Pith. sign in

REVIEW 4 major objections 4 minor 4 references

Mechanistic Interpretability of Antibody Language Models Using SAEs

T0 review · 4 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

Pith's one-line read In an antibody language model, sparse autoencoders that impose a hierarchical latent order yield features that can steer germline gene usage in generated sequences, whereas standard sparse autoencoders yield correlated but non-causal featur

desk verdict Useful negative result on TopK steerability; the O-SAE claim is real but the architecture comparison is confounded and the positive evidence is thin. read the letter →

arxiv 2512.05794 v3 pith:VJLPHIXB submitted 2025-12-05 cs.LG cs.AIq-bio.QM

classification cs.LGcs.AIq-bio.QM
keywords sparseautoencodersmechanisticinterpretabilityantibodylanguagemodelsgenerativesteeringgermlinegeneidentityIGHJ4featuresplittingp-IgGen
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks whether sparse autoencoders can do more than label what an antibody language model has learned—can they actually control what it generates. Using p-IgGen, a 17M-parameter antibody sequence model, the authors train two kinds of SAEs: standard TopK SAEs and hierarchically ordered SAEs. They find that TopK latents correlate cleanly with biological concepts such as CDR regions and germline J gene identity, but amplifying those latents does not reliably shift the model's output. Ordered SAEs, whose latents are forced into a nested hierarchy, produce less intuitive activation patterns but include latents that steer generation monotonically: positively steering latent 12 raises the fraction of IGHJ4 sequences, while steering latent 49 lowers it. The paper's claim is that feature-concept correlation is not causation, and hierarchical SAE structure is what makes generative steering reliable.

What carries the argument

The load-bearing object is the Ordered Sparse Autoencoder (O-SAE), an extension of the standard SAE in which the first m rows of encoder and decoder are reused for every truncation level m (per-index nested grouping) and each truncation contributes a loss weighted by a strictly decreasing probability p_M(m). This forces low-index latents to encode coarse, global structure and later indices to add finer detail. Steering is then performed by adding a scaled decoder vector to the hidden state (h_l* ← h_l + α d^(i)). The hierarchy is what makes latent 12 a 'steerable' feature: unlike TopK latents, which fire on single residues, this low-index latent activates broadly across the J region and, whe

What would settle it

Retrain both SAE types at identical dictionary size and expansion factor on the same p-IgGen layer-3 activations, then repeat the IGHJ4 steering protocol; if a matched TopK SAE also yields a latent whose positive steering monotonically increases IGHJ4 proportion (R≈0.9), the claim that Ordered SAEs are preferable for steering fails.

Watch

Extended reading notes

Core claim

The paper demonstrates that in the autoregressive antibody language model p-IgGen, the latents that best predict a biological label are not necessarily the latents that cause the label. TopK SAEs trained on layer-3 activations yield residue-level features—e.g., latents firing on CDRH3 or on conserved or IGHJ4-specific residues in the J region—that predict IGHJ4 with F1 scores up to 0.93, yet positive steering on these features fails to increase IGHJ4 usage in generated libraries. Ordered SAEs, by contrast, impose a nested hierarchy with strictly decreasing truncation weights; the resulting low-index latents activate broadly across the J region and steer generation in the expected direction.

Load-bearing premise

The central comparison assumes the steerability difference between the two SAE families comes from the hierarchical ordering, not from the different dictionary sizes and expansion factors (TopK: 24,576/32; O-SAE: 6,144/8), which are not matched in the paper's setup.

Editorial extensions

If this is right

  • TopK features with high F1 scores for a concept cannot be assumed to control that concept in generation; steering experiments, not correlation scores, are the test of causal relevance.
  • A single hierarchically ordered latent can push an antibody language model's germline gene usage up (latent 12) or down (latent 49), giving bidirectional control over IGHJ4 proportion.
  • Ordered SAE latents trade intuitive, localization-based interpretability for steerability, so researchers must choose between feature-mapping and generative control.
  • Sparse autoencoders can be applied to autoregressive antibody language models to identify CDR and germline-gene features without task-specific supervision beyond a linear probe.
  • Scaling SAE training on annotated antibody datasets could extend this from gene identity to developability properties such as solubility and specificity.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the two SAE types were trained with different dictionary sizes and expansion factors (TopK d_sae=24,576, r=32; O-SAE d_sae=6,144, r=8), the paper's central comparison would be stronger if the steerability advantage persists at matched capacity; this is a testable extension, not something the paper establishes.
  • The authors' proposed mechanism—feature splitting converts higher-order concepts into residue-level features—predicts that TopK latents trained with larger k or on earlier layers should show more steerability; this could be tested directly.
  • Steering success on IGHJ4 suggests the same protocol could identify steerable latents for other IGHJ/IGHV genes and for continuous properties like binding affinity, potentially enabling rational library design; the paper does not demonstrate this.
  • The observation that low-index O-SAE latents activate broadly across the J region hints that steerability may scale with the abstraction level of a feature, which could inform future architecture design.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper trains TopK and Ordered sparse autoencoders (SAEs) on activations of p-IgGen, an autoregressive antibody language model, and uses linear probes and activation-threshold F1 scores to identify latents correlated with CDR identity and heavy-chain J gene identity. The authors report that TopK SAE latents preserve biological information and are visually interpretable, but that steering on TopK features does not reliably shift the proportion of IGHJ4 in generated sequences. In contrast, for an Ordered SAE, positive steering on latent 12 increases and on latent 49 decreases the IGHJ4 proportion with significant Pearson/Spearman correlations. The paper concludes that Ordered SAEs are preferable for precise generative steering.

Significance. If the comparative claim were supported, this would be a useful contribution to mechanistic interpretability of domain-specific protein language models, and the negative TopK steering result would be a cautionary data point for the field. The paper makes good use of established tools (ANARCI, linear probes, F1-based feature selection) and reports some quantitative results, including per-class IGHJ prediction metrics. However, the central comparative claim is currently under-supported: the two SAE families are compared under different dictionary sizes, expansion factors, and training-data volumes, and the Ordered SAE steering evidence rests on only two latents for one gene in one layer, without a random-latent baseline or a reported success rate. The paper itself acknowledges in Section 5 that exhaustive steering analyses are still needed.

major comments (4)
  1. [§3.1.1–3.1.2 and §4.3] The central TopK-vs-Ordered comparison is confounded. TopK uses d_sae=24,576 (r=32) trained on 1.8M sequences, while O-SAE uses d_sae=6,144 (r=8) trained on a 100,000-sequence subsample, chosen 'to speed up training.' This simultaneously varies dictionary size, expansion factor, active-latent fraction (32/24576 vs 32/6144), and training-data volume. The observed steerability difference in §4.3 may therefore be due to capacity or data volume rather than the hierarchical architecture. The authors should either train both SAEs under matched capacity/data conditions or explicitly test whether TopK with r=8 on the subsample also becomes steerable; otherwise the headline claim is not established.
  2. [§4.3] The claim that Ordered SAEs 'reliably identifies steerable features' is based on only two latents (12 and 49) for a single gene (IGHJ4) in a single layer (layer 3). The paper does not report how many features were tested, the selection rule that singled out 12 and 49, the success rate, or confidence intervals for the steering proportions. Given that Feature '12' was selected from a ranked list, the reported Pearson R=0.939 and p=6.458e-7 are not evidence of reliability. Report all steering attempts, the total number of O-SAE features with F>0.5 that were tested, and the number that produced significant monotonic effects.
  3. [§4.2] The negative TopK steering result is not quantitatively documented. For IGHJ4, only three latents (463, 4720, 6276) are shown in Figure 3, and the text states that attempts across layers and gene identities failed, but this is explicitly relegated to '[data not shown]'. The paper cannot conclude that TopK SAEs are not steerable, or that Ordered SAEs are preferable, without a systematic summary of all steering experiments, including alpha values, effect sizes, and failure criteria. At minimum, tabulate the number of features tested per layer/gene and the fraction that showed any significant effect.
  4. [Equation (6) and §4.3] Steering is performed by adding a scaled decoder vector to the hidden state, but no null baseline is provided. A monotonic correlation between steering factor and IGHJ4 proportion could in principle arise from the scaling intervention itself (e.g., pushing the hidden state toward any decoder direction and shifting the output distribution). The authors should include a control condition with randomly chosen decoder vectors or with shuffled latent indices, and report the distribution of Spearman/Pearson correlations under such null steering. Without this, the specificity of the observed steering effects is unclear.
minor comments (4)
  1. [Abstract and §3.1.2] Typos: 'mecahnistic' in the abstract; 'architecutre' in §3.1.2; 'p-iGgen' in the Table 3 caption. Also, the model name is inconsistently written (p-IgGen vs pIg-Gen).
  2. [Figure 4] The x-axis is labeled 'alpha' but the specific steering factors are not listed, and the number of alpha points used for the correlation is not given. State the alpha values and consider adding error bars or confidence intervals on the proportions (if the 1000 sequences are a single generation, resampling or multiple seeds would strengthen the plot).
  3. [§3.2.1] The SARS-CoV-2 RBD binder/non-binder data are described at length but are irrelevant to the study. The claim that 'a dataset of equivalent size randomly sampled from OAS should produce the same results' is an unsupported conjecture; either remove this dataset description or justify its use.
  4. [General] The manuscript does not state whether code, trained SAEs, or the exact activation extraction pipeline will be released. For a mechanistic interpretability paper, code and model access are important for reproducibility; please add a Data Availability statement.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: steerability is an external causal test; self-cited O-SAE code is not load-bearing.

full rationale

The paper's derivation chain is not circular. The steering experiments are external causal tests: latents are selected via linear-probe correlation with IGHJ4, but the reported outcomes (Pearson R = 0.939 for latent 12 and R = -0.705 for latent 49) are measured from generated-sequence composition after adding the decoder vector. Correlation and steering outcome are distinct quantities; the TopK section itself demonstrates this by showing that highly correlated latents are not necessarily steerable (Figure 3). Thus the Ordered SAE steerability result does not reduce to the fitting procedure by construction. The O-SAE implementation is adopted from the authors' own repository (Parsan et al., 2025b), but this is code reuse, not an argument: the paper defines and explains the O-SAE architecture internally, and the steerability claim is benchmarked against actual generation rather than imported from the self-citation. The difference in dictionary size (24,576 vs 6,144) and expansion factor (r=32 vs r=8) between TopK and O-SAE is a real confound that undermines the comparative architecture claim, but that is a correctness/experimental-design concern, not circularity. No fitted parameter is renamed as a prediction, and no uniqueness theorem or ansatz is load-bearing from self-citations. Accordingly, no specific circular step can be quoted with a reduction; the appropriate score is at the low end.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central claim depends on several hand-picked hyperparameters (dictionary sizes, sparsity k, F1 thresholds, steering αs) and on domain assumptions about what makes a latent 'steerable'. No new physical or biological entities are introduced. The most consequential free parameter is the O-SAE dictionary size (6,144) relative to TopK (24,576), which confounds the comparison.

free parameters (8)
  • TopK expansion factor r = 32
    Sets dictionary size to 24,576; chosen from prior works; affects capacity and the steerability comparison with O-SAEs.
  • O-SAE expansion factor r = 8
    Sets dictionary size to 6,144; chosen to speed training; different from TopK, confounding the architecture comparison.
  • Sparsity k = 32
    Number of active latents for both SAEs; affects feature granularity and steering behavior.
  • Feature-selection F1 threshold = 0.5
    A latent is called a 'feature' if its max F1 across thresholds exceeds 0.5; ad hoc and used to define steerable candidates.
  • Activation thresholds = 0.1, 0.2, 0.5, 0.8, 0.9
    Multiple thresholds tested; the maximum F1 is used, leading to optimistic feature identification without multiple-testing correction.
  • Top-500 correlation weight cutoff = 500
    Only the top 500 positive-correlation latents from the linear probe are validated; arbitrary.
  • O-SAE training subsample size = 100,000
    Subsampled from 1.8M sequences to reduce training time; may affect feature quality and representativeness.
  • Steering factors alpha = not specified
    The α values used in Eq. 6 are not listed in the paper; without them the steering dose-response cannot be reproduced.
assumptions (5)
  • domain assumption Thresholded SAE activations define meaningful binary features
    Methods §3.2.3 assumes a latent is interpretable if thresholding its activation predicts a label with F1 > 0.5; this is the operational definition of 'feature'.
  • domain assumption Lower Ordered-SAE dictionary index corresponds to higher-level, more steerable features
    In §4.3, features are ranked by dictionary index to select latent 12 and latent 49; the hierarchy is assumed to reflect abstraction and steerability without independent validation.
  • domain assumption Steering via decoder-vector addition is a valid causal intervention
    Eq. 6 (Templeton et al. 2024) is applied directly; the paper relies on this to measure steerability.
  • domain assumption p-IgGen hidden activations from concatenated VH/VL sequences are representative of antibody-language-model internals
    Methods §3.1 trains SAEs on these activations and assumes they capture the model's relevant biology.
  • domain assumption The TopK vs O-SAE comparison isolates architecture effects
    Implicitly assumed in §4.3 despite differing dictionary sizes and expansion factors; this is the weakest assumption, as capacity may drive the observed differences.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mechanistic Interpretability of Antibody Language Models Using SAEs." pith.science (2026). https://pith.science/paper/VJLPHIXB

@misc{pith2026251205794,
  author       = {Pith},
  title        = {Pith review of: Mechanistic Interpretability of Antibody Language Models Using SAEs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VJLPHIXB}},
  note         = {Machine review of arXiv:2512.05794}
}
read the original abstract

Sparse autoencoders (SAEs) are a mechanistic interpretability technique that have been used to provide insight into learned concepts within large protein language models. Here, we employ TopK and Ordered SAEs to investigate autoregressive antibody language models, and steer their generation. We show that TopK SAEs can reveal biologically meaningful latent features, but high feature-concept correlation does not guarantee causal control over generation. In contrast, Ordered SAEs impose a hierarchical structure that reliably identifies steerable features, but at the expense of more complex and less interpretable activation patterns. These findings advance the mechanistic interpretability of domain-specific protein language models and suggest that, while TopK SAEs suffice for mapping latent features to concepts, Ordered SAEs are preferable when precise generative steering is required.

Figures

Figures reproduced from arXiv: 2512.05794 by the authors.

Figure 1
Figure 1. Latent activations (a) and neuron activations (b) for CDRH3 identity, and latent activations for IGHJ3 (c). The x-axis shows the amino-acid sequence of the VH region of a test antibody; the y-axis shows normalised activation. CDRs are coloured CDRH1 (red), CDRH2 (blue), and CDRH3 (green). Latent activations localise to the expected regions—CDRH3 in (a) and the heavy J region in (c)—whereas neuron activations (b) are… view at source ↗
Figure 2
Figure 2. Comparison of absolute positional (a) and IMGT (b) activations of top three IGHJ4 latents. The sequence/IMGT positions are shown on the x-axis. For the sequence positions, the amino acid sequences were end-padded to a constant length of 350. Percentage of total activations on any given position across validation IGHJ4 sequences is shown on the y-axis. The most frequent IMGT position for activation is highlighted for… view at source ↗
Figure 3
Figure 3. Results of IGHJ4 feature steering for latent 463 (a), 4720 (b), 6276 (c). Y-axis shows the proportion of generated sequences. Plots are coloured by heavy J gene identity. X-axis shows the steering factor used (alpha). Results are for a library of 1000 p￾IgGen-generated sequences. For each latent tested (a-c), steering did not result in a predictable change in library composition. ranked features with an F-score > 0.… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Results of IGHJ4 steering using Ordered latent 12 (a) and 49 (b). Y-axis shows the proportion of generated sequences. Plots are coloured by heavy J gene identity. X-axis shows the steering factor used (alpha). Results are for a library of 1000 p-IgGen￾generated sequenc…
Figure 5
Figure 5. Figure 5: IMGT activations of latent 12 (a) and 49 (b). Activation patterns of both latents show scattered distribution across the range of IMGT positions. 10 [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

4 extracted references · 2 linked inside Pith

  1. [2021]

    doi: 10.1084/jem.20210281

    ISSN 0022-1007. doi: 10.1084/jem.20210281. URL https://www.ncbi.nlm.nih.gov/pmc/ articles/PMC7933983/. Bussmann, B., Nabeshima, N., Karvonen, A., and Nanda, N. Learning Multi-Level Features with Matryoshka Sparse Autoencoders, March 2025. URL http://arxiv. org/abs/2503.17547. arXiv:2503.17547 [cs]. Chanin, D., Wilken-Smith, J., Dulka, T., Bhatnagar, H., a...

  2. [2022]

    doi: 10.1002/pro.4205

    ISSN 1469-896X. doi: 10.1002/pro.4205. URL https://onlinelibrary.wiley. com/doi/abs/10.1002/pro.4205. eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/pro.4205. Parsan, N., Yang, D. J., and Yang, J. J. Towards Inter- pretable Protein Structure Prediction with Sparse Autoen- coders, March 2025a. URL http://arxiv.org/ abs/2503.08764. arXiv:2503.08764...

  3. [3224]

    URL https://www.frontiersin.org/journals/ immunology/articles/10.3389/fimmu

    doi: 10.3389/fimmu.2024.1471396. URL https://www.frontiersin.org/journals/ immunology/articles/10.3389/fimmu. 2024.1471396/full. Publisher: Frontiers. Dunbar, J. and Deane, C. M. ANARCI: antigen receptor numbering and receptor classification.Bioinformatics, 32(2):298–300, January 2016. ISSN 1367-4803. doi: 10.1093/bioinformatics/btv552. URL https://doi. o...

  4. [4920]

    URL https: //doi.org/10.1007/s13194-024-00614-4

    doi: 10.1007/s13194-024-00614-4. URL https: //doi.org/10.1007/s13194-024-00614-4. Lefranc, M.-P., Pommi ´e, C., Ruiz, M., Giudicelli, V ., Foulquier, E., Truong, L., Thouvenin-Contet, V ., and Lefranc, G. IMGT unique numbering for immunoglob- ulin and T cell receptor variable domains and Ig super- family V-like domains.Developmental and Comparative Immuno...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.