acceptodds
Under review as a conference paper at ICLR 2027

When Trees Are Not Enough: Learning Mixed-Topology Feature Graphs with Adaptive Graph Sparse Autoencoders

10 views

Abstract

Sparse autoencoders (SAEs) expose interpretable features in large language model activations, yet existing structured SAEs impose single-parent trees or forests, while post-hoc graphs permit multiple parents but neither guide feature learning nor ensure reliable relation recovery. We introduce the Adaptive Graph Sparse Autoencoder (AG-SAE), a structure-guided training paradigm that treats each feature's complete parent set as an atomic structural hypothesis and lets evidence select zero, one, or multiple parents. By competing complete parent sets against null, subset, and alternative explanations, AG-SAE identifies jointly necessary multi-parent relations while rejecting redundant or spurious alternatives and verifying that each child contributes beyond its parents. The induced topology over SAE features then defines a differentiable structural loss that guides SAE training, while topology-guided refinement mitigates feature absorption and uses persistent reconstruction gaps exposed by the learned structure to initialize new features. The entire graph is then induced again from the revised dictionary by reassessing every feature's complete parent set, closing the dictionary–graph self-consistency cycle. Experiments demonstrate exact mixed-topology recovery in a controlled toy model, greater relational reliability and semantic validity than structured and post-hoc baselines on real LLM activations, and stronger feature-level causal interventions than conventional SAE features. AG-SAE thereby turns recovered mixed-topology feature structure into an unsupervised training signal that improves the dictionary, enables reliable feature organization beyond the topological limitations of trees, and exhibits stronger causal control beyond reconstruction.

Primary area:
interpretability and explainable AI
Keywords:
sparse autoencoders, mechanistic interpretability, dictionary learning, graph structure learning, feature hierarchies, language model interpretability

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_G8HLOpenReview ↗
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
  1. In the main experiment with 30,720 features, how much additional training time and GPU memory does AG-SAE require compared with Vanilla SAE?

  2. 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?

  3. 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
AuthorsOpenReview ↗

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.

MetricVanilla SAEAG-SAE
Total runtime (h)8.8215.66
Peak allocated GPU memory (GiB)3.28058.2149
Matrix FLOPs per training step1.7395 × 10¹²2.0294 × 10¹²
Final output size (MiB)540.17548.72

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.

AuthorsOpenReview ↗

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.

AuthorsOpenReview ↗

Response to Reviewer G8HL — Q2 (Part 2/3)

Let MM denote the maximum total number of features during a run, dd the activation dimension, BB the batch size, and kk the maximum parent-set size. The following analysis focuses on scaling with MM, holding dd, BB, the dataset size and sampling budgets, the candidate-pool size, kk, 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 O(BdM)O(BdM) computation per step under the standard SAE encoder–decoder formulation (Gao et al., 2025), scaling linearly with MM for fixed BB and dd. 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 (121)+(122)+(123)=298\binom{12}{1}+\binom{12}{2}+\binom{12}{3}=298 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 MM. Consequently, computing the structural loss over all features requires O(M)O(M) 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 kk parents, requiring O(kMd)O(kMd) operations over all features, which is also linear in MM for fixed kk and dd. Thus, the combined cost of dense encoding and reconstruction, the structural loss, and graph-conditioned decoding remains linear in MM. Including BatchTopK selection, a conservative sorting-based bound is O(Mlog⁡M)O(M\log M) for fixed BB. Combining these terms therefore gives an O(Mlog⁡M)O(M\log M) 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 M(M−1)M(M-1) feature pairs. With a fixed per-pair scoring cost, this requires O(M2)O(M^2) 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 O(M)O(M), and their fitting and evaluation require O(M)O(M) time under fixed sampling and numerical fitting budgets. Acyclic graph selection then sorts these scored candidate parent sets in O(Mlog⁡M)O(M\log M) time and performs reachability checks in order, checking each candidate at most once. Since the graph contains at most MM nodes and kMkM edges, each check requires at most O(M+kM)O(M+kM) time. Combined with O(M)O(M) candidates and fixed kk, all checks have an O(M2)O(M^2) time upper bound. Thus, candidate retrieval, parent-set evaluation, and acyclic graph selection together give an O(M2)O(M^2) upper bound for graph re-induction. Residual completion adds at most O(M2)O(M^2) time, provided that each of the O(M)O(M) relations produces a bounded number of candidate residual directions and each direction scans at most MM existing features for reuse matching. Combining these costs gives an O(M2)O(M^2) 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 UU training updates and RR global updates, with initial and final graph induction included as applicable, the total time upper bound is

O!(UMlog⁡M+RM2).O\\!\left(UM\log M+RM^2\right).

We provide the detailed space complexity analysis in the next comment.

Reviewer_UUe6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Adaptive Graph Sparse Autoencoder (AG-SAE) to learn mixed-topology feature graphs from LLM activations. AG-SAE identifies complete parent sets and iteratively refines feature representations using graph-guided training, absorption realignment, and residual completion. Experiments show improved structure recovery, relational reliability, semantic validity, and causal intervention effectiveness.

Critical strengths
  1. Complete parent-set induction enables multi-parent structure learning, while iterative dictionary–graph refinement integrates structural information into SAE training.
  2. The paper evaluates synthetic structure recovery, real-LLM relational reliability, semantic validity, causal interventions, and ablations, demonstrating consistent improvements over baselines.
Critical weaknesses

The individual contributions of structural loss, residual completion, and full re-induction remain unclear. Experiments are limited to relatively small LLMs and parent sets of at most three, without systematic runtime, memory, or scaling analysis.

Questions for authors

Please refer to the weaknesses above.

7 comments
AuthorsOpenReview ↗

Response to Reviewer UUe6 — Q1 (Part 1/2)

We sincerely appreciate your thoughtful review and your recognition of our work, as well as your valuable suggestions for strengthening our evaluation. We hope the additional experiments and analyses below address your concerns and provide further support for our approach.

Q1: The individual contributions of structural loss, residual completion, and full re-induction remain unclear.

A: We sincerely thank the reviewer for raising this important question. To address this concern and clarify the contribution of each component, we supplement the ablations already presented in the paper with additional end-to-end training ablations of the three components. All ablations use the same training configuration and budget as the main Gemma-2-2B experiment, with the corresponding component disabled in each setting and the dictionary size fixed at 30,720. The resulting graph structure counts, dictionary quality, and relation validity are reported below.

Table 1. Counts of recovered graph structures under different ablation settings. Reductions in multi-parent nodes are calculated relative to the full AG-SAE.

SettingTotal nodesSingle-parent nodesMulti-parent nodesNodes with parentsReduction in multi-parent nodes
Full AG-SAE30,7205,4211,6707,091—
w/o Structural Loss30,7205,2461,3996,64516.23%
w/o Residual Completion30,7202,8911,0033,89439.94%
w/o Full Re-induction30,7202,3278283,15550.42%

Table 2. Dictionary quality and relation validity under different ablation settings. PSV, NR, SV, and Joint are reported as percentages. Joint denotes the overall rate of simultaneously passing all three checks.

SettingL0L_0EV ↑MSE ↓Single-parent PSV ↑Single-parent NR ↑Single-parent SV ↑Multi-parent PSV ↑Multi-parent NR ↑Multi-parent SV ↑Joint ↑
Full AG-SAE50.130.6700.33099.7597.0076.8897.0091.8875.0069.56
w/o Structural Loss50.420.6370.36396.6396.8876.5084.1384.2577.2561.25
w/o Residual Completion50.340.6130.38799.0097.2576.0096.1391.7577.6369.69
w/o Full Re-induction50.560.6380.36283.6384.6360.6366.5080.1362.2554.88

We provide a detailed analysis of the three components in the follow-up comment below.

AuthorsOpenReview ↗

Response to Reviewer UUe6 — Q1 (Part 2/2)

Below, we analyze the ablation results reported in Tables 1 and 2 to clarify the individual contributions of the three components.

Structural Loss. Removing the structural loss reduces the number of features with parents by 6.29% and the number of multi-parent features by 16.23%, indicating that structural training signals help recover a richer set of feature relations. The decline in relation reliability is particularly evident in multi-parent PSV and NR. PSV decreases from 97.00% to 84.13%, and NR from 91.88% to 84.25%, corresponding to drops of approximately 12.9 and 7.6 percentage points, respectively. Overall Joint also decreases from 69.56% to 61.25%, a drop of approximately 8.3 percentage points. These results support the role of the structural loss in incorporating recovered relations into feature learning as training signals, helping the dictionary develop more reliable parent–child structures during optimization, particularly for complex relations that require joint support from multiple parents.

Residual Completion. Removing Residual Completion reduces the number of features with parents from 7,091 to 3,894, a decrease of 45.09%, and the number of multi-parent features by 39.94%. This reduction is accompanied by a decline in dictionary quality. At comparable sparsity, EV decreases from 0.670 to 0.613, an absolute drop of 0.057 and a relative decrease of 8.51%, while MSE increases from 0.330 to 0.387. Furthermore, the dead feature rate rises sharply from only 0.01% in the full AG-SAE to 35.24% without RC. Meanwhile, overall Joint among the remaining relations increases slightly from 69.56% to 69.69%, by approximately 0.13 percentage points. Thus, although removing Residual Completion does not reduce the joint pass rate of the remaining relations, it substantially reduces the number of recovered structures, dictionary reconstruction quality, and capacity utilization. These results support Residual Completion’s primary contribution to improving the effective use of dictionary capacity, reconstruction quality, and structural coverage during AG-SAE’s joint optimization of the graph–dictionary pair. Through residual-driven feature reuse and initialization, Residual Completion addresses representational gaps not yet captured by the current dictionary, bringing underutilized dictionary capacity into subsequent learning and supporting the recovery of more meaningful mixed-topology structures.

Full Re-induction. Removing full re-induction substantially reduces both the number and quality of the mixed-topology relations learned by AG-SAE. The number of features with parents decreases from 7,091 to 3,155, a reduction of 55.51%, and the number of multi-parent features decreases from 1,670 to 828, a reduction of 50.42%. PSV, NR, and SV also decline for both single-parent and multi-parent relations. Specifically, single-parent PSV decreases from 99.75% to 83.63%, and multi-parent PSV from 97.00% to 66.50%; overall Joint decreases from 69.56% to 54.88%, a drop of approximately 14.7 percentage points. As dictionary representations evolve during AG-SAE training, the existing graph can gradually become misaligned with the current feature relations. By reassessing and updating parent sets, full re-induction translates changes in dictionary representations into updated structural guidance, which is then fed back into subsequent feature learning to sustain the joint optimization of the graph and dictionary. These results indicate that continuing dictionary optimization alone is insufficient to attain the structural coverage and relation validity of the full method. Full re-induction provides a key link between dictionary updates and structural optimization, enabling updated representations to support the discovery of new structures and continually refreshing the structural guidance used to learn richer and more reliable feature relations.

AuthorsOpenReview ↗

Response to Reviewer UUe6 — Q2

Q2: Experiments are limited to relatively small LLMs.

A: We sincerely thank the reviewer for raising this important question. To further evaluate AG-SAE’s applicability to larger models, we have conducted additional experiments on Qwen3.5-9B-Base. These experiments follow the same experimental setup as our Qwen3.5-2B-Base experiments in Appendix D.3.2, using model activations from MiniPile for training. We directly reuse the AG-SAE hyperparameters from those experiments, without any additional tuning for the 9B model, and retain the same dictionary bank widths of 16,384 and 32,768. The recovered graph structures, relation validity, and dictionary quality are reported below.

Table 3. Graph structure statistics on Qwen3.5-9B-Base.

MetricValue
Total feature nodes49,152
Single-parent nodes6,089
Multi-parent nodes2,674
Total parented nodes8,763
Multi-parent share among parented nodes30.51%

Table 4. Feature relation validity on Qwen3.5-9B-Base. All metrics are reported as percentages.

Relation typePSV ↑NR ↑SV ↑Joint ↑
Single-parent99.3994.6766.0963.19
Multi-parent95.4293.9368.2964.57

Table 5. Reconstruction quality and sparsity on Qwen3.5-9B-Base.

Dictionary widthEV ↑L0L_0
16,3840.6750.56
32,7680.7150.65

On the 9B model, AG-SAE recovers 2,674 multi-parent features, accounting for 30.51% of all parented features, indicating that it continues to recover rich mixed-topology structures when applied to a larger model. Across all relations with sufficient activation samples for held-out evaluation, both single-parent and multi-parent relations achieve high PSV and NR, with Joint validity of 63.19% and 64.57%, respectively, supporting the validity of the recovered relations. Meanwhile, the two dictionaries achieve EV values of 0.67 and 0.71, respectively, while keeping L0L_0 close to the target of 50, indicating good reconstruction quality at the intended sparsity level.

These additional results extend our evaluation of AG-SAE from 2B to 9B, supporting its ability to learn effective dictionary–graph pairs on the evaluated larger model. Achieving these results by directly reusing the 2B experimental hyperparameters, without additional tuning for the 9B model, further supports AG-SAE’s generalization across model scales.

Reviewer_JLDrOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper focuses on the interpretation of LLMs using sparse autoencoders (SAEs). Previous structured SAE-based methods use single-parent trees or forests but the single-parent formulation is not enough for complex real-world concepts. The paper proposes a new method, AG-SAE, which considers full-set parents of each feature for training. The proposed AG-SAE utilizes a recursive way to update the dictionary-graph pair. The experimental results show that AG-SAE outperforms structured and post-hoc baselines.

Critical strengths

The paper clearly presents the current research progress and gaps of SAEs for LLMs. To address the single-parent issues of SAEs, the paper proposes a new adaptive method, AG-SAE, to iteratively identify the multi-parent relations and SAE training. The overall framework is well formulated in the paper. The performance of AG-SAE is validated in both toy and real LLM settings.

Critical weaknesses

The paper mainly conducts experiments on the Gemma-2-2B model, which is a small and legacy model. The generalization of AG-SAE needs to be further verified on newer LLMs.

Questions for authors

Is there any chance to use a more general form (i.e., a knowledge graph) to present the relations between concepts instead of multi-parent relations, and why?

2 comments
AuthorsOpenReview ↗

Response to Reviewer JLDr — Q1

We sincerely thank you for your recognition of our work and for your valuable comments and constructive suggestions. We hope the additional results and discussion below address your concerns and further substantiate our contributions.

Q1: The paper mainly conducts experiments on the Gemma-2-2B model, which is a small and legacy model. The generalization of AG-SAE needs to be further verified on newer LLMs.

A: Thank you for this helpful suggestion. Beyond Gemma-2-2B, our submission already includes experiments on Qwen3.5-2B-Base, released in 2026, in Appendix D.3.2. These results show that AG-SAE can also learn meaningful mixed-topology feature graphs on a newer model architecture.

To further evaluate the generalizability of AG-SAE, we have conducted additional experiments on Qwen3.5-9B-Base (Qwen Team, 2026), a larger model also released in 2026, using activations from MiniPile. We directly reused the AG-SAE hyperparameters from our Qwen3.5-2B-Base experiments, without any additional tuning for the 9B model, and retained the same dictionary bank widths of 16,384 and 32,768. The results are summarized in Tables 1–3.

Table 1. Graph structure statistics on Qwen3.5-9B-Base.

MetricValue
Total feature nodes49,152
Single-parent nodes6,089
Multi-parent nodes2,674
Total parented nodes8,763
Multi-parent share among parented nodes30.51%

Table 2. Feature relation validity on Qwen3.5-9B-Base.

Relation typePSV (%) ↑NR (%) ↑SV (%) ↑Joint (%) ↑
Single-parent99.3994.6766.0963.19
Multi-parent95.4293.9368.2964.57

Table 3. Reconstruction quality and sparsity on Qwen3.5-9B-Base.

Dictionary widthEV ↑L0L_0
16,3840.6750.56
32,7680.7150.65

The recovered graph contains 2,674 multi-parent features, accounting for 30.51% of all parented features, showing that AG-SAE continues to recover substantial mixed-topology structure when scaling to a larger model on the same MiniPile dataset. Across all relations with sufficient activation samples for held-out evaluation, both relation types achieve high PSV and NR, with Joint validity exceeding 63%, supporting the reliability and semantic meaningfulness of the recovered relations. Meanwhile, the dictionaries achieve EV values of 0.67 and 0.71 while keeping L0L_0 close to the target of 50, indicating good reconstruction quality and controlled sparsity.

These results support AG-SAE’s ability to learn high-quality dictionary–graph pairs on newer and larger models. Achieving this without additional hyperparameter tuning further supports the generalizability of our approach.

AuthorsOpenReview ↗

Response to Reviewer JLDr — Q2

Q2: Is there any chance to use a more general form (i.e., a knowledge graph) to present the relations between concepts instead of multi-parent relations, and why?

A: Thank you for this insightful suggestion. We consider using a more general knowledge graph to organize feature relations a valuable direction for extending AG-SAE. To explain our current choice and its connection to this potential extension, we first distinguish the different levels of abstraction associated with knowledge graphs and multi-parent relations. A knowledge graph is a knowledge representation framework for organizing entities, concepts, and their semantic relations, with a stronger emphasis on the application level, whereas a multi-parent structure is a topological form describing how nodes are connected. The former concerns what knowledge is expressed by relations between concepts; the latter concerns how a node establishes structural connections with multiple parent nodes.

For AG-SAE, the central problem addressed in this paper is that existing structured sparse autoencoders restrict each child feature to at most one parent feature, preventing them from fully expressing relations in which multiple parent features jointly support a child feature. This restriction directly concerns which connections are permitted between nodes and therefore lies at the level of topological structure. To address this problem, we adopt an organizational form that permits multiple parents and treats complete parent sets as the basic units of identification and validation. This allows multiple parent features to be assessed as a whole, overcoming the single-parent constraint and recovering reliable mixed-topology feature relations. A more general knowledge graph framework can accommodate this structure, but its capabilities for organizing diverse semantic relations and supporting downstream applications are not necessary for addressing the specific topological problem studied here. We therefore focus first on multi-parent relations because they provide a direct solution at the structural level corresponding to our core research problem.

However, your insightful suggestion points to a valuable extension for AG-SAE. When a research setting requires further organization and use of semantic knowledge among concepts for domain analysis or specific applications, a representation centered on a knowledge graph would be more suitable. For example, in a medical setting, we may wish to use disease–symptom associations and drug–disease treatment relations to query and analyze the medical knowledge represented by a model. This requires explicitly distinguishing the semantics of different relations and organizing the relevant concepts into a usable knowledge structure. In such settings, AG-SAE could be extended into a learning framework centered on a feature knowledge graph by establishing correspondences between features and domain concepts, using domain knowledge to propose candidate relations, developing appropriate model-internal validation criteria for different relation types, and exploring suitable structured training objectives. Such an extension could connect semantic knowledge at the concept level with representational evidence at the feature level, enabling the recovered feature structures to further support the organization, analysis, and application of domain knowledge.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Reviewer Feedback and Key Rebuttal Evidence

Dear Area Chair,

Thank you for overseeing the review of our submission. Since the rebuttal period ends soon and we haven’t heard back from Reviewer JLDr, we would like to briefly summarize the reviewer feedback and highlight the key evidence we added to address the questions raised in their review.

All three reviewers recognized our paper’s contribution and gave it an initial rating of 3 (weak acceptance). We provided detailed responses to all three reviews. In their follow-up comments, Reviewers UUe6 and G8HL both confirmed that we had addressed all of their concerns and that they would maintain their positive ratings in support of the paper.

Reviewer JLDr also expressed support for our paper in their initial review and gave it a rating of 3, although we haven’t yet received a follow-up from them. Below, we briefly summarize how we addressed the questions raised by Reviewer JLDr.

1. Generalization to newer and larger models. JLDr noted that our main experiments focused on Gemma-2-2B and asked whether AG-SAE would generalize to newer models. We clarified that the original submission already included results on Qwen3.5-2B-Base. During the rebuttal, we extended that evaluation to Qwen3.5-9B-Base using the same hyperparameters. AG-SAE recovered 2,674 multi-parent features, accounting for 30.51% of all features with at least one parent. Of the multi-parent relations evaluated on held-out data, 64.57% passed all three validity checks. These results show that AG-SAE can recover mixed-topology feature graphs on a newer and larger model.

2. Choice of representation for feature relations. JLDr asked whether a more general representation, such as a knowledge graph, could describe the relations between concepts. We explained that complete parent sets capture which features jointly support a child feature. This directly addresses the structural question studied in the paper and provides a foundation for a knowledge graph with typed semantic relations.

We appreciate the reviewers’ feedback and thank you for considering these results in your assessment.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.