acceptodds
Under review as a conference paper at ICLR 2027

Beyond Decoder Proxies: Causal Feature Selection for Sparse Autoencoder Control

Abstract

Scaling sparse autoencoder (SAE) dictionaries creates a feature-selection bottleneck: static decoder projections can screen thousands of candidate directions, but do not reveal which cause a desired probability shift in context. We introduce the target-selection margin (TSM), an intervention-based selection method that ranks features by changes in the log ratio of target to distractor probability mass. Across 18 concepts and five public model/SAE settings, TSM correlates more strongly with held-out margin than decoder proxies (Spearman - versus -). In matched 512-feature pools, paired held-out utility gains have concept-clustered 95% confidence intervals entirely above zero in all five settings. Lexical and grammatical audits broaden coverage beyond the original token sets, while severe target replacement exposes specification sensitivity. Two-dictionary audits identify retrieval losses before causal reranking. Separate behavioral audits show that improved contrast-specific decisions need not overcome full-vocabulary competition and do not establish reliable multi-step generation gains. These results support causal feature selection across prompts under a specified target–distractor contrast, distinct from retrieval coverage and generation.

Primary area:
interpretability and explainable AI
Keywords:
Sparse Autoencoders, Mechanistic Interpretability, Activation Steering, Causal Interventions, Feature Selection

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_gaEUOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper aims to identify sparse autoencoder features that have a targeted downstream effect when used for steering, in particular features that change the model’s generated text. The key idea is to select features that, when steered, increase the probability of target tokens on a set of prompts, while normalizing against the probability of a set of unrelated distractor tokens on those same prompts.

Critical strengths

I appreciate the scope across several SAE families, and I agree there is value in studying features by their downstream effects rather than static metrics.

Critical weaknesses

Fundamentally, the overall contribution of this paper feels relatively minor and unsurprising. If you explicitly select features by how much they increase the probability of some tokens and decrease the probability of others, it is unsurprising that the same metric looks better on held-out prompts than selecting on positive token mass alone or on logit-lens-style methods. Moreover, the held-out and selection prompts are relatively similar, and both only consider single-token behavior.

Methodologically, there are many other ways to select features that are not compared against — for example, gradient-based attribution or activating text examples.

When evaluating steering, a larger existing benchmark such as AxBench would make more sense than a small homemade dataset, especially because it lets you consider the effects of steering as a frontier over steering strengths: effect size against overall degradation and model coherence.

The writing and clarity are somewhat poor. Much of the prose reads as though it was primarily language-model-generated, and is hard to follow due to heavy use of jargon; simpler language when describing methods and metrics would help. I acknowledge that some of this is a matter of taste, but it is worth raising as a separate critique from the weakness of the results themselves.

Questions for authors

None

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

The paper proposes target-selection margin (TSM), an intervention-based criterion for selecting sparse autoencoder features. Each candidate decoder direction is added to the residual stream at the final prompt position using a feature-specific calibrated scale. Features are ranked by their average intervention-induced change in the log ratio of target to distractor probability mass on selection prompts, then evaluated individually on held-out prompts. Across 18 concepts and five model/SAE settings, TSM improves held-out margin utility over static decoder selection within matched 512-feature pools. Reported selection–evaluation Spearman correlations are 0.379–0.735 for TSM versus 0.155–0.191 for decoder proxies. Additional experiments examine lexical specification, grammatical categories, full-dictionary retrieval, and joint-vector behavior. These support conditional ranking improvements, while exposing specification sensitivity, model-dependent retrieval performance, and no established reliable improvement in multi-step target generation.

Critical strengths
  1. Consistent evidence across several model/SAE settings. Tables 1–2 show improvements over the reported decoder baseline across all five settings, with 52–54 wins out of 54 concept–pool comparisons at K=10 and positive paired confidence intervals. Shared candidate pools and intervention protocols, together with concept-clustered uncertainty estimates, make this stronger evidence than a small collection of successful steering examples. The target-mass-only ablation also helps establish the usefulness of explicitly distinguishing targets from distractors.
  2. Useful separation of ranking performance from other requirements for steering. The complementary lexical splits, harder distractors, grammatical-category experiments, and retrieval audits go beyond the primary fixed-token comparison. The behavioral audits explicitly show that improved target–distractor margins need not produce target tokens, and the paper reports heterogeneous and null outcomes rather than presenting them as successful control. This separation is consequential for interpreting and designing SAE evaluations.
Critical weaknesses
  1. The comparisons do not sufficiently isolate the contribution of causal scoring. Appendix A.1 explicitly states that the decoder baseline ranks the unscaled direction (d_i), whereas TSM and held-out evaluation use the feature-specific intervention (\alpha_i d_i). Consequently, part of the ranking gap could arise because the baseline ignores known intervention magnitudes, even without contextual nonlinear effects. A decoder proxy evaluated on (\alpha_i d_i) is an important missing control. Moreover, SAE4Steer’s Output Score is not evaluated, and DTC uses cached prompt-wide selection scores rather than a matched final-position protocol (Appendices A.1 and C.1). These omissions leave the improvement over appropriately controlled proxies and existing output-utility selection unresolved, which directly affects the assessment of the contribution.
  2. The demonstrated transfer remains narrow, limiting the significance of the result. The main experiment transfers between four selection and four held-out generic short prefixes while preserving the same margin objective (Table 8). Lexical audits help, but do not establish ranking transfer across substantially different contexts. Calibration texts also include extensions of held-out prefixes, although no held-out intervention utilities are used. The behavioral experiments do not provide stronger positive evidence: neither selector produces a target as the global next-token argmax in the 864 main paired cases, and multi-step target-mention improvements remain uncertain. Zero unsteered strict instruction adherence makes the latter audit difficult to interpret (Tables 31–32). These observations do not invalidate the bounded ranking claim, but leave practical utility and broader generalization under-established.
Questions for authors
  1. Baseline clarification and control: How much of the improvement in Tables 1–2 remains when the decoder proxy projects (\alpha_i d_i), rather than (d_i)? Please also report the fractions of candidate features with zero or capped intervention scales. This would help separate the effects of calibration from those of contextual causal scoring.
  2. Assessment-critical matched comparison: How does TSM compare with SAE4Steer Output Score and DTC when their selection scores are recomputed using the same final-position intervention protocol, candidate pools, scales, and selection prompts? This comparison would materially affect the assessment of the incremental contribution over existing output-utility selectors.
  3. Assessment-critical generalization evidence: Does the held-out ranking advantage persist on longer, substantively different contexts when calibration texts are disjoint from evaluation prompts? Such evidence would strengthen the central transfer claim without requiring the method to solve a different downstream steering task.
Reviewer_ZHbyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Target-Selection Margin (TSM), a method that ranks Sparse Autoencoder (SAE) features by how much an intervention on each feature shifts the log ratio of target to distractor probability mass. Unlike typical approaches based on the logit lens, which read out parametric knowledge independently of the input, TSM accounts for how the input contextualizes features. Experiments show that TSM outperforms the Decoder baseline on several criteria, including transfer across inputs.

Critical strengths
  • This work highlights the fact that SAE features should be interpreted in light of the input context.
Critical weaknesses
  • The evaluation compares the proposed method against only one simple baseline, which leaves its general effectiveness in doubt.
  • Similarly, existing work is covered only briefly and is largely absent from the experiments, which makes it difficult to judge the contribution of this paper against prior approaches.
  • The paper does not justify why TSM should take the specific form the authors propose. Since many variants are possible, a more systematic exploration of alternatives would strengthen the work.
Questions for authors
  • The baseline appears to be a logit lens-style computation applied to SAE features, yet this is described only in the Appendix and should probably be moved into the main text.
  • Moreover, since the logit lens has many variants, a single simple baseline seems insufficient; could more advanced variants, such as the tuned lens, be included as additional baselines?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.