Pith. Sign in

REVIEW

Probing and steering biology across Boltz-1s trunk-diffusion boundary

PaperPDF from arXiv

data files

ReviewThe objections and verdict, as JSON Paper breakdownThe authors' claims, conditions and measurements, as JSON Signed receiptProof this review has not been altered
X Bluesky LinkedIn Reddit Hacker News
Piotr Jedryszek, Tongmeng Xie, Adam Winnifrith, Alexander Hasson, Weronika Ślesak, George Wicks, Toby Winnifrith, Oliver M. Crook

Quantitative Methods · Submitted 11 Aug 2026 · 12 references

Pith's verdict: A careful, well-controlled probing and steering study of Boltz-1 that establishes a real trunk/diffusion representational split; the strand null is real but scoped, and the paper mostly says so.

T0 · AI review

How this review was made: An AI referee (deepseek-v4-flash) read the paper on August 15, 2026 and answered a fixed list of questions about it. T0 is the starting level every paper gets. The levels describe how deeply Pith has checked the paper, never the authors.

  1. T0 AI review An AI read the paper and answered a fixed list of questions.
  2. T1 Audit Automatic checks that the references and files the paper relies on exist and match.
  3. T2 Compute The paper's code or calculations re-run with the same results.
  4. T3 Formal statement The claims written out in a formal math language a computer can read.
  5. T4 Proof Those claims proved, and the proof checked by computer.

How Pith checks papers →

· 1 major objection · 4 minor · Ask about this paper · The compiled record · Read the full review →
Challenge this review Comment (0) Read on arXiv ↗
Cite ▾
arXiv ID2608.11475 (version 1)
Pith IDWFFMKCFI, Pith's permanent record of this paper and its review
CitationAdam Winnifrith, Alexander Hasson, George Wicks, Oliver M. Crook, Piotr Jedryszek, Toby Winnifrith, Tongmeng Xie, Weronika \'Slesak. "Probing and steering biology across Boltz-1s trunk-diffusion boundary." arXiv:2608.11475 (2026). https://pith.science/paper/WFFMKCFI
Cite this reviewBibTeX for Pith's review of this paper
Claim paper

Why Pith says this

Highlight any text to comment on it.

Referee report

How Pith checks papers → · Full audit of this paper →

The paper

The paper claims

In Boltz-1, a beta-strand signal that is highly decodable from the trunk (F1 0.82) cannot steer predicted structure through the single representation, while helix and coil directions do, showing that linear decodability does not imply…

The reading

Boltz-1, a protein structure predictor in the AlphaFold3 family, separates a trunk that builds representations from a diffusion module that generates coordinates. This paper asks what biological information survives that boundary and answers by decoding per-residue activations from both modules. It finds that geometry (helix, strand, coil, disorder) stays linearly decodable across the boundary, while sequence chemistry (amino-acid identity, signal peptides, disulfide bonds) is strongly attenuated in the diffusion module. A second result separates representation from use: adding a highly predictive beta-strand direction to the trunk's single representation does not change predicted strand content, even though helix and coil directions do steer structure dose-dependently. If these claims hold, interpretability claims about protein structure models must distinguish what is decodable from what is causally active.

What carries the argument

The argument is carried by the Boltz-1 architecture's two representation spaces: the Pairformer trunk's per-residue single representation s and residue-pair representation z, and the diffusion module that consumes them to generate coordinates. The paper reads these spaces with logistic-regression linear probes and TopK sparse autoencoders (k=256, 2048 latents) at matched fractional depth, and tests causal influence by adding a fitted direction u to the final trunk s at dose k while leaving z fixed, comparing against matched-norm random directions. The steering protocol applied to s alone is the key mechanism that separates decodability (probe F1) from steerability (own-state DSSP fraction change), and the z→s information flow hypothesis explains why strand, a pairwise property, is not reachable by single-representation steering.

Extended reading notes

Core claim

The paper's central discovery is a representational split across the trunk–diffusion boundary of Boltz-1, together with a dissociation between decodability and causal influence. Linear probes read both geometry and sequence chemistry from the Pairformer trunk (amino-acid identity at F1 0.996, signal peptide 0.76, helix/strand/coil 0.79–0.90), but in the diffusion module geometry is essentially unchanged (Δ ≤ 0.02) while sequence chemistry drops sharply (signal peptide 0.76→0.35, disulfide 0.43→0.10, amino-acid identity 1.00→0.66). Steering the final trunk single representation with a supervised probe direction raises helix and coil content dose-dependently against matched random controls, yet a strand direction with held-out F1 0.82 and precision ≥0.93 produces no measurable increase in strand content. The paper argues that linear decodability at a site does not imply causal use at that site, suggests that strand formation is carried by the pair representation z which the intervention did not touch, and shows that evaluating against sparse SwissProt annotations suppresses F1 because unannotated correct residues are counted as false positives.

Load-bearing premise

The result that a decodable strand signal is not causal rests on treating the steering protocol — adding a single direction to the trunk's per-residue representation while leaving the residue-pair representation untouched — as a fair test of causal use; if strand information acts through residue-pair structure or needs nonlinear combinations that additive steering cannot express, the strand null is an artifact of the intervention family rather than evidence that the model ignores the signal.

Editorial extensions

If this is right

  • Interpretability studies of structure predictors should report whether a feature is decodable, steerable, or both; the two properties do not coincide even within the same model and layer.
  • Steering sequence-chemistry concepts such as signal peptides or disulfide bonds is unlikely to succeed through the single representation; the pair representation z is the natural next intervention target, especially for strand and disulfide patterns.
  • Scores against sparse curated annotations (SwissProt) should be treated as lower bounds; controlling with dense labels such as DSSP is necessary to separate under-annotation from decoder failure.
  • For any concept with existing labels, supervised probes are the reliable readout and steering direction; single SAE features can overstate alignment on rare concepts through best-of-many selection, so SAEs are best used for hypothesis generation.
  • The trunk–diffusion split predicts that, in Boltz-1, diffusion activations will support geometric decoding but not sequence-chemistry decoding across sampling steps; this gives a concrete signature to test on other AlphaFold3-class models.

Reading between the lines

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

  • If the z→s flow explanation is right, steering z along a strand direction should raise strand content; this is a direct, testable extension that would also distinguish the paper's architectural hypothesis from a mere limitation of additive single-representation interventions.
  • The same decodability-versus-steerability probe could be applied to the sequence-chemistry concepts the trunk uniquely retains (signal peptide, disulfide): the paper's Limitations section notes this remains untested, and a null there would strengthen the claim that chemistry is consumed without being re-represented.
  • Activation patching between trunk and diffusion layers (for example, ablating a chemistry direction in the trunk and measuring diffusion output) could reveal whether attenuated decodability means the information is genuinely dropped or merely redistributed into nonlinear form.
  • Because the steering effect is several-fold larger in low-confidence structures, a practical design implication is that representation-level control of fold state is strongest where the model is uncertain; exploiting this in generative settings would be an extension beyond the paper's scope.
Original abstract

AlphaFold3-class structure predictors pair a representational trunk, which processes sequence and context, with a diffusion module, which generates atomic coordinates. How biological information changes as it crosses this architectural boundary remains poorly understood. We analyze per-residue activations from the Pairformer trunk and diffusion module of Boltz-1 using linear probes, sparse autoencoders (SAEs), and causal interventions. From the trunk, both geometry (secondary structure, disorder) and sequence chemistry (amino-acid identity, signal peptides, disulfide-bond annotations) are linearly decodable. In the diffusion module, the two diverge. Secondary structure transfers essentially unchanged, whereas sequence chemistry is strongly attenuated. We then test whether decodable directions can steer the model, intervening on the final trunk single representation that conditions the diffusion module. Helix and coil directions change predicted structure dose-dependently against matched-norm random controls, but a beta-strand direction that is highly predictive (F1 =0.82) produces no measurable increase in strand content: linear decodability does not imply causal influence at the site we tested. The same probes also score markedly lower against sparse SwissProt annotations than against dense DSSP labels, because unannotated residues that the model gets right are charged as false positives; such scores are therefore lower bounds. Finally, supervised probes outscore single SAE features wherever a label already exists. We release the trained trunk and diffusion SAEs, Boltz-1 per-residue activations, and the analysis code.

The compiled record

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 8,233 input tokens.

What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.95 · human participants 0.03 · controlled comparison 0.96 · randomized assignment 0.03 · states quantitative result 0.98 · compares with prior results 0.77 · mentions code or data availability 0.99.

The abstract, sentence by sentence

arXiv listing abstract v1. Results and the main claim are underlined; .

AlphaFold3-class structure predictors pair a representational trunk, which processes sequence and context, with a diffusion module, which generates atomic coordinates. How biological information changes as it crosses this architectural boundary remains poorly understood. We analyze per-residue activations from the Pairformer trunk and diffusion module of Boltz-1 using linear probes, sparse autoencoders (SAEs), and causal interventions. From the trunk, both geometry (secondary structure, disorder) and sequence chemistry (amino-acid identity, signal peptides, disulfide-bond annotations) are linearly decodable. In the diffusion module, the two diverge. Secondary structure transfers essentially unchanged, whereas sequence chemistry is strongly attenuated. We then test whether decodable directions can steer the model, intervening on the final trunk single representation that conditions the diffusion module. Helix and coil directions change predicted structure dose-dependently against matched-norm random controls, but a beta-strand direction that is highly predictive (F1 =0.82) produces no measurable increase in strand content: linear decodability does not imply causal influence at the site we tested. The same probes also score markedly lower against sparse SwissProt annotations than against dense DSSP labels, because unannotated residues that the model gets right are charged as false positives; such scores are therefore lower bounds. Finally, supervised probes outscore single SAE features wherever a label already exists. We release the trained trunk and diffusion SAEs, Boltz-1 per-residue activations, and the analysis code.

background contribution method result interpretation main claim

What the classifier reports for each sentence. The number is the probability it gave the label.
#sentencerolethe authors' own workstates the main conclusionstrength
s0 AlphaFold3-class structure predictors pair a representational trunk, which processes sequence and context, with a diffusion module, which generates atomic coordinates. background 1.00 0.05 0.02 asserted 1.00
s1 How biological information changes as it crosses this architectural boundary remains poorly understood. gap 1.00 0.05 0.03 asserted 0.91
s2 We analyze per-residue activations from the Pairformer trunk and diffusion module of Boltz-1 using linear probes, sparse autoencoders (SAEs), and causal interventions. method 0.86 0.97 0.04 asserted 0.97
s3 From the trunk, both geometry (secondary structure, disorder) and sequence chemistry (amino-acid identity, signal peptides, disulfide-bond annotations) are linearly decodable. result 1.00 0.92 0.12 asserted 1.00
s4 In the diffusion module, the two diverge. result 0.97 0.61 0.09 asserted 0.98
s5 Secondary structure transfers essentially unchanged, whereas sequence chemistry is strongly attenuated. result 1.00 0.94 0.27 asserted 1.00
s6 We then test whether decodable directions can steer the model, intervening on the final trunk single representation that conditions the diffusion module. method 0.74 0.91 0.05 asserted 0.51
s7 Helix and coil directions change predicted structure dose-dependently against matched-norm random controls, but a beta-strand direction that is highly predictive (F1 =0.82) produces no measurable increase in strand content: linear decodability does not imply causal influence at the site we tested. result 0.88 0.97 0.45 asserted 1.00
s8 The same probes also score markedly lower against sparse SwissProt annotations than against dense DSSP labels, because unannotated residues that the model gets right are charged as false positives; such scores are therefore lower bounds. limitation 0.43 0.92 0.03 asserted 1.00
s9 Finally, supervised probes outscore single SAE features wherever a label already exists. result 1.00 0.91 0.05 asserted 1.00
s10 We release the trained trunk and diffusion SAEs, Boltz-1 per-residue activations, and the analysis code. contribution 0.98 0.83 0.03 asserted 0.93

What kind of paper the classifier reports this is: empirical study 1.00 · evidence from observation or experiment 0.44.

What was examined, and what was not

sources read
  • arXiv listing abstract v1
not examined
  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers
levels run
  • P abstract
not run
  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. Every number in the measurements table is the authors', checked by code. Record as JSON

Figures
Figure 1. The diffusion module retains geometry while attenuating sequence chemistry. (A) For depth-matched module outputs (pf L47 vs. diff L22), geometric features are decodable from both the pairformer and diffusion module. In contrast, sequence-chemistry concepts are more decodable in the pairformer module compared to the diffusion module. (B,C) Diffusion probe-raw over depth × sampling step: helix (B) is bright across the whole grid, already resolved at the noisiest step; signal peptide (C) is weakly decodable and fades along both the layer and sampling-step axes. Helix (probe) Helix…
Figure 2. Which trunk directions causally steer, and one protein driven both ways. Per-concept population own-state effect (k=16, n=97) against a 19-direction random null; and held-out A6NI15 under baseline, helix, and coil steering. identical trunk activations, the same concepts decode far better against dense DSSP than against SwissProt’s sparse experimental annotations, with helix at 0.85 versus 0.42 and strand at 0.83 versus 0.33 (Fig. S1). As a self-consistency check, computing DSSP on Boltz-1’s own predicted structure rather than a third-party AlphaFold structure changes helix probe-raw…

Figures reproduced from arXiv:2608.11475 by the authors.

References
How the literature uses it

1 arXiv paper cite this in 3 sentences

Each citing sentence is taken from the citing paper's LaTeX source and typed by Pith's classifier. A label appears at probability 0.8 or more (0.95 for support), a contest at 0.7 or more; the quoted words are the citing authors'. How each label was checked · JSON

Papers citing it

Uses its method or tool 1 paper
  • Following [ref] we deduplicate each seed's dictionary (merging directions above cosine 0.95), then match every alive seed-1 direction one-to-one to seeds 2 and 3 by decoder cosine (Hungarian assignment).

    Efficient Auto-Interpretability of AI Models in Biology · Methods · 0.98

Uses its data or values 1 paper
  • We analyze SAEs (TopK, $k = 256$, dictionary size $2048$, input dimension $384$) trained on Boltz-1 Pairformer trunk activations (layer 20, recycle 1) https://huggingface.co/collections/evolve-away/boltz-saespublicly available Hugging Face collection of the SAEs and made available by a previous study [ref] .

    Efficient Auto-Interpretability of AI Models in Biology · Methods · 0.89

Cites it as support 1 paper

What it cites

Compares with 1 paper
  • Concurrently with our work, [ref] scale SAEs to the diffusion modules of Boltz-2 and AlphaFold3 and likewise find a geometric rather than biochemical focus, consistent with what we report here for Boltz-1's diffusion module; [ref] steer a generative structure model with SAE features, whereas our results suggest supervised probes are more reliable steering directions.

    FoldSAE: Learning to Steer Protein Folding Through Sparse Representations · Introduction · 0.84

Cites as background 1 paper
  • Related probing work has looked inside individual modules: [ref] argue AlphaFold3's Pairformer compresses evolutionary information into an increasingly linearly-decodable latent geometry, and [ref] use activation patching in ESMFold to separate an early biochemical stage from a later geometric one.

    Two Stages of Folding: Convergent Mechanisms in AI Protein Folding Trunks · Introduction · 0.96

Discussion

Disagree with the review, or know something it missed? Leave a comment, highlight any passage above to comment on it, or challenge this review formally.