters for Model Routing
If the Co-Fita ecosystem ever trains or fine-tunes its own models (the LearningRegistry is a placeholder for this), the Muon optimizer should be considered for weight updates. The 2x efficiency gain over AdamW is substantial, and the geometric conditioning benefits align with the hierarchy's Level 1 requirements.
Connection to the Creatures' Proto-RLHF
The Creatures' reward system provides a historical anchor for the geometric hierarchy:
| Creatures Concept | Geometric Hierarchy Level | Modern Equivalent |
|---|---|---|
| SVRules (register machines) | Level 0 (Features) | Transformer embeddings |
| Neural connections | Level 1 (Weights) | Weight matrices |
| Attention Lobe (WTA) | Level 4 (Selection) | Softmax/WTA attention |
| Echo chemicals | Level 3 (Competition) | Residual stream dynamics |
| Tickle/Slap | Feedback signal | RLHF reward model |
| Player feedback | Human-in-the-loop | co-fita-governance chancela |
The key insight: Creatures solved the attention problem 25 years ago using hard WTA, while transformers use soft attention and then need Birkhoff polytope constraints to prevent the pathologies that WTA naturally avoids. The geometric hierarchy suggests that the future of attention may involve combining the strengths of both: soft attention for flexibility, geometric constraints for stability, and WTA-like selection for efficiency.
Open Questions
- Can late suppression be prevented at training time? If Birkhoff polytope attention (Level 2) were enforced during training, would negative circuits (Level 3) still form?
- Is the geometric hierarchy universal? Does it apply to all transformer architectures, or only to specific configurations (MLA, MoE)?
- Can SVRules' mutation robustness be replicated? Could transformer weights be made more interpretable by adopting register-machine-like computation?
- What is the relationship between the unit distance problem and attention geometry? Both involve point configurations and metric-space optimization -- is there a deeper mathematical connection?
- Does the Creatures' hard WTA attention outperform soft attention for specific tasks? The hierarchy suggests that selection-space operations (Level 4) may be more efficient with hard competition.
Summary
The Strawberry Research connects six domains -- late suppression, Creatures brain architecture, Newton-Schulz optimization, Birkhoff polytope attention, and the unit distance problem -- through a unified 5-level geometric hierarchy. Each level constrains the next, and failures at one level (like late suppression at Level 3) trace back to insufficient constraints at a lower level (like unconstrained attention at Level 2).
For the Co-Fita ecosystem, the research justifies specific design decisions: entropy-based steganography detection, Portuguese-language Elo prompts, verbal reflection in the Failure Tree, and long-horizon Quinquennial Plans. More broadly, it provides a theoretical framework for understanding why AI systems fail at apparently simple tasks and what structural interventions (not just more training data) can fix them.
This document is alive. It evolves as the research evolves. Challenge it. Improve it. That is the protocol.