acceptodds
Under review as a conference paper at ICLR 2027

Skill Co-activation Graphs: A Functional Topology of Generalization in Large Language Models

Abstract

Large language models (LLMs) exhibit strong generalization, yet it remains unclear how reusable predictive mechanisms are organized internally to support this behavior. We introduce the Skill Co-activation Graphs (SCGs), a functional-topology view in which nodes are skill-indexed components induced from token-level loss-gradient alignment and edges encode their empirical co-activation across inputs. Starting from the autoregressive objective, we derive a loss-induced skill-alignment formulation and establish a task-conditioned generalization bound whose complexity depends on , where captures SCG resistance and captures task-induced skill usage. We further extend this analysis to empirically estimated SCGs through a stability-aware graph-estimation bound. Experiments on Pythia with ARC-Challenge and SciQ reveal systematic relationships between SCG structure and downstream generalization. Larger Pythia models tend to exhibit richer and more stable skill organization, while lower held-out task-conditioned resistance is associated with higher task accuracy. Intervention-based analyses support the functional relevance of the induced skill subspaces. Across both Pythia and Qwen2.5, SCG-guided fine-tuning further shows that high-resistance examples provide a useful structural signal for data selection under limited adaptation budgets.

Primary area:
interpretability and explainable AI
Keywords:
Large Language Models, Mechanistic Interpretability, Skill Co-activation Graphs, Functional Topology

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_GPg5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for analyzing how LLMs organize reusable predictive mechanisms, which the authors call “skills.” Rather than starting from predefined architectural units such as attention heads, circuits, or neurons, the approach is data-driven: token-level loss gradients with respect to the last Transformer block are clustered, and each cluster is treated as a skill. The authors then build a Skill Co-activation Graph (SCG), connecting two skills whenever they co-occur in the same example. On the theoretical side, the paper derives a generalization bound for a graph-smooth skill-alignment function class, in which the complexity term depends on the task-conditioned resistance Tr(L⁺Σ_T) of the SCG. A stability-aware extension accounts for the graph being estimated from data. Empirically, across six Pythia model scales on ARC-Challenge and SciQ, lower task-conditioned resistance is strongly correlated with higher held-out accuracy. Ablating task-relevant skill subspaces degrades gold-answer likelihood more than ablating random subspaces of the same rank. Finally, the authors use per-example SCG resistance as a data-selection criterion for budget-limited fine-tuning, and report that selecting high-resistance examples yields accuracy gains over random, high-loss, and skill-coverage baselines on Pythia-2.8B and Qwen2.5-3B.

Critical strengths
  1. The idea is interesting and offers a new perspective on how LLMs work internally. Inducing “skills” directly from the model’s own loss-gradient geometry, rather than from predefined architectural units or external task labels, is a fresh, data-driven angle.

  2. The theoretical development is clear and intuitive. The derivations are complete, and the resulting task-conditioned bound, with the spectral interpretation of Tr(L+ΣT)\textrm{Tr}(L^+\Sigma_T), gives a thoughtful way to reason about how a task’s skill demands interact with the structure of the skill graph.

  3. The empirical results are consistent with the theory. The authors further turn their observations into a practical data-selection metric (SCG resistance) for budget-limited fine-tuning, which shows a potential use of the framework beyond analysis.

Critical weaknesses
  1. Some modeling choices are oversimplified and may undermine the main empirical conclusion. For example, the SCG connects two skills whenever they co-occur in at least one example. With many examples, this binary rule likely makes the graph nearly fully connected. In a fully connected graph, Rnorm reduces to exactly 1/K~1/\tilde{K}, so its value is determined mainly by the number of retained skills rather than by the graph’s structure.

  2. The fine-tuning gains are small and may not support the conclusion. For Qwen2.5-3B on ARC-Challenge, the improvement from SCG High-R is below 0.1 percentage points, roughly one question on the test set. Across settings, the error bars of several strategies overlap, with only three seeds. Statements such as High-R “consistently exceeds” the baselines therefore seem too strong. More seeds, significance tests, or larger evaluation sets would make this result more convincing.

Questions for authors
  1. Can you report K~\tilde{K}, Rnorm × K~\tilde{K}, edge density, and the number of connected components for every model–task pair, and explain any cases where Rnorm < 1/K~1/\tilde{K}?

  2. How do the results change with a weighted graph (using N_ij) instead of the binary existence rule?

  3. How are skill subspaces constructed for ablation, and why are the 70M/160M effects an order of magnitude larger than for the other models?

1 comment

Appreciate the detailed response provided by the authors. Most of my concerns have been adequately addressed, and I will maintain my positive evaluation score.

Reviewer_To5EOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper constructs Skill Co-activation Graph (SCG) by token-loss gradient clustering. The authors name it skill. Next, the authors introduce a metric called task-conditioned resistance to connect task requirements with graph structure. Using this, they analyze smooth graph functions and test how model scaling, subspace intervention, and data selection perform under a limited fine-tuning budget.

Critical strengths
  1. The graph-based analysis establishes a theoretical link between the graph's topology, task-specific skill, and the generalization bounds of graph-smooth functions.
Critical weaknesses
  1. Does the gradient clustering in the last Transformer block really reflects reusable skills? The lack of cross-layer and activation-based controls makes it difficult to evaluate the rationale behind using gradient representations and this specific layer choice.

  2. I doubt that whether a single co-occurrence within the same input distinguishes stable functional dependencies and the metric is easily confounded by text length, skill frequency, and the sample size used for graph construction. Frequency-corrected or alternative graph construction baselines are needed to verify whether the SCG genuinely captures functional relationships.

  3. While the formulation is mathematically grounded under the graph-smoothness assumption, the paper does not sufficiently validate whether skill relationships satisfy this assumption.

Questions for authors
  1. What's the motivation using token-level loss gradients from the last Transformer block for skill induction?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.