acceptodds
Under review as a conference paper at ICLR 2027

MAxBench: A Multinomial concept recovery benchmark

Abstract

Concept localization and activation steering are increasingly common in interpretability research and its applications. Many existing methods implicitly assume binary or one-dimensional concepts, but many concepts are not binary: some contain many values, hierarchical structure, ordered values, among others. Recent work has explored multinomial concepts and found that their representation geometries are far more varied than those of binary concepts; this makes it unclear which methods or assumptions, and thus which localization and steering methods, are most appropriate. In this study, we introduce MAxBench, a geometry-agnostic evaluation framework for multinomial concept representations based on sampling from the recovered concept representation. Comparing localization methods across a range of geometry types, concepts, and models, we find the following: (i) where the concept representation is centered often matters as much as (and sometimes more than) the choice of basis and the dimensionality of the concept representation; (ii) affine subspaces outperform directions and linear subspaces (even at rank one); (iii) methods within the same geometry class tend to perform similarly; and (iv) no method consistently outperforms prompting, especially for larger models. These findings underscore the importance of expanding meta-evaluations of interpretability research to concepts with more varied internal structure.

Primary area:
interpretability and explainable AI
Keywords:
representation geometry, multinomial concepts, concept localization, affine subspaces, steering, interpretability

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_TRSqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MAxBench evaluates multinomial concept representations through sampled activation interventions. It compares directional, linear, affine, SAE-based and spline representations using concept, fluency, instruction-following and newly introduced diversity scores. Experiments cover six concepts, four instruction-tuned models and multiple training subsets. The main empirical findings are that centering strongly affects performance, affine variants generally outperform origin-centered variants under this protocol, and prompting remains a strong baseline for larger models.

Critical strengths
  1. The paper addresses a useful evaluation gap: recovering a concept with multiple values requires assessing coverage as well as successful concept injection. The proposed sampling-based framework makes this distinction explicit.

  2. The experiments separate several relevant factors, including rank, offset and basis, and include prompting, isotropic sampling, held-out concept instances/categories and supplementary layer sweeps. These controls provide informative evidence about when simple representations can compete with more complex methods.

  3. The evaluation spans four models and six structured concepts, with ten concept-training splits. The paper provides metric breakdowns, judge prompts and human agreement results for concept, fluency and instruction scores, supporting inspection of the reported findings.

Critical weaknesses
  1. The diversity normalization does not guarantee the claimed range. Appendix B.2 divides instance entropy by log(N), where N = 150 is the number of outputs, despite allowing multiple entities per output. For example, 300 distinct equally frequent entities give log(300)/log(150) = 1.138, exceeding the stated instance-score bound. The appendix acknowledges this possibility but supplies no fixed correction or evidence that every evaluated setting satisfies the assumed sparsity. Since diversity enters the benchmark ranking, its normalization needs a well-defined, consistently bounded rule.

  2. Diversity extraction does not consistently measure valid concept coverage. Appendix E.2 explicitly maps non-animal uses such as programming “Python” to an animal and “train your model” to a vehicle, while the concept judge distinguishes intended meanings from such uses. Generic references also count as instances. Thus lexical ambiguity can increase diversity without recovering additional valid concept values. Table 10 validates only concept/fluency/instruction judgments, not entity extraction or taxonomy assignment; an audit of the novel diversity component is missing.

  3. Sampling can depend on parameterization rather than the recovered subspace. Section 3.2 uses s = Lz + b with isotropic z, so sample covariance is sigma^2 L L^T. Factor-analysis loadings are not generally orthonormal, and Appendix A.2 explicitly leaves SAE directions non-orthogonal. Two bases spanning the same subspace can therefore receive different scores. Moreover, Appendix B.1 restricts sigma to [r_min, r_max], although the RMS norm for an origin-centered orthonormal rank-k sampler is sigma sqrt(k). Orthonormalization and reported norm-acceptance diagnostics are needed to disentangle geometric quality from sampling artifacts.

  4. The directional-versus-affine comparison changes the intervention semantics. Rank-1 directions are sampled with a zero-centered Gaussian, including both signs, and all methods use h′ = (1 - alpha)h + alpha s. This can push a signed concept-presence direction toward its negative side and suppress unrelated activation components. Footnote 3 reports diversity collapse under additive steering but provides no corresponding quantitative comparison. Positive-direction sampling, a mean-only baseline and context-preserving interventions would test whether the headline offset advantage persists beyond this specific operator.

  5. Prompt-level robustness is weakly established. Section 4 uses only five validation instructions and five test instructions, with 30 samples per instruction. The ten training splits measure sensitivity to concept-data selection, not uncertainty over new instructions. Consequently, repeated samples do not establish broad prompt generalization for the method rankings or prompting comparison. Appendix B.1 also omits the actual alpha grid and the selected decay parameters needed to reproduce this sensitive part of the protocol.

Questions for authors

Clarifications that could materially affect the assessment:

  1. What bounded normalization and semantic validity rules are actually implemented for diversity? Does a human-audited entity/taxonomy evaluation and corrected scoring preserve the main rankings?

  2. Do the offset and geometry conclusions persist with orthonormalized bases, rank-aware norm calibration, positive-only directional sampling and a mean-only control? Please report sampling norm-acceptance rates.

  3. How stable are the rankings across a larger held-out instruction set with prompt-level uncertainty? Please also specify the alpha grid and the initial/final injection strengths and decay duration.

4 comments

We thank the reviewer for the thorough feedback; we’re glad they found our main questions and findings useful and interesting.

Diversity normalization

This would be a good point to clarify. We clip all normalized diversity scores at 1.0, ensuring that the final scores remain within the range of [0, 1]. We now also measure the distribution of diversity scores; we report their distributions in Tab. 15 and Fig. 18 of Appendix K. The results demonstrate that only a small number of ranks naturally yield scores marginally above 1.0, so few values are clipped.

Diversity extraction does not consistently measure valid concept coverage....An audit of the novel diversity component is missing.

We agree that entity extraction can often be limited; there is indeed quite a bit of fuzziness in deciding what counts as a valid concept instance. This is particularly true for broad concepts, where lexical ambiguity can make entity extraction less straightforward. We now discuss this limitation in App. J in the revised PDF. We have also added an experiment to App. J, which reports the Jaccard similarity of extracted entities from a human (on 100 samples) vs. the GPT LLM judge. The scores are reported in the table below and in the appendix.

Comparisonanimalscountriesvehiclesplantsdaysyears
human–GPT0.6110.7690.6090.6070.8370.826

We observe good agreement between GPT and human annotations for COUNTRIES, DAYS, and YEARS, and reasonable (far better than random) agreement for ANIMALS, VEHICLES, and PLANTS. We also compare human and GPT entity extraction to regular expressions, which do not distinguish semantically relevant vs. irrelevant instances. We have added this analysis and a discussion to the revised manuscript in App. J.

Sampling can depend on parameterization; sampling artifacts

Thanks for pointing out the subtle influence of the L-(ortho)normalization on sampling in identical subspaces. We would like to clarify that we orthonormalize the basis vectors we obtain after factor analysis method using QR decomposition. We have added this L.1132 - L. 1135 in the PDF. For SAEs, we indeed keep the unit norm vectors as they are. Thus, since we used a single sampling method for z uniformly over all affine methods, we do not make isolated claims about the effectiveness of the sampling subroutine. Indeed, if we were to make such specific claims we would need to control for the orthonormalization. To keep the scope manageable, we leave such investigations for future work.

Zero-centered Gaussians for rank-1 methods

We now report additive steering results for additive methods including DiffMean-r1 and REFT-r1 in Table 13. We observe that both additive and convex combination steering strategies achieve similar diversity scores but vary in concept scores with replacement methods getting better overall MAxbench scores than additive methods. We have updated the paper to reflect this finding (L. 254 - L.256) and footnote update in page 5) and added all the results to Table 13 and App. I.

Prompt-level robustness is weakly established

We use a similar number of prompts during evaluation as in AxBench, and generally observe very small variations in performance across ranks for most methods (seen in Fig.2, which reports the variance across 5 test prompts as well as data splits), as well as sufficiently small variances to make claims about the importance of offsets and geometry classes (where confidence intervals often do not overlap). Based on these observations, we expect our findings to generalize to additional held-out prompts as well, as adding more prompts from a similar distribution would be likely to further reduce variance. We acknowledge that evaluating on a larger set of held-out instructions would further increase trustworthiness of the results, however, and we have noted this as a current limitation and direction for future work.

Appendix B.1 omits the actual alpha grid

We tune α0 and α1 in [0, 1], with the constraint that α0 > α1. To find the best alphas, we grid search over for both values (following the above constraint) in [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0]. We inject it throughout the length of generation and our max_tokens are set to 50. We have added these details in App. B.1 L.1135-L.1137.

What bounded normalization and semantic validity rules are actually implemented for diversity? Does a human-audited entity/taxonomy evaluation and corrected scoring preserve the main rankings?

See responses to weaknesses 1 and 2 for details, as well as App. K, Fig. 18 and Tab. 15. As requested, we now report Jaccard similarities of human annotations vs. GPT, as well as GPT vs. regular expressions, in App. J. Agreements are generally well above chance.

Do the offset and geometry conclusions persist with orthonormalized bases, rank-aware norm calibration, positive-only directional sampling and a mean-only control? Please report sampling norm-acceptance rates.

For all methods except for SAEs, the bases are orthonormal. We also tune the standard deviation selection for each rank separately (see L.1296 - L.1299). For directional methods, we also now show affine variant results (in App.H, Tab. 12). For mean-only control, we report isotropic results, where there is no notion of what the bases are (i.e., the only set parameter is the mean/centroid). We unfortunately did not understand what is meant by “norm-acceptance rates”, but we are happy to discuss further. We believe that sampling outside of orthonormal space is only an issue for SAEs, especially given our tuning of the hyperparameters of the sampling to optimize for good scores. We are happy to discuss further!

How stable are the rankings across a larger held-out instruction set

We observe very small variations in performance across ranks for most methods (seen in Fig. 2). Based on this observation, we expect our findings to generalize to additional held-out prompts as well. However, we acknowledge that evaluating on a larger set of held-out instructions would further strengthen our findings, and we have noted this as a direction for future work. We hope this convinces you that with our experiments we do establish a rather broad prompt generalization for the methods we analyse.

Alpha grid and the initial/final injection strengths

We now added these details in App. B.1 L.1135-L.1137.

Thank you for the revisions. Several aspects of my original concerns remain insufficiently addressed:

  1. The human–GPT agreement analysis does not establish whether correcting semantically invalid mentions and taxonomy assignments would preserve the main rankings.
  2. Per-rank tuning does not resolve the concern about the sigma search range, and norm-acceptance rates remain unreported. Positive-only directional sampling is also missing, while isotropic sampling is not equivalent to the requested mean-only control.
  3. Evaluation still uses only five test instructions. Small variation across training splits does not establish generalization across new prompts. My overall rating currently remains 2 (Weak rejection).

Semantically invalid mentions

We thank the reviewer for this suggestion. To directly address this, we conduct an additional experiment where we recompute diversity scores using human-annotated concept entities instead of GPT-extracted entities. Due to time constraints we select four methods from different geometry classes in our main results and annotate a subset of generated sentences for each method and concept. We report the resulting MAxBench scores in the new Tab. 16. We observe that affine version of DiffMean remains the highest-performing method and that the rankings across geometry classes remain highly similar. We also compare affine and linear variants using human annotations in Fig. 19; we observe that affine methods still outperform linear methods. These results provide additional evidence that our main findings are robust to these implementation choices.

sigma search range and norm-acceptance rates

We are unfortunately still unsure what these terms refer to; with clarification, we would be happy to engage with these points further.

positive only directional sampling and mean only control

We report these results now in Fig. 20 and Tab. 18. We find that positive-only sampling improves DiffMean-r1 performance but still performs worse than affine DiffMean (Pairs) (Tab. 18), while the mean-only control performs worse than the affine methods (Fig. 20).

Evaluation still uses only five test instructions.

We now conduct an additional experiment with 20 new test prompts. Due to time and compute constraints, we select four methods from different geometry classes: DiffMean-r1, DiffMean (Pairs), MFA, and SAE (Top-K). We see that ranking across methods remains mostly unchanged (Tab. 17) and affine methods still outperform linear methods (Fig. 20).

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

The paper proposes MAxBench, an extrinsic, geometry-agnostic benchmark for concept representations that are multinomial rather than binary. A recovered representation is evaluated by sampling points from it and interpolating them into the hidden state. The steered generations are then scored by an LLM judge. The scores used are the AxBench concept, fluency and instruction scores, plus a new entropy-based diversity score. The paper evaluates multiple localisation methods, including affine generalisations of existing methods. It finds that affine representations beat directions and linear subspaces, that the offset matters about as much as the basis or rank, that methods within a geometry class behave similarly, and that no method consistently beats prompting for larger models.

Critical strengths
  1. The paper addresses a well-motivated, important problem. Steering and localisation work is evaluated almost entirely on binary concepts, even though recent work shows far richer geometries. A way to compare across geometries is needed.
Critical weaknesses
  1. Some of the contributions seem fairly marginal (e.g., the MAxBench score is the AxBench score expanded with a diversity component), and new contributions could sometimes be separated more clearly from prior work. It isn't clear to me that sampling from and steering towards a concept representation will be the best strategy in all cases, maybe there will be new intervention strategies for different geometries?

  2. Some results could be presented more clearly: for example, Table 2 could have a dimensionality column to make the dimensionality results easier to follow. The differences in Table 2 and Figure 3 are small and often within the margin of error. The main results could be shown for all the models, at the moment Figures 2 and 3 only show a single model. When making claims in the results section, clearly point out which results support those claims (e.g., the Geometry type and Prompting paragraphs starting from line 399 don't mention any Figures or Tables that back up their claims). The claim in the last sentence of the Non-Linear Methods paragraph (starting on line 418) does not seem to be supported by the results in Table 3.

Questions for authors
  1. Could the results, and the evidence backing them up, be presented more clearly?
2 comments

We thank the reviewer for the constructive feedback, and for finding the main questions well-motivated and important!

Some contributions seem marginal

Please see the general response.

Not clear that sampling will be the best strategy

Different interventions could indeed be used for different geometries. A nice property of MAxBench is that it is compatible with future innovations in how steering itself is done, independent of the geometry. For example, one could imagine swapping out only the sampling procedure while maintaining a geometry, and recomputing the evaluation metrics. This is because MAxBench only requires a set of continuations for each prompt. For example, manifold steering does not use the same sampling strategy as other geometries, but we were able to include it; this geometry is quite particular to ordinal concepts represented by spline geometries, and sampling points along the spline. We have revised the PDF to clarify this at L.231 - L.233.

Table 2 could have a dimensionality column.

Table 2 reports scores averaged across concepts, with the best-performing rank selected for each concept and method. Since the best-performing dimensionality can vary across concepts, a single dimensionality column would not accurately represent the results. To provide a more detailed view of how performance varies with dimensionality, we report rank-wise results separately in Figure 2. We revised the PDF L.273 as well as L.2409-L.2413 to clarify this.

The differences in Table 2 and Figure 3 are small and often within the margin of error

While differences between affine methods are small, differences across geometry classes are generally larger. We think this is interesting in itself! As discussed at L.430 - L.433, we see these results as indicating that the geometry class is far more important than the choice of concept localization method. With enough samples, any of these differences could be significant, but we believe that the scale of the differences reflect that these are likely very similar.

The main results could be shown for all the models; at the moment, Figures 2 and 3 only show a single model.

We have added a new Figure (Fig. 4) showing MAxbench scores for all 4 models.

When making claims in the results section, clearly point out which results support those claims

Thank you for this suggestion. We have revised Section 5 to explicitly reference the relevant figures and tables when discussing our findings.

The claim in the last sentence of the Non-Linear Methods paragraph (starting on line 418) does not seem to be supported by the results in Table 3.

Thanks for pointing this out. You are right that spline-fitting with complete data works best only for the Gemma-3-27B model; for smaller models, affine methods remain competitive. Nonetheless, the claim about incomplete data not working as well is still supported by Table 3. We have revised these claims in the paper in L.453-L.457.

Remaining concerns/questions

Thank you for the suggestions. Following your suggestions and those of other reviewers, we have revised the text throughout the paper to improve clarity. We have also tried to make our results and the evidence supporting our claims more explicit wherever possible.

If any questions were not addressed sufficiently, we would be happy to discuss further.

Thank you for your response and for addressing my concerns. I remain somewhat positive about this paper, and will leave my score unchanged.

Reviewer_tNbYOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces MAxBench, a benchmark for evaluating how well concept localization methods recover multinomial concepts (concepts taking many values, possibly hierarchical or ordered) in LLM activations. A localization method returns a geometric object (direction, linear or affine subspace, spline manifold), points are sampled from it, and the residual stream is steered toward each point with a convex combination. Generations are scored with the AxBench concept, fluency and instruction-following LLM-judge scores, plus a new entropy-based Diversity score, combined with a harmonic mean. The benchmark covers 6 concepts (ANIMALS, COUNTRIES, VEHICLES, PLANTS, DAYS, YEARS), 4 instruction-tuned models (Gemma-3-1B/27B, Llama-3.1-8B, Qwen-3.5-4B) and about ten localization methods.

Critical strengths
  1. The paper addresses a relevant question. Moving steering evaluation beyond binary concepts is well motivated given recent evidence for multidimensional, polytope-like and manifold concept geometries. The Diversity score is an interesting way to measure whether a representation covers the whole concept rather than a single value.

  2. The offset finding is useful and supported by a direct ablation. Fig. 3 and Fig. 13 compare each affine method with its b=0 counterpart and consistently favour the mean-centred version. This is a practical message for the steering community, consistent with prior work on mean-centring (Jorgensen et al., 2023; Belrose et al., 2023). The isotropic baseline is also a valuable control for the contribution of the basis.

Critical weaknesses

Limited contribution and scope

  1. The novelty over AxBench is modest and not clearly stated. The concept, fluency and instruction scores, the LLM judge, the HM aggregation and the prompting baseline come from AxBench. The new ingredients are the multinomial setting, the Diversity score and the sampling-based steering. The paper should present itself explicitly as an extension of AxBench to multinomial concepts and state what changes.

  2. The scale is small. With only 6 concepts (vs. hundreds in AxBench), general claims such as "methods within the same geometry class perform similarly" or "manifold steering is a particularly strong method" lack support.

  3. The framework is called geometry-agnostic, but almost all methods are affine subspaces. A single non-linear method is evaluated, on one concept (DAYS). No polytope or convex-hull method is evaluated, although these geometries are cited as motivation (Park et al., 2025; Fel et al., 2026b).

  4. Results vary considerably across models and concepts, and some conclusions are not consistent. Sec. 5 first argues that the offset explains most of the gain (flat rank curves in Fig. 2), then that both offset and basis matter (isotropic baseline), then that the offset dominates for small models. The flat curves in Fig. 2 show that additional directions bring little, but not that the offset explains the gains: a better estimated first direction would give the same picture. Only Fig. 3 isolates the offset.

Implementation choices biasing the comparison across geometry classes 5) Offsets are assigned inconsistently. All 1D methods use b = 0, although they could use the global activation mean as the affine methods do. For SAEs, the decoder bias b_dec is discarded and b is set to 0 without justification. With b=0, all 1D and SAE samples are centred at the origin, far from the activation distribution, which unfairly penalizes these methods. b=0 only makes sense as an ablation in the spirit of Fig. 3, which could also be run for SAE and 1D methods.

  1. Sampling coefficients z from a normal distribution does not suit all methods. For a DiffMean_r1 direction, a negative coefficient steers away from the concept. SAE activations are non-negative ReLU outputs, so Gaussian coefficients on decoder columns produce combinations the SAE never generates. This procedure suits affine representations centred at mu_C, not one-sided representations with b=0, which undermines the geometry-agnostic claim.

  2. The authors claim that convex-combination steering improves diversity for 1D methods. This is surprising, since these methods are designed for additive steering. Results with additive steering should be reported, even in the appendix, to support this choice.

Questions for authors
  1. Spline: how many PCA dimensions are kept? Is this swept like the rank of other methods?

  2. Linear and Schatten probes (l.192-193): what does the classifier predict? Sec. 3.1 states R^{dxd}, while App. A.1 states R^{dx|C|} with concept-value labels, which makes more sense.

  3. Alpha schedule: what are the values or grid limits of alpha_0 and alpha_1? They are not given in Sec. 3.2 or App. B.1.

  4. Small-model claim (l.395): where are the results showing isotropic sampling outperforming linear methods for small models? Please point to the corresponding results (F.1?).

Minor:

  1. It would be useful to name the Linear Representation Hypothesis (Park et al., 2023) explicitly in the first sentence of the Introduction.

  2. Define "HM" at l.254 (used later in tables). Typo: "increaisngly" (l.485).

  3. Define "concept value" early with an example (Monday ... Sunday for DAYS). This would clarify the non-linear manifold method (l.207-210).

  4. Suggestion for future work, not required for this version: in SAE feature matching, top-k MCC tends to select collinear features. A selection method that favours complementary directions, such as FBMP in Klotz et al. (2026, ECCV, "Evaluating the interpretability of sparse autoencoders with concept annotations"), may give a better basis.

5 comments

We thank the reviewer for the constructive and thorough review, and for appreciating the relevance of the main research question!

Limited contribution and scope

Please see the general response.

Only 6 concepts

While there are 6 concepts in this study, we emphasize that these concepts are diverse in their internal structure (e.g., continuous vs. categorical, ordinal vs. non-ordered); the concepts were chosen in large part because they have been studied (individually) in multiple prior studies [1,2,3,4], and have developed interest in the interpretability community; and that devising multinomial concepts is not quite as straightforward as binary concepts (like those that have been studied in prior work, including AxBench). Binary concepts have a far simpler conceptual structure, and there are in theory infinitely many of them (e.g., many concepts in AxBench are highly specific auto-generated sentences that most humans would not naturally have come up with). Our focus was on creating a more well-understood and diverse set of concepts, which allowed us to interpret the results individually in addition to aggregate. We have revised the dataset section (L.134 - L.138) to add this motivation. We thank the reviewer for this feedback; we believe this has made the motivation underlying the dataset design clearer!

Is the framework geometry agnostic?

Please see the general response. This refers to the evaluation framework, and not the choice of methods evaluated in this work. For example, our framework is compatible in theory with any structure, as one could also imagine sampling from a convex hull or its surface. Most current steering methods are linear; we aim to evaluate methods more similar to the methods people actually use in current work. While polytopes have been used as descriptive objects, we are not aware of any convex hull-based steering methods; this would be another non-trivial contribution. But we could imagine using the points spanning the concept as the span of the hull, and sampling from the surface of the hull (which would be equivalent to sampling from linear interpolations between concept-related points). Park et al. 2025 do not propose a steering method, and the code for Fel et al. 2026b’s steering method is not public.

Flow of results could be improved

We now revised the manuscript to make the flow of the findings and takeaways clearer. Specifically:

  • L.414 - L415 and L.419 - L.421 - The contribution of the offset is now isolated by Fig. 3 in the revised PDF (and Fig. 14); here, we keep the bases fixed while comparing linear vs. affine methods.
  • L.425 - L.429 - The isotropic baseline addresses whether the choice of basis matters beyond the offset. For smaller models (i.e Gemma-3-1B, Llama-3.1-8B, Qwen-3.5-4B), the isotropic baseline is competitive with affine methods and outperforms linear methods (see Fig. 14). For Gemma-3-27B, the isotropic baseline is not as competitive; for larger models, perhaps the choice of basis is more important.

We have revised Section 5 to clarify the evidence underlying our claims.

Implementation choices

In prior work, methods such as DiffMean-r1, ReFT-r1, and SAE are typically formulated as linear directions without any offset; thus, using b = 0 was motivated by comparability to commonly used methods. That said, we agree that introducing an offset to these methods would be likely to work far better. We have therefore conducted additional experiments (see Fig. 12, Fig. 17 in App.H and Fig. 4 in the revised manuscript). Introducing a non-zero offset improves the MAxBench score for all three methods, further supporting the importance of the offset (even for 1D methods!). However, we see that these 1D methods still underperform other multi-dimensional affine methods, like Factor Analysis or Schatten Probes. We have added these results to Table 12 and App. H.

Sampling coefficients

In Section 2, we define concept representations based on precision (whether the representation captures only information relevant to the target concept) and recall (whether it covers the full range of concept values). When we say geometry agnostic, we mean that sampling-based evaluations do not assume nor force any particular representation geometry. It would indeed be possible to evaluate DiffMean_r1 or SAE features with more optimized sampling methods. In the response to the previous concern, affine variants of SAE, ReFT-r1 and DiffMean-r1 show that sampling can be applied to directions even with non-zero offsets (and these do work better!).

Additive vs. convex combination steering

We now report steering results for additive methods, including DiffMean-r1 and ReFT-r1, in Table 13. We observe that both additive and convex combination steering strategies achieve similar diversity scores but vary in concept scores with replacement methods getting better overall MAxbench scores than additive methods. See the update at L.254 - L.256 and footnote update on page 5; see also new results in Table 13 and App. I.

How many PCA dimensions are kept?

We used 64 dimensions following Wurgaft et al. (2026), as this worked best in their setting. But we agree that it would be interesting to vary the dimensions, an interesting possible analysis for future work.

What do linear and Schatten probes predict?

Thank you for pointing this out, the actual dimensions are R^{dx|C|} and R^{dxd}. This is now updated in the paper (L.202 - L.204). For both probes, the classifier predicts concept instances (examples in Tab. 6 in App. C.1).

What are the values or grid limits of alpha_0 and alpha_1?

We tune α0 and α1 in [0, 1] with the constraint that α0 > α1. We grid search over both values (following the above constraint) in [0.1, 0.2, 0.3, …, 0.9, 1.0]. We have added these details in App. B.1 L.1315 - L.1317.

Where are the results showing isotropic sampling outperforming linear methods for small models?

This can be seen in Fig. 14, where for Llama-3.1-8B, Qwen-3.5-4B and Gemma-3-1B, Isotropic performs better than linear methods. We have updated the text at this claim to refer to the figures in Sec. 5 (L.426-L.427).

Remaining minor concerns

Minor We have addressed minor missing citation and typo issues as suggested. Thanks also for the reference on Klotz et al. (2026), which we added to the related work. We are happy to discuss further if there are any remaining questions or concerns!

References

[1] Engels et al. (2025). “Not all language model features are one-dimensionally linear.” ICLR. https://openreview.net/forum?id=d63a4AM4hb.

[2] Park et al. (2025). “The geometry of categorical and hierarchical concepts in large language models.” ICLR. https://openreview.net/forum?id=bVTM2QKYuA.

[3] Wurgaft et al. (2026). “Manifold steering reveals the shared geometry of neural network representation and behavior.” COLM. https://openreview.net/forum?id=pf3HKAGT2M.

[4] Lee et al. (2026). “Decomposing query-key feature interactions using contrastive covariances.” ICML. URL https://openreview.net/forum?id=t5cfrJfuzD

Dear Reviewer tNbY,

Thank you again for the detailed and constructive feedback! As today is the last day of the discussion period, we would greatly appreciate if you could review and engage with the response. We believe our response and additional experiments address your concerns, but we are happy to discuss further if you have any further questions. If we have successfully addressed your concerns, we kindly ask whether the score could be adjusted accordingly.

Comments on the paper

1 comment

We thank all reviewers for their constructive comments! Here, we address shared concerns. We have also revised the manuscript; changes are highlighted in blue in the PDF.

Novelty

Reviewer tNbY and gYGT state that they believe the contributions seem marginal over AxBench. While it is clear that MAxBench builds on AxBench (hence the choice of name), the multinomial setting poses unique challenges and requires new methods that are beyond an extension of AxBench. Specifically, we needed to:

  • Adapt existing methods for multidimensional and multinomial concepts

  • Propose new concept localization methods, as many prior methods (including difference in means) could not be easily adapted to this multinomial setting in a fair way

  • Curate appropriate multinomial concept datasets; the binary concepts used in prior work are far easier to curate and are often clearly artificial (e.g., the full sentence auto-interp descriptions from AxBench)

  • Devise a new method for sampling several continuations for the same input query

  • Add a diversity metric, as prior work only considered whether the output was related to the concept and had no notion of concept coverage Prior work often focuses on only one of these aspects at a time. For example, [1] focuses on concept geometry specifically in sparse autoencoders, while [2] focuses on proposing one new steering method. Finally, we view the integration of AxBench’s existing metrics as a strength of this work: Rather than introducing an entirely independent measure in a new setting, we provide a framework for evaluating multidimensional concepts while retaining direct comparability with metrics that the community has already gained experience in interpreting in other contexts. We believe this continuity facilitates interpretation and, ultimately, improves the usability and adoption of MAxBench.

Is the framework geometry agnostic?

Our sampling-based evaluation framework is geometry-agnostic in the sense that we can sample from any representation geometry. We do not make claims about the effectiveness of a particular steering strategy. Whether a sampled intervention successfully steers the model toward the concept depends on the quality of the recovered geometry, as well as the intervention strategy used.

References

[1] Engels et al. (2025). “Not all language model features are one-dimensionally linear.” ICLR. https://openreview.net/forum?id=d63a4AM4hb.

[2] Wurgaft et al. (2026). “Manifold steering reveals the shared geometry of neural network representation and behavior.” COLM. https://openreview.net/forum?id=pf3HKAGT2M.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.