MATH · IN · MODELS
structures / Manifolds / Concept Cluster Heterogeneity

Concept Cluster Heterogeneity

PROPERTYempirical-patternintermediatehow it's classified →

A nominal concept class's activations, discovered via unsupervised clustering (k-means/GMM) of real hidden states, form multiple distinguishable sub-clusters rather than a single unimodal blob around one class mean — each sub-cluster corresponding to a semantically coherent subgroup (e.g. distinct instruction types within a 'harmful' class).

Replicationcomputed from the corpus — never hand-assigned
3 papers · no shared authors3 architecture classes · across papers2 domains · across papers6 model families · across papers
Filled = two or more values reported by papers that share no author — replication. Outlined = two or more values, but all from a single study — breadth, not replication. Grey = a single value. Derived from paper authorship and each model's architecture class, domain and family; it updates itself when a paper is added.

Definition

Given a labeled concept class CC (e.g. “harmful instructions,” “toxic text”), cluster its activations {xiC}\{x_i \in C\} with an unsupervised method (k-means, Gaussian Mixture Model) into K>1K>1 groups. Concept cluster heterogeneity is confirmed when K>1K>1 clusters are needed to fit the data well (vs. a single Gaussian), and post-hoc inspection shows each cluster corresponds to a coherent semantic subgroup rather than an arbitrary partition — meaning the class is a mixture of distinguishable sub-populations, not a single tight blob around one centroid.

Intuition

A single steering/editing vector computed as one global difference-in-means implicitly assumes a concept is a single point (its class mean) that can be reached by one fixed translation. If “harmful instructions” is really several distinguishable sub-types (violence, fraud, self-harm, etc.) occupying different regions of activation space, a single mean-difference vector is a poor summary of any of them individually — it is instead the centroid of a mixture, which may lie in a low-density region between the true sub-clusters.

Properties

  • Discovered, not assumed. The clustering is fit on real model activations and validated (e.g. cluster count selected via information criteria, or clusters shown to align with human-labeled subgroups on manual inspection) rather than a modeling choice imposed a priori.
  • Degenerate case recovers the single-direction picture. When K=1K=1 best fits the data, the concept is unimodal and a single Difference-in-means direction direction is an adequate summary — heterogeneity is a generalization of, not a replacement for, the single-direction case.
  • Low-rank steering-covariance structure. Empirically, the covariance of the per-cluster steering directions needed to reach a heterogeneous target concept can itself be low-rank (e.g. concentrated in the top 2(K1)2(K-1) principal components for KK clusters), meaning the heterogeneity, while real, is still a compact structure rather than requiring full-rank freedom.
  • Causal payoff. Exploiting the discovered cluster structure (e.g. via a per-cluster optimal-transport steering map, see GMM-Optimal-Transport steering (CHaRS)) rather than a single global mean-difference vector can produce a measurably stronger causal steering effect than the single-direction baseline.

Key evidence

Abdullaev, Wong, Lee et al. (2026) show PCA/t-SNE visualizations of last-token hidden states in Llama-3.2-3B-Instruct and Qwen2.5-Instruct models revealing clustered, multimodal structure within nominally single concept classes, with manual inspection confirming clusters correspond to coherent semantic subgroups (e.g. distinct categories of harmful instruction). They further show the steering-direction covariance implied by their per-cluster optimal-transport map is concentrated in the top 2(K1)2(K-1) principal components, matching a derived theoretical rank bound, and that exploiting this heterogeneity (via GMM-Optimal-Transport steering (CHaRS)) raises jailbreak attack-success rate and improves toxicity-mitigation trade-offs relative to single-direction (diff-in-means/ActAdd) steering baselines.

Anderson (2026) applies K-means + silhouette-score mode discovery (PCA dimensionality reduction, k-means sweep, maximum silhouette score) to chain-of-thought reasoning trajectories in real Llama-3-8B-Instruct and Llama-3.1-70B-Instruct on code-reasoning tasks, finding the population of trajectories increasingly resolves into a small number of discrete, well-separated clusters (“strategic modes”) with scale — maximum silhouette rising 0.13->0.42 from the 8B to the 70B model — the paper’s own framing being a discrete “Lattice” of modes, contrasted with legal reasoning’s smooth dimensional-collapse pattern at the same two scales (see Dimensional collapse). Here the “nominal concept class” being found heterogeneous is a single reasoning-trajectory population within one task domain, rather than a labeled semantic concept class. See anderson-2026-code-reasoning-trajectories-in-real-llama-3-8b-instruct-and-llama-3-1-70b-instruct-organize-into-a-small-number-of-discrete-silhouette-validated-clusters-that-sharpen-with-scale.

A real trained contrastive vision MoE-CNN’s experts cluster into a stable animate/inanimate partition via RSA, across expert counts and independently trained seeds. Tangtartharakul & Storrs (2026) train a custom sparsely-gated contrastive MoE-CNN (ResNet backbone, SimCLR, E in {4,8,16} experts) from scratch on STL10, and find agglomerative clustering of per-expert representational-similarity matrices (built on the independent THINGS dataset) recovers a robust 2-cluster animate/inanimate partition — here the “nominal concept class” being unexpectedly heterogeneous is not a single labeled behavior category but an entire population of expert representations, which nonetheless resolves into exactly the same kind of discovered, RSA-validated sub-cluster structure — stable across expert counts and 10 independently trained seeds, with complementary Lasso regression showing individual experts are broadly tuned to continuous dimensions beyond hard category boundaries. See tangtartharakul-storrs-2026-a-real-trained-contrastive-vision-moe-cnns-experts-cluster-into-a-stable-animate-inanimate-partition-via-rsa-across-expert-counts-and-independently-trained-seeds.

Found in (3 observations · 6 families)

Gemma

Concept Heterogeneity-aware Representation Steering (2026)measured

Concept classes are multi-cluster; GMM-OT steering beats single directions

Details

- Nominal concept classes used for activation steering (e.g. "harmful instructions") are internally heterogeneous: last-token hidden states form multiple semantically coherent sub-clusters (via k-means + PCA/t-SNE), not single Gaussian blobs. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - CHaRS steers by modeling source and target activations as Gaussian mixtures and applying an entropy-regularized Wasserstein optimal-transport map between cluster centroids (a barycentric combination of per-cluster diff-in-means shifts), degenerating to plain diff-in-means at one cluster. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - The steering-direction covariance is low-rank, rank(Sigma) <= 2K-2, motivating a spectral variant (CHaRS-PCT) that uses fewer directions. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - As a causal intervention (Activation Addition / Directional Ablation), CHaRS raises jailbreak attack-success rate up to 95.19% (Qwen2.5-7B-Instruct) over single-direction baselines, sweeping K=1 to 15 clusters. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - On RealToxicityPrompts it cuts toxicity up to 43% (CLS) and 42% (zero-shot) versus Linear-AcT while preserving perplexity/MMLU. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - Tested on seven instruct LLMs for jailbreak (Gemma2-9B, Llama3.1-8B, Llama3.2-3B, Qwen2.5-3B/7B/14B/32B), three for toxicity (Gemma2-2B, Llama3-8B, Qwen2.5-7B), plus FLUX.1 for image-style control (Mistral-7B as perplexity scorer). [abdullaev-etal-2026-concept-heterogeneity-representation-steering]

models: Gemma-2-9B-it, Gemma-2-2B · method: GMM-Optimal-Transport steering (CHaRS)

Llama

Concept Heterogeneity-aware Representation Steering (2026)measured

Concept classes are multi-cluster; GMM-OT steering beats single directions

Details

- Nominal concept classes used for activation steering (e.g. "harmful instructions") are internally heterogeneous: last-token hidden states form multiple semantically coherent sub-clusters (via k-means + PCA/t-SNE), not single Gaussian blobs. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - CHaRS steers by modeling source and target activations as Gaussian mixtures and applying an entropy-regularized Wasserstein optimal-transport map between cluster centroids (a barycentric combination of per-cluster diff-in-means shifts), degenerating to plain diff-in-means at one cluster. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - The steering-direction covariance is low-rank, rank(Sigma) <= 2K-2, motivating a spectral variant (CHaRS-PCT) that uses fewer directions. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - As a causal intervention (Activation Addition / Directional Ablation), CHaRS raises jailbreak attack-success rate up to 95.19% (Qwen2.5-7B-Instruct) over single-direction baselines, sweeping K=1 to 15 clusters. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - On RealToxicityPrompts it cuts toxicity up to 43% (CLS) and 42% (zero-shot) versus Linear-AcT while preserving perplexity/MMLU. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - Tested on seven instruct LLMs for jailbreak (Gemma2-9B, Llama3.1-8B, Llama3.2-3B, Qwen2.5-3B/7B/14B/32B), three for toxicity (Gemma2-2B, Llama3-8B, Qwen2.5-7B), plus FLUX.1 for image-style control (Mistral-7B as perplexity scorer). [abdullaev-etal-2026-concept-heterogeneity-representation-steering]

models: Llama-3.1-8B-Instruct, Llama-3.2-3B-Instruct, Llama-3-8B · method: GMM-Optimal-Transport steering (CHaRS)
The Geometry of Thought: How Scale Restructures Reasoning in Large Language Models (2026)measured

Code-reasoning trajectories form discrete clusters that sharpen with scale

Details

- Chain-of-thought hidden-state trajectories on code-reasoning tasks organize into a small number of discrete, silhouette-validated clusters ("strategic modes"), found via PCA plus a k-means sweep maximizing silhouette. [anderson-2026-geometry-of-thought] - The clustering sharpens with scale: the optimal cluster count rises k=2 (Llama-3-8B-Instruct) to k=5 (Llama-3.1-70B-Instruct) and the maximum silhouette from 0.133 to 0.417 (a 213% gain, the study's highest). [anderson-2026-geometry-of-thought] - This "Lattice" regime is one of three in the paper's scale-dependent phase diagram; other domains behave differently (Law collapses/aligns, Math and Science are scale-invariant), so the discrete-mode result is specific to code. [anderson-2026-geometry-of-thought] - Observational only, on two Llama-3 instruct checkpoints (8B, 70B); the author flags correlation-not-causation and runs no activation-level intervention. [anderson-2026-geometry-of-thought]

models: Llama-3-8B-Instruct, Llama-3.1-70B-Instruct · method: K-means + silhouette-score mode discovery, PCA

Qwen

Concept Heterogeneity-aware Representation Steering (2026)measured

Concept classes are multi-cluster; GMM-OT steering beats single directions

Details

- Nominal concept classes used for activation steering (e.g. "harmful instructions") are internally heterogeneous: last-token hidden states form multiple semantically coherent sub-clusters (via k-means + PCA/t-SNE), not single Gaussian blobs. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - CHaRS steers by modeling source and target activations as Gaussian mixtures and applying an entropy-regularized Wasserstein optimal-transport map between cluster centroids (a barycentric combination of per-cluster diff-in-means shifts), degenerating to plain diff-in-means at one cluster. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - The steering-direction covariance is low-rank, rank(Sigma) <= 2K-2, motivating a spectral variant (CHaRS-PCT) that uses fewer directions. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - As a causal intervention (Activation Addition / Directional Ablation), CHaRS raises jailbreak attack-success rate up to 95.19% (Qwen2.5-7B-Instruct) over single-direction baselines, sweeping K=1 to 15 clusters. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - On RealToxicityPrompts it cuts toxicity up to 43% (CLS) and 42% (zero-shot) versus Linear-AcT while preserving perplexity/MMLU. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - Tested on seven instruct LLMs for jailbreak (Gemma2-9B, Llama3.1-8B, Llama3.2-3B, Qwen2.5-3B/7B/14B/32B), three for toxicity (Gemma2-2B, Llama3-8B, Qwen2.5-7B), plus FLUX.1 for image-style control (Mistral-7B as perplexity scorer). [abdullaev-etal-2026-concept-heterogeneity-representation-steering]

models: Qwen2.5-3B-Instruct, Qwen2.5-7B-Instruct, Qwen2.5-14B-Instruct, Qwen2.5-32B · method: GMM-Optimal-Transport steering (CHaRS)

FLUX

Concept Heterogeneity-aware Representation Steering (2026)measured

Concept classes are multi-cluster; GMM-OT steering beats single directions

Details

- Nominal concept classes used for activation steering (e.g. "harmful instructions") are internally heterogeneous: last-token hidden states form multiple semantically coherent sub-clusters (via k-means + PCA/t-SNE), not single Gaussian blobs. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - CHaRS steers by modeling source and target activations as Gaussian mixtures and applying an entropy-regularized Wasserstein optimal-transport map between cluster centroids (a barycentric combination of per-cluster diff-in-means shifts), degenerating to plain diff-in-means at one cluster. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - The steering-direction covariance is low-rank, rank(Sigma) <= 2K-2, motivating a spectral variant (CHaRS-PCT) that uses fewer directions. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - As a causal intervention (Activation Addition / Directional Ablation), CHaRS raises jailbreak attack-success rate up to 95.19% (Qwen2.5-7B-Instruct) over single-direction baselines, sweeping K=1 to 15 clusters. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - On RealToxicityPrompts it cuts toxicity up to 43% (CLS) and 42% (zero-shot) versus Linear-AcT while preserving perplexity/MMLU. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - Tested on seven instruct LLMs for jailbreak (Gemma2-9B, Llama3.1-8B, Llama3.2-3B, Qwen2.5-3B/7B/14B/32B), three for toxicity (Gemma2-2B, Llama3-8B, Qwen2.5-7B), plus FLUX.1 for image-style control (Mistral-7B as perplexity scorer). [abdullaev-etal-2026-concept-heterogeneity-representation-steering]

models: FLUX.1 · method: GMM-Optimal-Transport steering (CHaRS)

Mistral

Concept Heterogeneity-aware Representation Steering (2026)measured

Concept classes are multi-cluster; GMM-OT steering beats single directions

Details

- Nominal concept classes used for activation steering (e.g. "harmful instructions") are internally heterogeneous: last-token hidden states form multiple semantically coherent sub-clusters (via k-means + PCA/t-SNE), not single Gaussian blobs. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - CHaRS steers by modeling source and target activations as Gaussian mixtures and applying an entropy-regularized Wasserstein optimal-transport map between cluster centroids (a barycentric combination of per-cluster diff-in-means shifts), degenerating to plain diff-in-means at one cluster. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - The steering-direction covariance is low-rank, rank(Sigma) <= 2K-2, motivating a spectral variant (CHaRS-PCT) that uses fewer directions. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - As a causal intervention (Activation Addition / Directional Ablation), CHaRS raises jailbreak attack-success rate up to 95.19% (Qwen2.5-7B-Instruct) over single-direction baselines, sweeping K=1 to 15 clusters. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - On RealToxicityPrompts it cuts toxicity up to 43% (CLS) and 42% (zero-shot) versus Linear-AcT while preserving perplexity/MMLU. [abdullaev-etal-2026-concept-heterogeneity-representation-steering] - Tested on seven instruct LLMs for jailbreak (Gemma2-9B, Llama3.1-8B, Llama3.2-3B, Qwen2.5-3B/7B/14B/32B), three for toxicity (Gemma2-2B, Llama3-8B, Qwen2.5-7B), plus FLUX.1 for image-style control (Mistral-7B as perplexity scorer). [abdullaev-etal-2026-concept-heterogeneity-representation-steering]

models: Mistral-7B · method: GMM-Optimal-Transport steering (CHaRS)

Custom Research Contrastive MoE-CNN (sparsely-gated ResNet trained with SimCLR)

Beyond Routing: Characterising Expert Tuning and Representation in Vision Mixture-of-Experts (2026)measured

Vision MoE-CNN experts partition into a stable animate/inanimate split

Details

- A custom contrastive vision MoE-CNN (ResNet backbone, SimCLR objective, ~46M params, E in {4,8,16} experts, top-2 routing) trained from scratch on STL10 develops experts whose representations partition into a stable animate/inanimate split. [tangtartharakul-storrs-2026-vision-moe-representation] - The partition is discovered by RSA: per-expert RDMs over the THINGS images (1,854 concepts) are second-order-correlated and agglomeratively clustered (silhouette-selected), recovering a robust two-cluster animate-vs-inanimate structure. [tangtartharakul-storrs-2026-vision-moe-representation] - The split is stable across all tested expert counts (4, 8, 16) and 10 independently trained seeds, even though top-2 routing does not select the same expert pair, mirroring a known organizational axis of human inferior temporal cortex. [tangtartharakul-storrs-2026-vision-moe-representation] - Nested-CV Lasso against 66 human visual/semantic dimensions shows individual experts are broadly tuned to continuous dimensions (e.g. "animal-related", "movement-related") rather than narrowly domain-specialized; observational, small-scale (STL10, self-supervised only). [tangtartharakul-storrs-2026-vision-moe-representation]

models: Sparsely-gated contrastive MoE-CNN, E=8 experts (Tangtartharakul & Storrs 2026) · method: Representational Similarity Analysis (RSA)