- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper investigates how to better organize features extracted by Sparse Autoencoders (SAEs) from Large Language Models (LLMs) and use the relationships between these features to improve SAE learning. Traditional SAEs primarily learn a set of features without explicitly modeling their relationships. While some structured SAE approaches organize features into tree structures, allowing each feature to have at most one parent, they may fail to capture relationships involving multiple parents. In addition, post-hoc graph construction methods can identify relationships between features, but these relationships do not directly guide SAE training. To address these limitations, the authors propose AG-SAE. The method constructs a directed acyclic graph (DAG) that supports multiple parents by comparing candidate parent sets based on feature co-activation, representation matching, and the additional contribution of each child feature. The resulting graph is then used to guide SAE training and feature refinement. The dictionary and graph are updated iteratively until both become stable. The authors evaluate AG-SAE on synthetic data and real LLM activations, reporting improvements in feature relationship recovery, structural reliability, reduced feature absorption, and feature-level interventions compared with baseline methods.
Critical strengths
(1) I think this paper presents an interesting methodological contribution. The authors go beyond the single-parent constraint of some structured SAEs by constructing feature graphs through the joint evaluation of candidate parent sets. The learned graph is then used to guide SAE training. After each round of dictionary refinement, the graph is reconstructed and used to guide further training. This process continues until both the dictionary and graph structure become stable. This iterative approach, which combines feature relationship learning with dictionary refinement, is meaningful.
(2) The experiments are comprehensive and provide good support for the proposed method. The authors evaluate parent-set recovery on synthetic data and further examine structural reliability, semantic validity, and feature-level interventions using real LLM activations. In particular, the ablation study in Section 4.5 shows that evaluating complete parent sets, rather than selecting parent-child edges independently, substantially improves the reliability of multi-parent relationships.
Critical weaknesses
None.
Questions for authors
-
In the main experiment with 30,720 features, how much additional training time and GPU memory does AG-SAE require compared with Vanilla SAE?
-
Could you provide an analysis of the time and space complexity of AG-SAE, particularly how the computational cost scales with the number of SAE features?
-
Could you provide an additional ablation study showing how performance changes when Residual Completion is removed? Since this module can introduce new SAE features, could you also clarify whether the improvements are partly due to increased dictionary capacity?
7 comments
Response to Reviewer G8HL — Q2 (Part 1/3)
Q2: Could you provide an analysis of the time and space complexity of AG-SAE, particularly how the computational cost scales with the number of SAE features?
A: Thank you for raising this important question about computational cost and scalability. As reported in our response to the previous question, under the main experimental setting with Gemma-2-2B and 30,720 features, AG-SAE takes 15.66 hours on a single NVIDIA A100-SXM4-80GB GPU, with 8.2149 GiB of peak allocated GPU memory. These measurements support the practical feasibility of AG-SAE at the evaluated scale. To further address this concern, we provide an analysis of how its time and space requirements scale with the number of features.
We provide the detailed time complexity analysis in the next comment, followed by the space complexity analysis in a separate comment.
Response to Reviewer G8HL — Q2 (Part 2/3)
Let denote the maximum total number of features during a run, the activation dimension, the batch size, and the maximum parent-set size. The following analysis focuses on scaling with , holding , , the dataset size and sampling budgets, the candidate-pool size, , and the numerical fitting budgets fixed.
Time complexity. We divide AG-SAE into two parts, (a) dictionary training with a fixed graph and (b) global refinement and graph re-induction between training phases.
(a) During dictionary training with a fixed graph, dense encoding and reconstruction require computation per step under the standard SAE encoder–decoder formulation (Gao et al., 2025), scaling linearly with for fixed and . AG-SAE additionally computes the structural loss and graph-conditioned decoding. For the structural loss, AG-SAE uses an exact-search pool of 12 features and a maximum parent-set size of 3 (Appendix E.4), so each child feature considers at most nonempty candidate parent sets, together with the empty-set comparator. These sets remain fixed during the training phase, and the numbers of parents and comparison sets associated with each feature do not grow with . Consequently, computing the structural loss over all features requires time under the fixed budgets above. For graph-conditioned decoding, features are processed in parent-before-child order, allowing already computed parent decoder directions to be reused. Each feature is computed once by combining at most parents, requiring operations over all features, which is also linear in for fixed and . Thus, the combined cost of dense encoding and reconstruction, the structural loss, and graph-conditioned decoding remains linear in . Including BatchTopK selection, a conservative sorting-based bound is for fixed . Combining these terms therefore gives an time upper bound per fixed-graph training update, while the main matrix computations remain linear in the number of features.
(b) For each global update, we first analyze graph re-induction and then account for residual completion. Graph re-induction begins with all-pairs scoring to retrieve candidate parents, comparing each feature with the remaining features and involving at most feature pairs. With a fixed per-pair scoring cost, this requires time. During retrieval, each child maintains only a fixed-size pool of the best candidates, avoiding a full sort of all pairwise scores and preserving this time bound. This fixed-size pool also bounds the subsequent parent-set search. Specifically, with an exact-search pool of 12 features and a maximum parent-set size of 3, each child evaluates at most 298 nonempty parent sets. The total number of candidate parent sets across all features is therefore , and their fitting and evaluation require time under fixed sampling and numerical fitting budgets. Acyclic graph selection then sorts these scored candidate parent sets in time and performs reachability checks in order, checking each candidate at most once. Since the graph contains at most nodes and edges, each check requires at most time. Combined with candidates and fixed , all checks have an time upper bound. Thus, candidate retrieval, parent-set evaluation, and acyclic graph selection together give an upper bound for graph re-induction. Residual completion adds at most time, provided that each of the relations produces a bounded number of candidate residual directions and each direction scans at most existing features for reuse matching. Combining these costs gives an time upper bound per global update.
The bound for acyclicity checking is conservative, as it assumes that every candidate requires traversing the entire graph. In practice, candidate filtering can reduce the number of checks, traversal visits only the relevant reachable region, and a check can terminate early once a cycle-forming path is found. These mechanisms help reduce the actual overhead of global updates and contribute to manageable practical costs. If a run contains training updates and global updates, with initial and final graph induction included as applicable, the total time upper bound is
We provide the detailed space complexity analysis in the next comment.
Response to Reviewer G8HL — Q1
We greatly appreciate your positive assessment of our method and experiments. Thank you for your thoughtful review and for the helpful questions and suggestions. We hope the additional results and analyses below address your questions and further strengthen our paper.
Q1: In the main experiment with 30,720 features, how much additional training time and GPU memory does AG-SAE require compared with Vanilla SAE?
A: Thank you for this important question. To address this concern, we report the computational resource usage of Vanilla SAE and AG-SAE in the main experimental setting with 30,720 features in Table 1. Both methods were run on a single NVIDIA A100-SXM4-80GB GPU.
Table 1. Computational resource usage in the main experimental setting.
AG-SAE does require more computational resources than Vanilla SAE, primarily in longer runtime and higher peak allocated GPU memory. However, the complete run takes 15.66 hours on a single GPU, with 8.2149 GiB of peak allocated GPU memory, indicating that its computational cost remains manageable and the method is practical at the evaluated scale.
Moreover, for a fixed model and layer, the learned dictionary and graph can be reused for subsequent feature organization and model analysis without repeating AG-SAE training or graph induction. This allows the additional offline computation to support multiple downstream analyses, providing continued value from the learned features and their relations.