acceptodds
Under review as a conference paper at ICLR 2027

From Atoms to Trees: Building a Structured Feature Forest with Hierarchical SAEs

Abstract

Sparse autoencoders (SAEs) have proven effective for extracting monosemantic features from large language models (LLMs), yet these features are typically identified in isolation. Feature splitting and cross-feature organization suggest that SAE dictionaries may admit useful coarse-to-fine structure, although splitting alone does not establish a semantic hierarchy. To study such structure, we propose the **Hierarchical Sparse Autoencoder (HSAE)**, which jointly learns a series of SAEs and the parent-child relationships between their features. HSAE strengthens the alignment between parent and child features through two novel mechanisms: a structural constraint loss and a random perturbation of eligible non-root features. At matched , HSAE retains nearly the same variance explained as JumpReLU while achieving the best RAVEL, SCR, and sparse-probing scores among the evaluated methods, including clear gains over the framework-controlled H-SAE and Tree SAE reimplementations. Across layers and models, HSAE improves cross-level activation alignment over post-hoc hierarchies built from independently trained SAEs, while case studies and large-scale hierarchy judgments show coarse-to-fine semantic distinctions. Finally, we demonstrate two applications of the learned hierarchy: reusing parent descriptions for feature interpretation and steering broad concepts with coarse-level features.

Primary area:
interpretability and explainable AI
Keywords:
Large Language Model, Mechanistic Interpretability, Sparse AutoEncoder

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_8YArOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes HSAE, which jointly trains 4 JumpReLU SAEs (2K to 16K latents) and links features of adjacent levels into a partial forest. It has 3 parts: (1) an L2 loss L_PC that asks each parent's decoder contribution to equal the sum of its children's contributions, (2) a random perturbation that replaces some non-root parents with their children during reconstruction, and (3) re-assignment of parents by encoder cosine every 5000 steps, with the bottom 20% of children left unassigned. On Gemma-2-2B layer 13 at L0 ~50, the 16K level keeps the variance explained of JumpReLU (0.710 vs 0.712) and gets the best scores on RAVEL, SCR and sparse probing. HSAE also has lower parent-child Hamming distance than post-hoc hierarchies (21.3 vs 36.1). An LLM judge rates 70.5% of 5000 sampled edges as "parent subsumes child", vs 60.5% for JumpReLU post-hoc and 67.1% for Tree SAE.

Critical strengths
  1. The method is simple and each level remains a usable SAE. At matched L0 reconstruction does not drop, so the gains on RAVEL, SCR and sparse probing over framework-controlled baselines are useful if they hold across seeds. Also, the activation metrics show the same pattern on 3 layers of Gemma-2-2B, 1 layer of Qwen3-4B and 3 L0 values.

  2. The paper is careful about what an edge means. It separates functional substitution, activation containment and semantic subsumption, it does not overclaim. Besides, the LLM judge is checked against human labels (kappa = 0.72).

Critical weaknesses
  1. The evidence for a semantic hierarchy is weak. Hamming distance and P(P|C) closely track the loss, and Sec. 4.2 itself says they are not an independent test of a semantic hierarchy. L_PC directly penalizes a parent firing without children and a child firing without its parent, so these metrics should improve even on data with no hierarchy. Also, with 1 active child, L_PC is 0 only when d_child points in the same direction as d_parent. So the loss pushes children to copy the parent direction, not to specialize it (Park et al., ICLR 2025). The main support is then the LLM judge, but it has no reversed-order test and no random-pair baseline. Since child top examples likely overlap with the parent's, a "yes" may only reflect this overlap.

  2. It is unclear where the SAEBench gains come from. These gains are the strongest result of the paper. But the ablations report only Hamming distance, plus variance explained in Fig. 5, and none of them reports RAVEL, SCR or sparse probing. Also, no result reports seeds or error bars, while the gaps over JumpReLU are small (0.02-0.04). So we cannot tell if the gains come from L_PC, the perturbation, the reassignment, or run-to-run noise. In addition, the paper should explain more clearly what it adds over Tree SAE, which also combines a reconstruction condition with dynamic parent assignment. I treat Tree SAE as concurrent work, so this is a clarity point, not a novelty penalty.

Questions for authors
  1. What is the parent-child decoder cosine distribution for HSAE vs post-hoc hierarchies? How many edges are near copies (cos > 0.9)?
  2. What is the judge's "yes" rate when parent and child are swapped? What is the rate for random same-level pairs and for sibling pairs?
  3. Can you report Table 1 over 3+ seeds, and also RAVEL, SCR and sparse probing for the ablations (constraint only, perturbation only, fixed tree)?
  4. How many parents have children in each method? Does the ranking change if Hamming distance is normalized by this number, or if the post-hoc baseline assigns parents by co-activation?

Clarification:

  1. Beyond the loss form and the 20% filter, what is the main difference from Tree SAE?
1 comment

We thank the reviewer for their detailed review and concrete suggestions for strengthening the semantic evaluation and comparisons with related methods.

Q1

We provide the distribution of parent–child decoder cosine similarities in Appendix E.2 (Figure 12) of the revised version. For HSAE, 48.7% of pairs have a cosine similarity greater than 0.9. For JumpReLU SAE, this rate is 19.2%. However, these cosine > 0.9 pairs does not necessarily mean that they are near copies. For these pairs, the mean activation-support Jaccard similarity is only 0.6, and the child activates with a probability of approximately 62% when its parent is active. This is more consistent with the child being a more selective version of its parent feature than an identical copy.

W1/Q2

We provide additional results for reversed-order, random same-level, and sibling pairs in Appendix C.2 of the revised version. Random pairs (0.4%) and sibling pairs (23.2%) have much lower acceptance rates than pairs in the assigned direction (71.6%). The reversed-order acceptance rate remains relatively high (42.7%). We suspect that this partly reflects the limited set of 15 examples used for each judgment: the sampled parent and child examples may share substantial overlap.

Q3

We have updated Table 1 in the revised version to include the mean results over 4 training seeds and their standard deviations. HSAE gains a clear advantage in RAVEL (0.661 vs 0.636; HSAE SD: 0.005) and Sparse Probing (0.893 vs 0.874; HSAE SD: 0.004), while maintaining the lead in mean SCR (0.339 vs 0.332; HSAE SD: 0.013).

We have also added Table 3 in Appendix B.2 of the revised version to report the SAEBench results of the ablation studies. The explicit constraint loss is the most effective component of HSAE. This is consistent with the improvement in parent–child relationships.

Q4

For HSAE, 99.99% of features outside the final level have at least one child feature. The corresponding proportion for JumpReLU SAE is 99.6%. Sparsity encourages the coarse SAE to allocate its limited feature budget to useful directions, which are therefore likely to remain represented, possibly through more specialized features, in the larger dictionary.

The ranking remains unchanged when we normalize the Hamming distance: HSAE's normalized distance is still approximately 40% lower than that of JumpReLU SAE + post-hoc. Replacing the post-hoc assignment rule with co-activation does not improve the Hamming distance; instead, it makes it slightly worse (36.1 → 36.4). This is similar to the HSAE hierarchy-update ablation, in which different assignment methods yield similar hierarchy quality.

Clarification: Difference from Tree SAE

We appreciate the reviewer's recognition of Tree SAE as concurrent work. HSAE was developed slightly earlier within the same period, and we regard the two approaches as contemporaneous contributions.

Both methods allow the hierarchy to change during training, but HSAE revises the hierarchy more actively. Tree SAE's dynamic allocation mechanism only reallocates dead child features according to parent capacities. In contrast, HSAE periodically recomputes assignments across adjacent levels, allowing active as well as inactive features to change parents as their representations evolve. Thus, HSAE treats the hierarchy itself as an object of continued optimization, rather than limiting reassignment to dead features. This broader rematching mechanism is a core distinction beyond the loss formulation and the orphan filter.

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

This work presents a new Sparse Autoencoder architecture that aims at giving hierachical structure to features found, such that 'natural' interpretrations are easier; features are grouped by classes and relationships between features mean something about the underlying structure. They show that this new architecture has about the same or better performance on some downstream tasks as well as reconstruction

Critical strengths

The new hierarchical architecture matches standard architectures which have harder to interpret relationships between features.

Having a dynamically assigned hierarchy is more natural than a fixed one.

The hierarchical validation tests, while not perfectly convincing, show that the recovered hierarchy from their method is larger that other methods, even when compared with methods that have explicit hierarchies.

The hierarchy based automated interpretability results are interesting and show that the children are correctly matched with parents.

Critical weaknesses

Assigning parents through children via cosine similarity seems like wasted capacity. It is unclear to me if the 'functional substitution' is the most natural way to do hierarchy, but I guess the results speak for themselves.

The authors fail to perform experiments with models with know hierarchical representations, for instance models trained on syntetic data generated by Probabilistic context-free grammar, so the hierarchy extraction experiments are less convincing. The post hoc analysis is not directly fair to the other methods which are not optized to have children parent structures, and we don't know if the underlying data should even suport children parent structures.

As the authors point out, the 'perfect' SAE architecture probably would have both hierarchical and non hierarchical features, or at least dynamic levels of hierarchy.

Questions for authors

Have you investigated the stability of the learned hierarchy? Do you find your architecture to find the same trees, or do trees depend on the training condition?

Your training loss seems to force children feature to be 'subclasses' of the parent feature, instead of 'specializing' to broader concepts (basically you require 'poodle' + 'labrador'+ etc. = 'dog', instead of 'dog' as a parent + 'labrador definiting features' as one children). What is the motivation behind this?

How does your iterative hierarchy optimization compare to having a fixed hierarchy under your loss? What about starting with a fixed hierarchy and updating it?

Why the 80-20 split of orphan features? Is there a reason that this threshold is not adaptive? Do we need that most features are part of a hierarchy?

2 comments

We thank the reviewer for their encouraging assessment and insightful questions about hierarchy stability, the motivation for our design, and the structural assumptions of HSAE.

W2

We have added a synthetic example with a known generating hierarchy in Appendix E.1 of the revised version. In this controlled setting, JumpReLU SAE can learn an alternative representation in which a middle-level decoder bias absorbs a branch mean, leaving that branch without an active parent feature. Reassigning edges alone cannot repair this missing representation. HSAE's parent–child constraint penalizes the unmatched contributions and encourages recovery of the generating hierarchy. Under the feature-reconstruction and hierarchy-recovery criteria described in the appendix, HSAE succeeds in 13 out of 16 initialization seeds, compared with 4 out of 16 for JumpReLU SAE with post-hoc matching.

Q1

We have empirical observations of hierarchy stability. We have not quantified this stability because matching and comparing different trees is known to be difficult, especially at a large scale. However, we observe hierarchies related to certain concepts emerging consistently across all training seeds. For example, the temporal concept tree shown in Figure 6 appears in all the HSAEs we trained, although its exact structure varies.

Q2

We believe that certain features are meaningful only in the context of their parent features. For example, round or square is meaningful only when the shape feature is activated. We therefore choose this hierarchical organization of features.

Building on this idea, we want subclasses to be interpretable on their own, rather than representing only increments relative to their parents. For example, what would "square" minus "shape" mean? Such a subtraction does not have a clear interpretation. Squares form a subset of shapes, rather than an additional modification of the shape concept.

Nevertheless, features at each level retain the original SAE decomposition, so independent features can still be represented.

Q3

Our ablation study in Figure 3(b) demonstrates that keeping the hierarchy fixed throughout training can impair HSAE's ability to capture feature relationships. Starting from a fixed hierarchy and then updating it would constitute an alternative structural initialization. We expect that, given sufficient training, it would achieve results similar to those of the current method.

Q4

The 80–20 split was tuned on the validation set to obtain the best performance. An adaptive threshold could improve the results in principle, and we will explore this in future work. We aim to include most features in the hierarchy because HSAE is designed to capture relationships between features. Leaving too many features unassigned would weaken this effect.

I thank the authors for their time answering my questions. I think it would be hard for me to raise my score more as I think the work is interesting but is maybe slightly incremental. I will wait for other reviewers updates before changing my score.

Reviewer_hjC9OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Sparse autoencoders (SAEs) extract monosemantic features from LLM representations, and SAEs of different sizes tend to recover features at different granularities, which suggests a coarse-to-fine organization of the feature space. This paper proposes the Hierarchical Sparse Autoencoder (HSAE), which jointly trains a series of SAEs of increasing dictionary size together with parent–child assignments between their features. Training combines a structural constraint loss, a random parent–children feature perturbation, and alternating optimization between the model parameters and the hierarchy. The paper evaluates SAE quality, cross-level alignment, interpretability, and presents two preliminary applications.

Critical strengths
  • Learning coarse-to-fine structure over SAE features is an interesting problem, with practical value for navigating large dictionaries.
  • The paper is careful about its claims, and it states its limitations openly.
Critical weaknesses
  • The gains on SAEBench are not clear and reported without variance.
  • The evidence that the learned hierarchy is semantic is weak. The alignment metrics largely measure what the training objective optimizes.
  • The practical value of the hierarchy is not yet established, since proposed applications are preliminary.
Questions for authors
  • Could the authors report standard deviations or error bars, to show whether HSAE's gains over the baselines are significant? Could the SAEBench results also be reported at several L0 values?
  • Sec. 4.2 mainly tests the training objective, and the authors acknowledge that it does not show that the learned edges recover a pre-existing semantic hierarchy. Prior work [1] uses a toy model with a known ground-truth hierarchy to test this. Could the authors run a similar experiment, for example by checking whether HSAE recovers the true parent–child edges and their direction?
  • Sec. 4.3 states that sampling is not matched across methods by hierarchy level, activation frequency, branch size, or filtering status. Could the authors re-run the comparison with matched sampling? Could they also report the judge's rate on randomly paired features to show how much of the results reflects the learned assignment rather than general feature similarity?
  • For Sec. 5.1, does the description of an unrelated feature yield a similar gain? For Sec. 5.2, could the steering results be evaluated on generated text?

[1] Bussmann, Bart, et al. "Learning multi-level features with matryoshka sparse autoencoders." arXiv preprint arXiv:2503.17547 (2025).

1 comment

We thank the reviewer for their thoughtful review and constructive suggestions on the reliability of our results.

W1/Q1

We have updated Table 1 in the revised version to include results from training runs with 4 seeds and to report the standard deviations. Compared with the second-best method for each metric, HSAE gains a clear advantage in RAVEL (0.661 vs 0.636; HSAE SD: 0.005) and Sparse Probing (0.893 vs 0.874; HSAE SD: 0.004), while maintaining the lead in mean SCR (0.339 vs 0.332; HSAE SD: 0.013). We believe this is clear evidence of the gain achieved by our method.

We have also added Table 11 to report the SAEBench results for HSAE and JumpReLU SAE at different observed L0 values.

W2/Q2

We have added a synthetic example with a known generating hierarchy in Appendix E.1 of the revised version. In this controlled setting, JumpReLU SAE can learn an alternative representation in which a middle-level decoder bias absorbs a branch mean, leaving that branch without an active parent feature. Reassigning edges alone cannot repair this missing representation. HSAE's parent–child constraint penalizes the unmatched contributions and encourages recovery of the generating hierarchy. Under the feature-reconstruction and hierarchy-recovery criteria described in the appendix, HSAE succeeds in 13 out of 16 initialization seeds, compared with 4 out of 16 for JumpReLU SAE with post-hoc matching.

Q3

We have conducted additional experiments on our hierarchy judgments and report them in Sections C.1 and C.2 of Appendix C. The results remain consistent across sampling methods. We also report the standard deviations associated with pair sampling, which further support HSAE's advantage over the baseline. We originally used uniform sampling because exact matching across different architectures is difficult; for example, H-SAE has only two levels.

For randomly paired features, the subsumption rate is only 0.4%, much lower than the 71.6% rate for parent–child pairs identified by HSAE.

Q4

In Sec. 5.1, providing the description of an unrelated feature has no effect: 15 child examples are still required to reach the same AutoInterp score. For Sec. 5.2, evaluation on generated text is possible, but quantifying the results would be more difficult because it would require evaluating natural-language outputs. We therefore use the normalized target-token logit increase as a continuous metric.

Comments on the paper

1 comment

We apologize for the delay in responding. We needed additional time to complete the experiments and analyses requested by the reviewers. In the revised version, we added 4-seed SAEBench quality results with standard deviations, SAEBench evaluations for different sparsity and ablations, ablation studies for the hierarchy judgement method, and many other results requested by the reviewers. We thank the reviewers for their constructive suggestions, which motivated these additional experiments and analyses and made the paper more rigorous and empirically grounded. We would welcome any further feedback or questions on our responses and the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.