Pith. sign in

REVIEW 2 major objections 2 minor 45 references

In Mamba-2, single-bucket probes recover only a small execution layer of the state sink while missing a much larger detection layer with the same representational signature.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.3

2026-06-28 18:28 UTC pith:6H7ZXG7Q

load-bearing objection Single-bucket probes catch only the small execution set in the Mamba-2 state sink while missing the larger detection set, but shared Delta makes the ablation separation less clean than claimed. the 2 major comments →

arxiv 2606.00930 v1 pith:6H7ZXG7Q submitted 2026-05-30 cs.CL cs.AIcs.LG

Detection vs. Execution: Single-Bucket Probes Miss Half the Mamba-2 State Sink

Yuhang Jiang This is my paper
classification cs.CL cs.AIcs.LG
keywords Mamba-2state sinkmechanistic interpretabilityprobesablationDelta gatecircuit discoveryattention sink
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper establishes that probes locating a representational signature for a computation do not necessarily locate the circuit that executes it. In Mamba-2 the state sink, a pattern of strong Delta-gate activation on boundary tokens, splits into BOS-specialist heads that causally drive relevant predictions and dual heads that share the signature but contribute little when removed. Single-bucket probes recover only the small execution set while multi-class aggregation of the same probe recovers the larger detection set. This distinction is shown by ablation: removing the specialist heads destroys retrieval performance while removing the dual heads does not. The result indicates that representational similarity alone cannot be taken as evidence of functional equivalence.

Core claim

In Mamba-2 the state sink decomposes into two functional head sets. Single-bucket BOS-specialist heads, about 5 percent of heads at 2.7B scale, causally support both BOS-context and newline-target predictions across scales and corpora. Dual heads, 27-35 percent of heads recovered by multi-class aggregation of the same probe, exhibit stronger BOS-newline representational similarity yet substantially weaker causal effects under ablation. Representational similarity therefore does not imply functional equivalence, and separating the layers requires class-conditional ablation rather than class-conditional cosine.

What carries the argument

The state sink (disproportionate Delta-gate activation on boundary tokens) decomposed by single-bucket versus multi-class probe aggregation into BOS-specialist and dual head sets, with causal contributions isolated by targeted ablation.

Load-bearing premise

Ablation of the identified head sets accurately isolates their causal contributions without confounding effects from the head-shared Delta projection or post-hoc classification choices.

What would settle it

If ablating the dual heads recovered by multi-class probes also collapses NIAH retrieval accuracy to zero at 1024 context length, the claim that they form a distinct detection layer separate from execution would be falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The paper claims that in Mamba-2, the state sink phenomenon decomposes into distinct functional head sets: a small set of single-bucket BOS-specialist execution heads (~5% at 2.7B scale) that causally support BOS-context and newline-target predictions, and a larger set of dual heads (27-35%) recovered only via multi-class probe aggregation that exhibit stronger representational similarity to BOS-newline but weaker causal effects under ablation. It argues that representational similarity does not imply functional equivalence, shows that ablating the specialist heads collapses RULER NIAH accuracy while size-matched controls do not, and attributes the distinction to Mamba-2's head-shared Delta projection (supported by a random channel-bucketing control).

Significance. If the central empirical distinction holds, the result is significant for mechanistic interpretability of state-space models: it demonstrates that probe-identified representational signatures can correspond to detection rather than execution circuits, necessitating class-conditional ablation to separate them. Strengths include consistent findings across model scales and corpora, use of a random bucketing control to rule out granularity artifacts, and a concrete downstream behavioral demonstration on RULER NIAH retrieval. This cautions against equating probe outputs with circuit identification in architectures with shared components.

major comments (2)
  1. [Ablation experiments (methods and §4)] Ablation experiments (methods and §4): because Delta is head-shared, the procedure of zeroing or masking the dual-head set may still perturb the shared projection matrix for the remaining BOS-specialist heads (and vice versa). The random channel-bucketing control addresses substrate granularity but does not isolate this interaction effect, so the reported difference in causal effect sizes between the two sets could partly reflect confounding through the shared component rather than intrinsic functional separation. This is load-bearing for the claim that the sets are functionally distinct.
  2. [Head classification and probe aggregation (§3)] Head classification and probe aggregation (§3): the post-hoc assignment of heads to BOS-specialist vs. dual sets is downstream of the same probe outputs used to define the representational signature. While the paper invokes a control, it remains necessary to show that the classification cutoffs are not sensitive to the precise decision boundary or that an independent verification (e.g., via activation patching on the Delta gate itself) confirms the sets before ablation.
minor comments (2)
  1. [Results tables/figures] Table or figure reporting head percentages across scales: clarify whether the 5% / 27-35% figures are averaged or reported per model, and include variance or exact counts for reproducibility.
  2. [Methods] Notation for 'single-bucket' vs. 'multi-class' probes: define the exact aggregation rule (e.g., cosine threshold or clustering) in a dedicated methods subsection to avoid ambiguity when readers attempt replication.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive comments. The points raised highlight important nuances in interpreting ablations under shared components and in validating post-hoc classifications. We respond to each major comment below, indicating where revisions will be made.

read point-by-point responses

  1. Referee: [Ablation experiments (methods and §4)] Ablation experiments (methods and §4): because Delta is head-shared, the procedure of zeroing or masking the dual-head set may still perturb the shared projection matrix for the remaining BOS-specialist heads (and vice versa). The random channel-bucketing control addresses substrate granularity but does not isolate this interaction effect, so the reported difference in causal effect sizes between the two sets could partly reflect confounding through the shared component rather than intrinsic functional separation. This is load-bearing for the claim that the sets are functionally distinct.

    Authors: We agree that the head-shared Delta projection introduces a potential interaction: ablating the dual-head set can affect the projection matrix seen by the BOS-specialist heads. The random channel-bucketing control was intended only to rule out that distinctions arise from arbitrary granularity in channel selection rather than functional grouping. We acknowledge that it does not fully isolate the shared-component interaction. However, the size-matched random controls (which also operate on subsets of the shared projection) do not reproduce the performance collapse observed when ablating the specialist set, providing some evidence that the grouping carries functional information beyond the shared matrix alone. We will revise the methods and §4 to explicitly discuss this limitation and to qualify the causal claims as showing a difference in effect size that survives granularity controls but remains subject to possible shared-projection confounding. revision: yes

  2. Referee: [Head classification and probe aggregation (§3)] Head classification and probe aggregation (§3): the post-hoc assignment of heads to BOS-specialist vs. dual sets is downstream of the same probe outputs used to define the representational signature. While the paper invokes a control, it remains necessary to show that the classification cutoffs are not sensitive to the precise decision boundary or that an independent verification (e.g., via activation patching on the Delta gate itself) confirms the sets before ablation.

    Authors: The sets are indeed defined from the probe outputs, with the multi-class aggregation used to recover the full representational signature before testing functional equivalence via ablation. The random bucketing control already addresses one form of arbitrariness in grouping. To address cutoff sensitivity, we will add supplementary material showing that varying the classification threshold over a plausible range produces head sets whose ablation effects remain qualitatively consistent. Independent verification via activation patching on the Delta gate itself is a valuable suggestion; performing it would require new experiments that lie outside the current revision. We will therefore note this as a limitation and a direction for future work rather than claiming the current classification is fully independently verified. revision: partial

Circularity Check

0 steps flagged

No significant circularity; claims rest on empirical ablations and controls

full rationale

The paper's central distinction between detection and execution layers is established through probe-based head classification followed by independent ablation experiments and a random channel-bucketing control. No equations, fitted parameters, or predictions reduce to their inputs by construction. No self-citations are load-bearing for the core result, and the methodology does not rely on self-definitional mappings or ansatzes smuggled via prior work. The derivation chain is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The work is primarily empirical and relies on standard assumptions in interpretability research rather than new mathematical axioms or invented entities.

free parameters (1)
  • Head classification cutoffs
    Percentages such as 5% BOS-specialist and 27-35% dual heads are observed from probe results and may depend on specific selection thresholds.
axioms (1)
  • domain assumption Causal ablation of model components reveals their functional role in computation
    Used to distinguish execution heads from detection heads.

pith-pipeline@v0.9.1-grok · 5813 in / 1264 out tokens · 33647 ms · 2026-06-28T18:28:46.931017+00:00 · methodology

0 comments
read the original abstract

Mechanistic interpretability often assumes that probes identifying a representational signature also identify the circuit executing the corresponding computation. We show that this assumption can fail systematically in Mamba-2. Studying the state sink (disproportionate Delta-gate activation on boundary tokens, analogous to the attention sink), we find that single-bucket probes recover only a small execution layer while missing a much larger detection layer with the same representational signature. In Mamba-2, the state sink decomposes into two functional head sets. Single-bucket BOS-specialist heads (about 5% of heads at 2.7B) causally support both BOS-context and newline-target predictions across model scales and corpora. Dual heads (27-35% of heads, recovered by multi-class aggregation of the same probe) show stronger BOS-newline representational similarity but substantially weaker causal effects under ablation. Representational similarity does not imply functional equivalence. This distinction matters for downstream behaviour: ablating BOS-specialist heads collapses RULER NIAH retrieval accuracy from 1.00 to 0.00 at 1024 context length in both Mamba-1 2.8B and Mamba-2 2.7B, while size-matched complements preserve baseline performance. A random channel-bucketing control rules out substrate granularity alone, implicating Mamba-2's head-shared Delta projection. Probe-derived specialty can identify execution circuits; at coarse granularity the same probe also recovers detection circuits, and separating them requires class-conditional ablation rather than class-conditional cosine.

Figures

Figures reproduced from arXiv: 2606.00930 by Yuhang Jiang.

Figure 1
Figure 1. Figure 1: Granularity-conditional selectivity on gate_zero ablation of bos-specialist sets. Size-matched specialist– complement ∆NLL on first-eight-tokens-after-BOS for (a) Mamba-1 channels, (b) Mamba-2 heads (30-seed random-complement bank), and (c) Pythia attention heads. Bars: wikitext-2 (dark), Pile-10k (paler). Labels in nats. to BOS or other initial positions; we test the anal￾ogous label-to-locus move in sele… view at source ↗
Figure 2
Figure 2. Figure 2: F1 functional decomposition across three M-2 scales: spec [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Cross-method phenomenology evidence. (a) LongMamba– [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 3 canonical work pages · 2 internal anchors

  1. [1]

    Conference on Language Modeling (COLM) , year=

    Mamba: Linear-Time Sequence Modeling with Selective State Spaces , author=. Conference on Language Modeling (COLM) , year=

  2. [2]

    Transformers are

    Dao, Tri and Gu, Albert , booktitle=. Transformers are

  3. [3]

    International Conference on Learning Representations (ICLR) , year=

    Efficiently Modeling Long Sequences with Structured State Spaces , author=. International Conference on Learning Representations (ICLR) , year=

  4. [4]

    Zamba: A Compact 7B

    Glorioso, Paolo and Anthony, Quentin and Tokpanov, Yury and Whittington, James and Pilault, Jonathan and Ibrahim, Adam and Millidge, Beren , journal=. Zamba: A Compact 7B

  5. [5]

    International Conference on Learning Representations (ICLR) , year=

    Efficient Streaming Language Models with Attention Sinks , author=. International Conference on Learning Representations (ICLR) , year=

  6. [6]

    International Conference on Learning Representations (ICLR) , year=

    When Attention Sink Emerges in Language Models: An Empirical View , author=. International Conference on Learning Representations (ICLR) , year=

  7. [7]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL) , pages=

    Spectral Filters, Dark Signals, and Attention Sinks , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL) , pages=

  8. [8]

    International Conference on Learning Representations (ICLR) , year=

    Quamba: A Post-Training Quantization Recipe for Selective State Space Models , author=. International Conference on Learning Representations (ICLR) , year=

  9. [9]

    Ye, Zhifan and Xia, Kejing and Fu, Yonggan and Dong, Xin and Hong, Jihoon and Yuan, Xiangchi and Diao, Shizhe and Kautz, Jan and Molchanov, Pavlo and Lin, Yingyan , booktitle=

  10. [10]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  11. [11]

    Yin, Qingyu and He, Xuzheng and Zhuang, Xiang and Zhao, Yu and Yao, Jianhua and Shen, Xiaoyu and Zhang, Qiang , booktitle=

  12. [12]

    Computational Linguistics , volume=

    Probing Classifiers: Promises, Shortcomings, and Advances , author=. Computational Linguistics , volume=

  13. [13]

    Empirical Methods in Natural Language Processing (EMNLP) , year=

    Designing and Interpreting Probes with Control Tasks , author=. Empirical Methods in Natural Language Processing (EMNLP) , year=

  14. [14]

    Tenney, Ian and Das, Dipanjan and Pavlick, Ellie , booktitle=

  15. [15]

    Empirical Methods in Natural Language Processing (EMNLP) , year=

    Information-Theoretic Probing with Minimum Description Length , author=. Empirical Methods in Natural Language Processing (EMNLP) , year=

  16. [16]

    Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics (EACL) , pages=

    Probing the Probing Paradigm: Does Probing Accuracy Entail Task Relevance? , author=. Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics (EACL) , pages=

  17. [17]

    Interpretability in the Wild: A Circuit for Indirect Object Identification in

    Wang, Kevin Ro and Variengien, Alexandre and Conmy, Arthur and Shlegeris, Buck and Steinhardt, Jacob , booktitle=. Interpretability in the Wild: A Circuit for Indirect Object Identification in

  18. [18]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Towards Automated Circuit Discovery for Mechanistic Interpretability , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  19. [19]

    How to use and interpret activation patching

    How to Use and Interpret Activation Patching , author=. arXiv preprint arXiv:2404.15255 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  20. [20]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Investigating Gender Bias in Language Models Using Causal Mediation Analysis , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  21. [21]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Causal Abstractions of Neural Networks , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  22. [22]

    Transformer Circuits Thread , year=

    In-context Learning and Induction Heads , author=. Transformer Circuits Thread , year=

  23. [23]

    Locating and Editing Factual Associations in

    Meng, Kevin and Bau, David and Andonian, Alex and Belinkov, Yonatan , booktitle=. Locating and Editing Factual Associations in

  24. [24]

    Scaling Monosemanticity: Extracting Interpretable Features from

    Templeton, Adly and Conerly, Tom and Marcus, Jonathan and Lindsey, Jack and Bricken, Trenton and Chen, Brian and Pearce, Adam and Citro, Craig and Ameisen, Emmanuel and Jones, Andy and Cunningham, Hoagy and Turner, Nicholas L and McDougall, Callum and MacDiarmid, Monte and Tamkin, Alex and Durmus, Esin and Hume, Tristan and Mosconi, Francesco and Freeman,...

  25. [25]

    International Conference on Learning Representations (ICLR) , year=

    Sparse Autoencoders Find Highly Interpretable Features in Language Models , author=. International Conference on Learning Representations (ICLR) , year=

  26. [26]

    International Conference on Learning Representations (ICLR) , year=

    Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models , author=. International Conference on Learning Representations (ICLR) , year=

  27. [27]

    The Hidden Attention of

    Ali, Ameen and Zimerman, Itamar and Wolf, Lior , booktitle=. The Hidden Attention of. 2025 , doi=

    2025

  28. [28]

    Do Transformer Interpretability Methods Transfer to

    Paulo, Gon. Do Transformer Interpretability Methods Transfer to. Proceedings of the AAAI Conference on Artificial Intelligence , volume=. 2025 , doi=

    2025

  29. [29]

    Locating and Editing Factual Associations in

    Sharma, Arnab Sen and Atkinson, David and Bau, David , booktitle=. Locating and Editing Factual Associations in. 2024 , note=

    2024

  30. [30]

    Investigating the Indirect Object Identification circuit in Mamba

    Ensign, Danielle and Garriga-Alonso, Adri. Investigating the Indirect Object Identification Circuit in. arXiv preprint arXiv:2407.14008 , year=

    work page Pith review arXiv

  31. [31]

    Biderman, Stella and Schoelkopf, Hailey and Anthony, Quentin Gregory and Bradley, Herbie and O'Brien, Kyle and Hallahan, Eric and Khan, Mohammad Aflah and Purohit, Shivanshu and Prashanth, USVSN Sai and Raff, Edward and Skowron, Aviya and Sutawika, Lintang and van der Wal, Oskar , booktitle=

  32. [32]

    International Conference on Learning Representations (ICLR) , year=

    Pointer Sentinel Mixture Models , author=. International Conference on Learning Representations (ICLR) , year=

  33. [33]

    Gao, Leo and Biderman, Stella and Black, Sid and Golding, Laurence and Hoppe, Travis and Foster, Charles and Phang, Jason and He, Horace and Thite, Anish and Nabeshima, Noa and Presser, Shawn and Leahy, Connor , journal=. The

  34. [34]

    2024 , note=

    Hsieh, Cheng-Ping and Sun, Simeng and Kriman, Samuel and Acharya, Shantanu and Rekesh, Dima and Jia, Fei and Zhang, Yang and Ginsburg, Boris , booktitle=. 2024 , note=

    2024

  35. [35]

    Needle In A Haystack — Pressure Testing

    Kamradt, Greg , year=. Needle In A Haystack — Pressure Testing

  36. [36]

    Conference on Language Modeling (COLM) , year=

    Massive Activations in Large Language Models , author=. Conference on Language Modeling (COLM) , year=

  37. [37]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Quantizable Transformers: Removing Outliers by Helping Attention Heads Do Nothing , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  38. [38]

    International Conference on Learning Representations (ICLR) , year=

    Vision Transformers Need Registers , author=. International Conference on Learning Representations (ICLR) , year=

  39. [39]

    International Conference on Machine Learning (ICML) , year=

    Repeat After Me: Transformers are Better than State Space Models at Copying , author=. International Conference on Machine Learning (ICML) , year=

  40. [40]

    An Empirical Study of Mamba-based Language Models

    An Empirical Study of Mamba-based Language Models , author=. arXiv preprint arXiv:2406.07887 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  41. [41]

    International Conference on Learning Representations (ICLR) , year=

    Retrieval Head Mechanistically Explains Long-Context Factuality , author=. International Conference on Learning Representations (ICLR) , year=

  42. [42]

    Amnesic Probing: Behavioral Explanation with Amnesic Counterfactuals , author=. Trans. Assoc. Comput. Linguist. , volume=

  43. [43]

    Belrose, Nora and Schneider-Joseph, David and Ravfogel, Shauli and Cotterell, Ryan and Raff, Edward and Biderman, Stella , booktitle=

  44. [44]

    The Annals of Statistics , volume=

    The Dip Test of Unimodality , author=. The Annals of Statistics , volume=

  45. [45]

    Journal of the Royal Statistical Society: Series B (Methodological) , volume=

    Using Kernel Density Estimates to Investigate Multimodality , author=. Journal of the Royal Statistical Society: Series B (Methodological) , volume=