Pith. sign in

REVIEW 3 major objections 5 minor 42 references

Induction in Both Directions: A Mechanistic Analysis of In-Context Learning in Masked Diffusion Language Models

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Masked diffusion language models implement in-context learning with a bidirectional induction circuit, copying answers from past or future context.

desk verdict A genuinely new and carefully executed circuit analysis of induction in masked diffusion LMs; the main caveat is that the mechanistic claims are established on folded no-LayerNorm surrogates, not the original trained checkpoints. read the letter →

arxiv 2607.15893 v2 pith:IBFTKATI submitted 2026-07-17 cs.CL cs.AIcs.LG

Andy Catruna , Emilian Radoi This is my paper
classification cs.CLcs.AIcs.LG
keywords diffusionlanguagemodelsinductionheadsin-contextlearningmechanisticinterpretabilitybidirectionalattentionmaskedimplicittimestepQK/OVdecomposition
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 tries to establish how absorbing-mask diffusion language models (DLMs) perform induction, the copy-from-repeated-context operation that underlies in-context learning. Using matched attention-only autoregressive and diffusion transformers, it claims to identify a bidirectional induction circuit: layer-0 previous-token and next-token heads write the immediate left and right neighbours of a masked position into the residual stream, and layer-1 induction heads use QK composition to locate the matching source occurrence and a near-diagonal OV map to copy the answer token. The paper also argues that the DLM's advantage over an autoregressive model appears only when both sides of the mask are visible, and that DLMs compute and causally use the global fraction of masked tokens as an implicit timestep. A sympathetic reader would care because DLMs are an increasingly practical parallel-decoding alternative to autoregressive models, and this is a concrete mechanistic account of how they learn from context.

What carries the argument

The central object is the bidirectional induction circuit. It has two halves: layer-0 'neighbour heads' (previous-token and next-token heads) that attend to positions i−1 and i+1 around a mask and deposit token cues in the residual stream, and layer-1 'induction heads' that read those cues. The mechanism is carried by QK composition—an induction head's attention logit from mask to source is dominated (2.7–8.3 logits versus under 2.4 for other terms) by the term in which the same layer-0 pathway supplies both the query at the mask and the key at the source—together with a near-diagonal OV map that raises the logit of the attended token at the mask. A second piece of machinery is the folded no

What would settle it

Train an unfolded, LayerNorm-keeping DLM with MLPs and more layers, then run the same mean ablations and QK decomposition: if mean-ablating the dominant i−1 / i+1 layer-0 heads no longer drops induction by roughly the full score, or if the layer-1 induction head's QK logit is no longer dominated by the layer-0 pathway, the circuit is an artifact of the folded attention-only setting. A second test: run the same repeated-token task on natural-language-like sequences; if the circuit does not appear or scores collapse, the transfer from synthetic induction to in-context learning fails.

Watch

Extended reading notes

Core claim

The paper's central discovery is that absorbing-mask DLMs do not simply reuse the autoregressive induction circuit; they learn a direction-symmetric two-pathway version. Around each mask, one layer-0 head attends to the previous token and another to the next token, writing these local cues into the residual stream; in layer 1, induction heads compose with those outputs on both the query side (at the mask) and the key side (at the source answer), so the largest QK term comes from the same pathway at both positions, and their OV matrix is near-diagonal in token space, raising the logit of the attended source token at the mask. The circuit is causal: mean-ablating either layer-0 head drops indu

Load-bearing premise

The load-bearing premise is that the two-pathway circuit and the implicit-mask-rate feature found in small attention-only DLMs with folded LayerNorm (briefly fine-tuned to recover performance) are faithful to the original LayerNorm models and would persist in larger DLMs with MLPs; the paper's own Limitations section states this transfer is left open.

Editorial extensions

If this is right

  • If the circuit is real, DLM induction emerges abruptly around 8–24k training steps, needs at least two layers, and is direction-symmetric: reverse induction scores match forward scores (e.g. 4.07 vs 4.13 nats at L2).
  • The DLM's advantage over a matched AR model vanishes when only left context is visible (2.23 vs 3.46 nats at L2); it appears when both sides are visible, with scores close to the sum of the two one-sided scores, supporting the 'bidirectional access, not stronger mechanism' reading.
  • Conflict-prompt tests imply the two pathways can operate independently: with a left source for A and a right source for B, the model splits probability nearly equally, and ablating either layer-0 head shifts the answer toward the other side.
  • Local-context tests imply an effective induction window of about five tokens: matching one neighbour on each side yields 2.28 nats and two yield 4.29, with longer segments adding little.
  • The implicit-timestep evidence implies that DLMs do not need an explicit timestep embedding to track corruption level; the mask-fraction feature is causally used, since patching it moves prediction entropy toward the target corruption condition in all six models.

Reading between the lines

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

  • Beyond the paper: if the local five-token window is the real induction mechanism, then DLM training schedules that keep short spans around masks un-corrupted (e.g. span masking rather than uniform masking) may strengthen in-context learning at inference; this is a testable prediction.
  • Beyond the paper: the near-diagonal OV copy map predicts a specific failure mode—tokens whose OV map is not diagonal should be copied less reliably, and the reported top-1 rate of about one in five on random vocabulary suggests measuring copy accuracy per token could expose which embeddings are least copyable.
  • Beyond the paper: the implicit-timestep result suggests a design principle—DLMs may not need explicit timestep embeddings at all if the residual stream can infer corruption level from the input; an ablation that removes all timestep information from larger DLMs would be a direct extension.
  • Beyond the paper: a direction-symmetric induction circuit implies that DLMs could do 'retrospective' in-context learning, where a later prompt example changes the prediction for an earlier position; a behavioural test on natural text would show whether this is more than a synthetic-task property.
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

3 major / 5 minor

Summary. This paper presents a controlled mechanistic comparison of induction in attention-only autoregressive (AR) transformers and absorbing-mask diffusion language models (DLMs). Using matched architectures and synthetic repeated-token sequences, the authors report that DLMs learn a bidirectional induction circuit: layer-0 'previous-token' and 'next-token' heads write local context into the residual stream, and layer-1 'induction heads' use QK composition to match those features between the masked position and a source answer in either the past or the future, with an OV path that copies the answer. The paper also claims that DLMs compute the global fraction of masked tokens as an implicit timestep, supported by linear probing and activation patching. Behavioral comparisons show that DLMs outperform AR models only when both sides of the mask are visible, and that induction emerges abruptly at L2. The mechanistic analysis, however, is performed on folded no-LayerNorm variants obtained by transforming trained LayerNorm models and briefly fine-tuning them.

Significance. If the central claims hold, this is a valuable contribution to the nascent mechanistic understanding of diffusion language models. The paper is careful in its experimental design: matched AR/DLM architectures, multiple seeds, control sequences, mean ablations, QK/OV decompositions, conflict prompts, and a causal patch test together provide converging evidence for the described circuit. The authors are also unusually transparent about limitations, explicitly noting that the analysis uses folded no-LayerNorm models and synthetic tasks. The implicit-timestep finding is potentially important for understanding how DLMs cope without explicit timestep embeddings. The paper ships reproducible methodology and quantitative detail (Table 1), which strengthens confidence in the behavioral results. The main weakness is that the mechanistic circuit is established only for the folded surrogate, not for the originally trained LayerNorm models that the paper claims to explain; this gap is acknowledged but not closed.

major comments (3)
  1. [§3.3, §4.2–4.4, §5] The central mechanistic claims—head localization, mean ablations, QK/OV decompositions, and the mask-rate probe/patch—are all performed on folded no-LayerNorm checkpoints obtained by replacing LayerNorm with a constant and briefly fine-tuning until performance is recovered. Section 5 explicitly states that this 'leaves open if the mechanisms appear in the original LayerNorm models.' This is load-bearing because performance recovery (perplexity 117.2 vs. 117.5 at L2) does not imply circuit identity: the brief fine-tuning could in principle recruit a different circuit with the same loss. The QK/OV decompositions are only well-defined in the folded linear model. To support the paper's central claim, the authors should either validate the same circuit in the original LayerNorm models (e.g., by showing that the same layer-0/layer-1 heads are causally important via activation patching or atten
  2. [§4.3, Figure 6] The 'OV copy map' is claimed to be 'diagonal' and to implement copying, but the supporting evidence is weak. Figure 6 evaluates only 26 vocabulary tokens and reports that the map 'tends to raise' the logit of the attended token; over the full vocabulary, the copied token becomes top-1 for only about one in five source tokens. A diagonal map that is top-1 for 20% of tokens is a weak copy mechanism. The paper should quantify the strength of the diagonal component more rigorously (e.g., distribution of diagonal vs. off-diagonal logit differences over the full vocabulary, or top-5 accuracy) and temper the claim that the OV path 'copies the answer.' Otherwise, the mechanics section overstates the precision of the copying step.
  3. [§4.4, Figure 8] The implicit-timestep causal patch is performed on the folded no-LayerNorm models, and the probe is also trained on their residual streams. This inherits the same surrogate problem as the circuit analysis. Additionally, the patch recovers only 13–88% of the entropy shift (mean ≈40%), with the authors attributing the partial effect to later layers recomputing the mask rate. This is plausible, but the claim that 'DLMs compute the global fraction of masked tokens and use it as an implicit timestep' would be stronger if the behavioral entropy shift (Figure 8a) were also causally tied to the mask-rate direction in the original LayerNorm models, or if the authors showed that the probe and patch are stable across more architectural variations. As it stands, the implicit-timestep claim is supported only for the surrogate.
minor comments (5)
  1. [§3.4] The reported AR validation perplexity (64.5 vs. 117.5 at L2) is described as a direct comparison, but the DLM value is an ELBO upper bound. The authors note this, but the presentation could more explicitly avoid implying that the DLM is worse in a model-quality sense, since the two quantities are not directly comparable.
  2. [§4.2, Table 1] The table reports 'depth run' but the paper does not define the run labels A/B/C beyond seeds. It would help to state explicitly that these are the three seeds used in the behavioral averages, and to report whether the same folded seeds are used for both behavioral and mechanistic results.
  3. [§4.3, Figure 5] The QK decomposition figure would be clearer if it showed the same scale for all panels or if the paper explicitly stated the range of attention logits across the six models. Currently the reader cannot easily compare the dominance of the 'own pathway' term across runs.
  4. [§4.2] The conflict-prompt experiment is described only briefly. It would help to specify how the two source windows are placed in the sequence (e.g., both in the first copy? overlapping?) and to report the control condition where both windows support the same answer, as the +2.1 to +2.8 range is only mentioned in passing.
  5. [§5] The limitations section is commendable, but it could explicitly note that the QK/OV decomposition and the mask-rate probe are not only simplified but possibly not well-defined for the original LayerNorm models, which is precisely the concern raised above. Adding one sentence acknowledging that the mechanistic claims are about the folded surrogate would improve transparency.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central circuit and implicit-timestep claims rest on independent empirical measurements and control comparisons, not on fitting or self-citation.

full rationale

The paper's derivation chain is empirical rather than definitional. The bidirectional induction circuit is supported by multiple independent measurements on trained checkpoints: attention fingerprints, mean ablations, QK/OV decompositions, and conflict prompts. None of these measurements define the target conclusion in terms of the inputs; ablation drops and QK logit terms are computed from learned weights and compared against control conditions. The induction score is the log-probability difference between repeated and non-repeated control sequences, so the behavioral claims are not fitted parameters renamed as predictions. The implicit-timestep claim uses a linear probe trained to predict the global mask fraction from residual-stream positions, but the causal claim is separately tested by patching the probe direction and measuring entropy shifts, with a random-direction baseline; this is not circular because the probe is a measurement instrument, not an input to the loss or a fitted parameter of the claimed mechanism. The folded no-LayerNorm transformation is a model simplification, and Section 5 explicitly states that it leaves open whether the mechanisms appear in the original LayerNorm models or larger DLMs; this is an acknowledged limitation about external validity, not a case where the analysis reduces to its own assumptions by construction. There are no load-bearing self-citations, no imported uniqueness theorems, and no ansatz smuggled in via citation. The paper compares DLMs against matched AR models and against non-repeated controls, and all load-bearing claims are supported by weight-based and causal evidence independent of the claims themselves.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central claims rely on domain assumptions about transfer from small attention-only models to larger models and from synthetic tasks to natural language. The paper itself names both in the Limitations section. No ad hoc free parameters or invented entities are introduced.

assumptions (4)
  • domain assumption The LayerNorm-folded models faithfully represent the mechanisms of the original LayerNorm models.
    Section 3.4 states 'we briefly fine-tune the resulting no-LayerNorm models until they recover the performance of the LayerNorm models,' and Section 5 admits the analysis 'leaves open if the mechanisms appear in the original LayerNorm models.' This is a load-bearing domain assumption for the circuit analysis.
  • domain assumption Induction in synthetic random-token repeated sequences is the same mechanism as in-context learning in natural language.
    Section 5 states the task 'does not show that the same circuit underlies in-context learning in natural language.' The paper's claims about in-context learning rest on this transfer assumption.
  • standard math Mean ablation over non-repeated control prompts is a valid estimator of a head's causal contribution.
    Used in Section 3.3 and Section 4.2; standard practice in mechanistic interpretability but an assumption that the ablation does not create pathological states.
  • domain assumption The mask fraction is monotonically related to the true timestep under the cosine schedule, so recovering mask fraction from the residual stream counts as an implicit timestep.
    The model is trained without a timestep embedding; the paper's claim that it 'uses it as an implicit timestep' relies on the interpretational step that the learned mask-rate feature is equivalent to a timestep.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Induction in Both Directions: A Mechanistic Analysis of In-Context Learning in Masked Diffusion Language Models." pith.science (2026). https://pith.science/paper/IBFTKATI

@misc{pith2026260715893,
  author       = {Pith},
  title        = {Pith review of: Induction in Both Directions: A Mechanistic Analysis of In-Context Learning in Masked Diffusion Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IBFTKATI}},
  note         = {Machine review of arXiv:2607.15893}
}
read the original abstract

While the internal mechanisms of autoregressive (AR) transformers have been studied extensively, much less is known about diffusion language models (DLMs), an emerging alternative that generates text by iterative denoising. In this work, we study how DLMs implement induction, a mechanism behind in-context learning in which the model finds a repeated context and copies the token that followed it. Our analysis compares attention-only AR models and absorbing-mask DLMs with matched architectures. We find that DLMs learn a bidirectional induction circuit, where previous-token and next-token heads write local context into the residual stream and later induction heads use it to find and copy the answer from the matching source position. The circuit is direction-symmetric, working whether the source appears in the past or in the future. When only left context is visible, matching what an AR model sees, the DLM does not outperform its AR counterpart in induction capabilities. However, we observe it has stronger induction when both sides of the masked token are visible, pointing to bidirectional context access rather than a stronger one-sided mechanism. Beyond induction, we provide causal evidence that DLMs compute the global fraction of masked tokens and use it as an implicit timestep, even though they are given no explicit timestep embedding.

Figures

Figures reproduced from arXiv: 2607.15893 by the authors.

Figure 1
Figure 1. The bidirectional induction circuit identified in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Induction phase change. The induction score [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Context access at L2, averaged over three [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Circuit localization in one representative DLM [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: QK mechanics in representative induction [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 7
Figure 7. Figure 7: Induction under corruption at L2, averaged [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 6
Figure 6. Figure 6: OV mechanics in one representative induction [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 8
Figure 8. Figure 8: The implicit timestep. Gray lines show indi [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 15 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2209.11895 , year=

    In-context learning and induction heads , author=. arXiv preprint arXiv:2209.11895 , year=

    arXiv

  2. [2]

    Transformer Circuits Thread , volume=

    A mathematical framework for transformer circuits , author=. Transformer Circuits Thread , volume=

  3. [3]

    arXiv preprint arXiv:2211.00593 , year=

    Interpretability in the wild: a circuit for indirect object identification in gpt-2 small , author=. arXiv preprint arXiv:2211.00593 , year=

    arXiv

  4. [4]

    arXiv preprint arXiv:2304.05969 , year=

    Localizing model behavior with path patching , author=. arXiv preprint arXiv:2304.05969 , year=

    arXiv

  5. [5]

    Advances in neural information processing systems , volume=

    Structured denoising diffusion models in discrete state-spaces , author=. Advances in neural information processing systems , volume=

  6. [6]

    Advances in Neural Information Processing Systems , volume=

    Likelihood-based diffusion language models , author=. Advances in Neural Information Processing Systems , volume=

  7. [7]

    arXiv preprint arXiv:2310.16834 , year=

    Discrete diffusion modeling by estimating the ratios of the data distribution , author=. arXiv preprint arXiv:2310.16834 , year=

    arXiv

  8. [8]

    Advances in Neural Information Processing Systems , volume=

    Large language diffusion models , author=. Advances in Neural Information Processing Systems , volume=

Show all 42 references
  1. [9]

    arXiv e-prints , pages=

    Mercury: Ultra-fast language models based on diffusion , author=. arXiv e-prints , pages=

  2. [10]

    arXiv preprint arXiv:2602.05859 , year=

    DLM-Scope: Mechanistic Interpretability of Diffusion Language Models via Sparse Autoencoders , author=. arXiv preprint arXiv:2602.05859 , year=

  3. [11]

    arXiv preprint arXiv:2601.07894 , year=

    Revealing the Attention Floating Mechanism in Masked Diffusion Models , author=. arXiv preprint arXiv:2601.07894 , year=

  4. [12]

    arXiv preprint arXiv:2601.14758 , year=

    Mechanism Shift During Post-training from Autoregressive to Masked Diffusion Language Models , author=. arXiv preprint arXiv:2601.14758 , year=

    arXiv

  5. [13]

    arXiv preprint arXiv:2511.19152 , year=

    Masked Diffusion Models are Secretly Learned-Order Autoregressive Models , author=. arXiv preprint arXiv:2511.19152 , year=

  6. [14]

    Advances in Neural Information Processing Systems , volume=

    Simple and effective masked diffusion language models , author=. Advances in Neural Information Processing Systems , volume=

  7. [15]

    Advances in neural information processing systems , volume=

    Simplified and generalized masked diffusion for discrete data , author=. Advances in neural information processing systems , volume=

  8. [16]

    International Conference on Learning Representations , volume=

    Your absorbing discrete diffusion secretly models the conditional distributions of clean data , author=. International Conference on Learning Representations , volume=

  9. [17]

    International Conference on Learning Representations , volume=

    Masked diffusion models are secretly time-agnostic masked models and exploit inaccurate categorical sampling , author=. International Conference on Learning Representations , volume=

  10. [18]

    International Conference on Learning Representations , volume=

    Scaling up masked diffusion models on text , author=. International Conference on Learning Representations , volume=

  11. [19]

    International Conference on Learning Representations , volume=

    Scaling diffusion language models via adaptation from autoregressive models , author=. International Conference on Learning Representations , volume=

  12. [20]

    arXiv preprint arXiv:2508.15487 , year=

    Dream 7b: Diffusion large language models , author=. arXiv preprint arXiv:2508.15487 , year=

    arXiv

  13. [21]

    International Conference on Learning Representations , volume=

    Block diffusion: Interpolating between autoregressive and diffusion language models , author=. International Conference on Learning Representations , volume=

  14. [22]

    arXiv preprint arXiv:2502.06768 , year=

    Train for the worst, plan for the best: Understanding token ordering in masked diffusions , author=. arXiv preprint arXiv:2502.06768 , year=

    arXiv

  15. [23]

    arXiv preprint arXiv:2209.10652 , year=

    Toy models of superposition , author=. arXiv preprint arXiv:2209.10652 , year=

    arXiv

  16. [24]

    arXiv preprint arXiv:2605.29358 , year=

    Scaling monosemanticity: Extracting interpretable features from claude 3 sonnet , author=. arXiv preprint arXiv:2605.29358 , year=

    arXiv

  17. [25]

    International Conference on Learning Representations , volume=

    Sparse autoencoders find highly interpretable features in language models , author=. International Conference on Learning Representations , volume=

  18. [26]

    Advances in Neural Information Processing Systems , volume=

    Towards automated circuit discovery for mechanistic interpretability , author=. Advances in Neural Information Processing Systems , volume=

  19. [27]

    arXiv preprint arXiv:2301.05217 , year=

    Progress measures for grokking via mechanistic interpretability , author=. arXiv preprint arXiv:2301.05217 , year=

    arXiv

  20. [28]

    2026 , month = jun, howpublished =

    DiffusionGemma: 4x faster text generation , author =. 2026 , month = jun, howpublished =

    2026

  21. [29]

    arXiv preprint arXiv:2409.13710 , year=

    You can remove GPT2's LayerNorm by fine-tuning , author=. arXiv preprint arXiv:2409.13710 , year=

    arXiv

  22. [30]

    2022 , howpublished =

    TransformerLens , author =. 2022 , howpublished =

    2022

  23. [31]

    Advances in Neural Information Processing Systems , volume=

    The fineweb datasets: Decanting the web for the finest text data at scale , author=. Advances in Neural Information Processing Systems , volume=

  24. [32]

    OpenAI blog , volume=

    Language models are unsupervised multitask learners , author=. OpenAI blog , volume=

  25. [33]

    arXiv preprint arXiv:1711.05101 , year=

    Decoupled weight decay regularization , author=. arXiv preprint arXiv:1711.05101 , year=

    arXiv

  26. [34]

    Science , volume=

    Experimental evidence on the productivity effects of generative artificial intelligence , author=. Science , volume=. 2023 , publisher=

    2023

  27. [35]

    arXiv preprint arXiv:2302.06590 , year=

    The impact of ai on developer productivity: Evidence from github copilot , author=. arXiv preprint arXiv:2302.06590 , year=

    arXiv

  28. [36]

    Nature , volume=

    Autonomous chemical research with large language models , author=. Nature , volume=. 2023 , publisher=

    2023

  29. [37]

    arXiv preprint arXiv:2404.14082 , year=

    Mechanistic interpretability for AI safety--a review , author=. arXiv preprint arXiv:2404.14082 , year=

    arXiv

  30. [38]

    Proceedings of the 2022 CHI conference on human factors in computing systems , pages=

    Coauthor: Designing a human-ai collaborative writing dataset for exploring language model capabilities , author=. Proceedings of the 2022 CHI conference on human factors in computing systems , pages=

    2022

  31. [39]

    Computational Linguistics , volume=

    Probing classifiers: Promises, shortcomings, and advances , author=. Computational Linguistics , volume=

  32. [40]

    Findings of the Association for Computational Linguistics: EMNLP 2022 , pages=

    On the role of bidirectionality in language model pre-training , author=. Findings of the Association for Computational Linguistics: EMNLP 2022 , pages=

    2022

  33. [41]

    arXiv preprint arXiv:2209.14500 , year=

    Bidirectional language models are also few-shot learners , author=. arXiv preprint arXiv:2209.14500 , year=

    arXiv

  34. [42]

    Computational Linguistics , volume=

    The quest for the right mediator: Surveying mechanistic interpretability for nlp through the lens of causal mediation analysis , author=. Computational Linguistics , volume=. 2026 , publisher=

    2026

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.