Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

This paper claims that rewriting the Kalman filter in information form makes its nonlinear per-token updates parallelizable, giving a sequence mixer with explicit uncertainty that outperforms linear state-space and gated-linear-attention la

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 →

KLA reparameterises information-form Kalman filtering as an associative Möbius scan, giving a parallelisable nonlinear state-space mixer that tracks A5 permutations with one or two layers.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection The Möbius-scan theory is sound and the primitive is genuinely interesting, but the A5 evidence as presented is too weak to carry the expressivity claim, and the abstract oversells what the body delivers. the 4 major comments →

arxiv 2602.10743 v2 pith:C6YVXQIK submitted 2026-02-11 cs.LG

Kalman Linear Attention: Parallel Bayesian Filtering For Efficient Language Modelling and State Tracking

classification cs.LG MSC 62M2068T50
keywords Kalman filterinformation formMöbius transformparallel prefix scanlinear attentionBayesian filteringstate trackinglanguage modelling
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 reading

The paper asks whether sequence mixing in language models can be exact Bayesian filtering rather than a linear recurrence. It answers yes: when the Kalman filter is written in information form, the precision update is a fractional-linear (Möbius) map, and fractional-linear maps compose by 2x2 matrix multiplication. That makes the recurrent update nonlinear yet perfectly parallel via a prefix scan, at the same asymptotic cost as current linear-time mixers. The resulting layer carries an explicit belief-state uncertainty that acts as a learned, history-dependent gate, and it solves permutation-composition state-tracking problems that linear recurrences and attention cannot. A sympathetic reader would care because this is a route out of the linear-update expressivity ceiling without giving up parallel training.

Core claim

The central claim is that Kalman filtering can serve as a drop-in sequence-mixing primitive for language modelling. In a diagonal linear-Gaussian state-space model, the posterior precision evolves by a Möbius transformation, so the sequence of precision updates is equivalent to a chain of 2x2 matrices under multiplication; the posterior information mean evolves affinely with a forget gate set by the precision ratio. These two recursions can be computed by associative parallel scans in O(T) work and O(log T) depth. The nonlinearity—the fractional-linear denominator—is exactly what makes the gate history-dependent and strictly more expressive than linear/affine gating, while the explicit preci

What carries the argument

The Möbius (fractional-linear) transform of the precision recursion: λ_t = M_t(λ_{t−1}) = (α_t λ_{t−1} + β_t)/(γ_t λ_{t−1} + δ_t), with M_t = [[1 + p_t ϕ_t, a_t^2 ϕ_t],[p_t, a_t^2]], where ϕ_t is token-evidence precision weighted by the observation operator. Because composing Möbius transforms is 2x2 matrix multiplication, the entire precision path is a parallel prefix scan; the affine information-mean recursion runs as a second scan. This object carries the argument: it is the point where nonlinear Bayesian filtering meets associative scan parallelism and where the uncertainty-ratio gate emerges.

Load-bearing premise

The argument assumes a diagonal, per-coordinate state model: every feature channel runs an independent scalar Kalman filter, so any advantage rests on tasks where the needed information lives within coordinates rather than in correlations between them.

What would settle it

Train a single-layer KLA on a task whose answer requires combining two different state coordinates, such as adding two numbers encoded in separate channels or a selective-copy variant where the token to copy is the XOR of two channel values. If accuracy stays at parity with a full-covariance Bayesian filter, the diagonal assumption is benign; if it collapses while a transformer solves it, the central expressivity claim is bounded to per-coordinate state tracking.

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

If this is right

  • KLA is a drop-in mixer with O(T) work and O(log T) depth, matching the training profile of linear-time state-space and gated-attention layers.
  • The precision-ratio forget gate is history-dependent and nonlinear, so the model can selectively down-weight new evidence as accumulated confidence grows—mechanistically similar to gating but derived from inference.
  • KLA solves the A5 permutation-composition task with 1-2 layers, while linear state-space models and transformers need depth that grows with sequence length.
  • On long-context associative recall at sequence length 2048, KLA exceeds baseline accuracy and reaches near-perfect performance at model dimension 256.
  • Uncertainty-weighted updates reduce state saturation under heavy key-value load, and Monte Carlo decoding under the learned posterior gives further gains on noise-corrupted synthetic tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial: since all recursions are per-coordinate, the layer is effectively a bank of independent scalar filters; tasks that require binding or correlating across feature coordinates would likely need a full-covariance extension, which this paper does not provide.
  • Editorial: the explicit posterior variance is a naturally interpretable uncertainty signal that could serve hallucination detection, abstention, or out-of-distribution prompting; the authors note these as untested future work.
  • Editorial: the 2x2-matrix Möbius associativity is a general transferable mechanism—any nested fractional-linear recursion, such as time-varying or controlled dynamics or Bayesian smoothing, could inherit the same parallel-scan trick.
  • Editorial: a clean empirical test of whether the nonlinearity matters is to fix the precision path to an affine approximation and measure A5 accuracy; if the result persists, the extra expressivity has a different source than the fractional-linear denominator.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 6 minor

Summary. The paper introduces Kalman Linear Attention (KLA), a sequence-mixing layer that reparameterises a diagonal linear-Gaussian Kalman filter in information form. The central theoretical claim is that the posterior precision recursion is a Möbius (fractional-linear) map and the information-mean recursion is affine, so both update families compose associatively and can be evaluated with parallel prefix scans in O(T) work and O(log T) depth, matching Mamba/GLA-style efficiency while retaining a nonlinear, uncertainty-driven gating mechanism. The paper also claims empirical gains on synthetic language-modelling tasks (MAD suite), long-context associative recall (MQAR), and the A5 permutation-composition state-tracking benchmark, and states in the abstract that KLA is among the first stacked Bayesian-filtering primitives trained at billion-token scale.

Significance. If the central claims hold, KLA is a conceptually appealing and practically relevant primitive: it offers a nonlinear recurrence that remains scan-parallel, explicit belief-state uncertainty, and a principled probabilistic interpretation of gating. The theoretical core (Theorems 1 and 2, Corollaries 1 and 2) is self-contained and appears correct: the derivation of the Möbius precision update from standard information-filter equations is clean, and the scan parallelisation follows directly from associativity of matrix multiplication. The paper also provides a useful comparison with Särkkä and García-Fernández's lifting construction, showing that no 5-tuple augmentation is needed in this case. However, the empirical evidence for the headline expressivity claim—that KLA 'solves' A5 tasks that linear SSMs and attention cannot—is currently weak, and the manuscript's own limitations section disclaims the web-scale claims made in the abstract. The central theoretical contribution is sound, but the empirical demonstration supporting 'strictly more expressive' requires substantial strengthening before the paper's claims are fully credible.

major comments (4)
  1. [§5.5, Fig. 1, Appendix G.5] The A5 state-tracking result is the only direct evidence for the paper's central expressivity claim, but as reported it cannot be independently verified. Success is defined as '≥90% accuracy in at least one of five seeds' (Appendix G.5), a criterion that can be satisfied by a single lucky run and is not a standard measure of reliable performance. Figure 1 has no numerical table, no error bars, and no per-seed results. Baselines are taken from Merrill et al. (2024) rather than trained under the same protocol, parameter counts, data, and early-stopping. Given that the entire 'strictly more expressive' claim rests on this experiment, a matched re-evaluation under identical training conditions with all five seeds reported (mean, std, min/max) is required before the claim is supportable.
  2. [Section 3 'Notational conventions'; Theorem 1 proof (Appendix D)] The derivation assumes a diagonal/per-coordinate model: 'we identify diagonal matrices with their diagonal terms and apply scalar recursions elementwise.' Consequently, the KLA layer is a collection of independent scalar filters per coordinate, and the Möbius associativity argument does not extend to cross-coordinate interactions. The A5 permutation-composition task is non-abelian and may require cross-coordinate mixing in the state update, yet the paper does not test or justify whether diagonal structure suffices. This compounds the fragility of the A5 evidence: the empirical claim is doing critical work that neither theory nor robust experiment currently supports. Please add an explicit discussion and, ideally, an ablation with non-diagonal/matrix-state updates or a concrete argument for why per-coordinate independence still captures the required composition.
  3. [Abstract vs. Section 7 (Limitations)] The abstract states that KLA is 'among the first stacked Bayesian-filtering primitives trained at the billion-token scale,' but Section 7 explicitly says 'we do not claim state-of-the-art performance on web-scale corpora' and the body contains no billion-token experiments. This is an internal inconsistency between the abstract and the manuscript's own limitations. The abstract should be tempered to match the actual empirical scope, or the billion-token experiments should be included.
  4. [§5.3, Table 3, Appendix G.2] Table 3 reports a single accuracy number per task and per method, with no error bars, despite Appendix G.1 stating results are averaged over 5 seeds. Moreover, KLA+ uses 10 Monte Carlo samples for probabilistic decoding (Appendix G.2), which is a different inference/training procedure than the baselines; comparing KLA+ against standard cross-entropy baselines may conflate the benefit of the Bayesian primitive with the benefit of marginalisation. Please report per-seed variance for all methods and clearly separate 'KLA with posterior-mean decoding' from 'KLA with Monte Carlo marginalisation', and ensure baselines receive the same decoding treatment where applicable.
minor comments (6)
  1. [§5.3] 'Section 5.3 shows that KLA performs competitively...' is a self-referential phrase; it should say 'Table 3 shows...'.
  2. [Appendix A] The text mentions QK-Norm and expansion factor 1, but this is not described in the main method or hyperparameter tables. Please add a sentence explaining where QK-Norm is applied and why.
  3. [§5.2, Fig. 5] The runtime benchmark compares KLA implementations against each other but not against Mamba/GLA on the same hardware. Since the paper claims 'same computational cost' as those models, a direct wall-clock comparison with at least one strong baseline would make the claim concrete.
  4. [References] The workshop name 'EurIPS 2025' appears in the footnote; this is likely a typo for NeurIPS. Please verify.
  5. [Appendix E.3/E.4] The 'Kalman Attention Matrix' visualisations are qualitative. A quantitative comparison with the attention matrices of GLA/Mamba on the same tasks would strengthen the claimed interpretability.
  6. [Theorem 3 / Appendix D] The convolutional form for deterministic LTI systems is derived correctly, but the statement 'computable in O(T log T)' and 'O(log T) parallel time' conflates serial and parallel complexity. Please clarify that FFT gives O(T log T) serial work and O(log T) parallel depth.

Circularity Check

0 steps flagged

No significant circularity: the Kalman/information-filter derivation is self-contained; empirical caveats exist but do not reduce to construction.

full rationale

The paper's central derivation is mathematically self-contained. Theorem 1 derives the Möbius precision recursion directly from the standard information-form Kalman update (predict: λ_prior = λ_{t-1}/(a^2 + p λ_{t-1}), update: λ_t = λ_prior + k^2 Λ^v_t), and Theorem 2 derives the affine mean update from the same equations. These are not fitted to any target result; they are algebraic consequences of the stated diagonal linear-Gaussian model. The associativity enabling parallel scans follows from the 2×2 matrix representation of Möbius transformations, an external mathematical fact. No fitted parameter is later called a prediction; the MAD, MQAR, and A5 results are empirical evaluations, not derived from the trained parameters in a way that forces the outcome. Self-citations (Shaj et al. 2021, Shaj Kumar et al. 2023) appear only in related work on deep Kalman filters and are not load-bearing: the key theorems and proofs do not rely on them. The claim of 'strictly more expressive' nonlinear updates rests on the algebraic form (nonlinear vs. linear recurrence) plus the A5 experiment; while the experiment itself has validity concerns—Appendix G.5 defines success as ≥90% accuracy in at least one of five seeds, and baselines are copied from Merrill et al. rather than re-run under the same protocol—this is an empirical robustness issue, not a construction-level circularity. The diagonal (per-coordinate) assumption is explicitly stated and limits the theory's generality, but again does not make the derivation circular. Neither the abstract's unsupported 'billion-token scale' claim nor the discrepancy with Section 7's disclaimer are circularity; they are consistency/support issues. Overall, the derivation chain is not equivalent to its inputs, and no step reduces to its own conclusion.

Axiom & Free-Parameter Ledger

6 free parameters · 7 axioms · 0 invented entities

The theoretical result is self-contained: it uses only standard Gaussian filtering, Möbius composition, and prefix-scan facts. The empirical claims rest on learned parameters (p, a, Δ, Λ^v, projections), on the diagonal-state assumption, and on taking the A5 benchmark as the expressivity testbed. No new physical or ontological entities are introduced.

free parameters (6)
  • process_noise_scale_p = learned; initial 0.01
    Learnable OU process-noise scale in Eq. (8). It is what makes the precision update Möbius rather than linear; the ablation (Fig. 10) shows p=0 degrades performance by ~49.8 points, so the central claim depends on it.
  • OU_decay_a = learned, time-invariant
    Controls the forget/decay rate and, via Eq. (8), the process noise; unlike Mamba's token-dependent a, KLA's a is fixed across tokens.
  • discretization_timestep_delta = learned in [0.001, 0.1]
    Scales the OU discretisation; range chosen by hand (Appendix G.2).
  • value_precision_lambda_v = learned from input
    Token-dependent precision Lambda^v_t is the confidence assigned to each observation; it is a learned function of the input and is central to the uncertainty gating behaviour.
  • observation/readout_operators_k_q = learned from input
    Keys and queries are linear projections of the input; they define the observation geometry and readout, analogous to GLA.
  • state_expansion_N = 8 (MAD) / 16 (MQAR, A5)
    Design choice for per-channel memory slots; matched across baselines as 'effective state size' but the matching is approximate (GDN gets slightly larger state).
axioms (7)
  • standard math Standard Kalman/information-filter recursions for linear-Gaussian state-space models.
    Starting point for Theorems 1-2; cited to Anderson & Moore and Khan.
  • standard math Möbius transformations compose through 2x2 matrix multiplication, and matrix multiplication is associative.
    Basis of Corollaries 1-2; no proof needed.
  • standard math Parallel prefix (associative) scan has O(T) work and O(log T) depth.
    Blelloch's prefix-sum result; used for all scan claims.
  • domain assumption Diagonal parameterisation is sufficient for language-model sequence mixing.
    All recursions are elementwise scalar; no cross-dim coupling. The paper does not justify this for language data (Section 3 Notation; Theorem 1 proof).
  • domain assumption OU process is an appropriate continuous-time prior for the latent state.
    Chosen for mean reversion and exact discretisation (Eq. (8)); only ablation is p=0, not alternative priors.
  • domain assumption A5 permutation composition is a valid proxy for sequential state tracking and distinguishes TC0 from NC1.
    Takes Merrill et al. (2024) theoretical framework at face value; baselines are not re-run.
  • domain assumption Observations vt are generated as noisy linear measurements of the latent state; training uses discriminative cross-entropy, so the generative interpretation is metaphorical.
    The likelihood model in Eq. (9) underpins the filtering semantics but is not tested as a generative model.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Kalman Linear Attention: Parallel Bayesian Filtering For Efficient Language Modelling and State Tracking." pith.science (2026). https://pith.science/paper/C6YVXQIK

@misc{pith2026260210743,
  author       = {Pith},
  title        = {Pith review of: Kalman Linear Attention: Parallel Bayesian Filtering For Efficient Language Modelling and State Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C6YVXQIK}},
  note         = {Machine review of arXiv:2602.10743}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

State-space language models such as Mamba and gated linear attention (GLA) offer linear-complexity, parallelisable alternatives to transformers, but their linear state updates limit expressivity and robust state tracking. We close this gap from a probabilistic angle, casting sequence mixing as exact Bayesian filtering with the Kalman filter as the core primitive. Classical Kalman filters give principled state and uncertainty estimates but are viewed as inherently sequential; we show that reparameterising them in information form turns their updates into an associative scan - so the per-token recurrent update is non-linear (a M\"obius/precision recursion) yet remains temporally parallel. The resulting Kalman Linear Attention (KLA) layer is a drop-in sequence mixer that performs time-parallel probabilistic inference, carries an explicit belief-state uncertainty, and is strictly more expressive than GLA-style linear updates at the same computational cost. This expressivity translates directly into stronger state tracking: KLA solves permutation-composition ($A_5$) tasks that linear SSMs and attention cannot, while staying scan-parallel. As a drop-in primitive it also matches or improves on modern SSMs and GLAs across synthetic token-manipulation and zero-shot commonsense benchmarks, and is among the first stacked Bayesian-filtering primitives trained at the billion-token scale.

Figures

Figures reproduced from arXiv: 2602.10743 by Aidan Scannell, Amos Storkey, Andras Szecsenyi, Cameron Barker, Elliot J. Crowley, Vaisakh Shaj.

Figure 1
Figure 1. Figure 1: Minimum number of layers required to solve the A5 (alter￾nating group on 5 elements) permutation composition task (Merrill et al., 2024). KLA’s fractional linear updates fall between a fully nonlinear RNN and linear SSMs/transformers, requiring only 1 or 2 layers to solve the task without loss of parallelism. successors, achieve this with O(log T) depth, O(T) work, and linear or sublinear memory. This is e… view at source ↗
Figure 2
Figure 2. Figure 2: From OU dynamics to parallel inference. (Top) Continuous-time OU prior. (Middle) Discrete linear-Gaussian SSM. (Bottom) Möbius scan for parallel posterior state estimation. 4. Method We introduce Kalman Linear Attention (KLA) as a probabilis￾tic sequence mixer. Instead of deterministically updating a hidden state as in modern SSM mixers, KLA maintains a belief state over a latent representation, consisting… view at source ↗
Figure 3
Figure 3. Figure 3: Block architecture. The block follows the fused-MLP design of Mamba, with the Kalman Filter as a drop-in replacement for any SSM/Attention primitive. fractional (Möbius) transform: λt = Mt(λt−1) = αt ⊙ λt−1 + βt γt ⊙ λt−1 + δt , (16) Mt =  αt βt γt δt  =  1 + ¯pt ⊙ ϕt ¯a2 t ⊙ ϕt ¯pt ¯a2 t  . (17) Interpretation: precision track as uncertainty-driven gat￾ing. While Equation (16) is nonlinear in λt−1, it… view at source ↗
Figure 4
Figure 4. Figure 4: Ablation of OU prior dynamics and discretisation (Section 4.1) on Selective Copy (T = 256). OU discretisation improves accuracy and learning stability, especially for deeper models. Special cases: Under deterministic (pt = 0) and linear time￾invariant (LTI) settings, the KLA updates reduce to convolu￾tions computable in O(log T) time via FFT (see Theorem 3 in the Appendix). 4.3. KLA as a Drop-in Probabilis… view at source ↗
Figure 5
Figure 5. Figure 5: Training-time runtime scaling. Wall-clock runtime of KLA implementations across sequence lengths. Torch Scan uses torch._higher_order_ops.associative_scan; Triton Scan uses custom forward/backward kernels. additional speedups unlocked by hardware-level optimisation beyond the pure mathematical reparameterisation. While our Triton kernel already narrows the gap between theory and throughput, additional hard… view at source ↗
Figure 7
Figure 7. Figure 7: reveals distinct scaling behaviours across archi￾tectures. KLA consistently outperforms Mamba across all dimensions and substantially outperforms GLA, which fails to learn the task under this extreme setting (T = 2048, V = 256). At lower dimensions (d = 64, d = 128), GDN performs well; its delta-rule mechanism is specifically de￾signed for strong associative recall at limited capacity (Yang et al., 2024), … view at source ↗
Figure 6
Figure 6. Figure 6: Posterior variance as uncertainty-driven gating. Selec￾tive Copy task. Top: per-channel variance across state dimensions. Middle: average variance over channels. Bottom: task structure with green copy tokens. Variance dynamics align with task-relevant tokens, providing an interpretable signal of filtering/selection. 5.4. Long-Context Associative Recall Multi-Query Associative Recall (MQAR) (Arora et al., 2… view at source ↗
Figure 5
Figure 5. Figure 5: Model Architecture Diagram [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 8
Figure 8. Figure 8: Model architecture used in our experiments. The KLA block is a drop-in replacement for Transformer or SSM blocks. We follow a similar block structure to Mamba, fusing the sequence mixing and gated MLP into a single block. We additionally employ QK-Norm and use an expansion factor of 1. Tokens are embedded and passed through a sequence mixer, then decoded to logits. The sequence mixer is a drop-in primitive… view at source ↗
Figure 9
Figure 9. Figure 9: Two views of state-space sequence models. Left: Bayesian/observation view—tokens are measurements, inference recovers hidden state. Right: Control view (Mamba)—tokens are inputs that deterministically update state. C.4. Kalman and Information Filters C.4.1. BAYESIAN FILTERING AS POSTERIOR INFERENCE. Bayesian filters (Kalman, information-filter variants) are inference schemes—procedures that compute posteri… view at source ↗
Figure 10
Figure 10. Figure 10: compares KLA with learnable process noise against a variant where process noise is fixed to zero (deterministic dynamics, but retaining observation variance (Λv ) −1 ). The results show that removing process noise leads to significantly worse performance across all MADLAB tasks, with accuracy degrading by an average of 49.8 percentage points. The degradation is particularly severe on memorisation (93.6% d… view at source ↗
Figure 11
Figure 11. Figure 11: Forward pass (prompt processing) runtime scaling. Wall-clock runtime of KLA implementations across sequence lengths during forward-only pass. Torch Scan uses torch._higher_order_ops.associative_scan; Triton Scan uses custom forward kernels. Conv1D Linear Linear Linear Linear Kalman Filter silu silu W = ×   k1Λ v 1 0 · · · 0 f2 k1Λ v 1 k2Λ v 2 · · · 0 . . . . . . . . . . . . QT s=2 fs k1Λ v 1 QT s=3… view at source ↗
Figure 12
Figure 12. Figure 12: Block architecture and equivalent attention form. Left: The block architecture follows the fused-MLP design of Mamba, with the Kalman Filter as a drop-in replacement for any SSM/Attention primitive. Right: Unrolling the information-mean recurrence (Equation (19): ηt = ft ηt−1 + kt Λ v t vt) yields a lower-triangular matrix W whose entries are products of history-dependent forget gates fs, keys kj , and ob… view at source ↗
Figure 13
Figure 13. Figure 13: Attention maps for Selective Copying (sequence length 256). The model learns sparse, intermittent vertical bands, attending strongly to a small set of task-relevant positions (mostly copy positions). They either activate or suppress these relevant positions. Differences across channels indicate channel-level specialisation in which positions are retrieved [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Attention maps for In-Context Recall (sequence length 128). The patterns are sparse and “pointer-like”: within each channel, attention concentrates on a small set of source positions (vertical bands) across many target steps, with a mostly low-activation background. Channels show specialisation trends. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Attention maps for Memorization (sequence length 32). The smoother, more gradual decay reflects the task’s requirement to maintain information uniformly across the sequence. All channels show similar patterns, indicating that memorization benefits from redundant, distributed storage rather than selective attention. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_15.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Memory by Design: Probabilistic Sequence Layers

    stat.ML 2026-05 unverdicted novelty 6.0

    The design-model framework unifies sub-quadratic sequence models as Bayesian filters and introduces a covariance-tracking Bayesian Layer that improves retrieval robustness beyond training regimes on MQAR and RULER benchmarks.

  2. Memory for Large Language Models

    cs.CL 2026-07 conditional novelty 4.0

    The paper organizes LLM memory architectures along three axes—implicit vs explicit representation, offline vs online update, and short vs long-term persistence—plus finer-grained update rules.

Reference graph

Works this paper leans on

17 extracted references · 10 linked inside Pith · cited by 2 Pith papers

  1. [3]

    More extensive exploration of scaffolding configurations optimally suited to the KLA filtering mixer can be a direction for future work

    and is set to 16 (N or sometimes called d_state in SSM codebases) as default similar to (Gu & Dao, 2023). More extensive exploration of scaffolding configurations optimally suited to the KLA filtering mixer can be a direction for future work. We implemented KLA with PyTorch (Paszke et al.,

  2. [8]

    and Garcia, V

    Nielsen, F. and Garcia, V . Statistical exponential families: A digest with flash cards.arXiv preprint arXiv:0911.4863,

  3. [9]

    Gated linear attention transformers with hardware-efficient train- ing.arXiv preprint arXiv:2312.06635,

    Yang, S., Wang, B., Shen, Y ., Panda, R., and Kim, Y . Gated linear attention transformers with hardware-efficient train- ing.arXiv preprint arXiv:2312.06635,

  4. [11]

    and Erichson, N

    Yu, A. and Erichson, N. B. Block-biased mamba for long-range sequence processing.arXiv preprint arXiv:2505.09022,

  5. [13]

    Each transform is represented (up to multiplication by a nonzero scalar) by a 2×2 matrix M= a b c d and acts on a scalar z via the linear–fractional rule M(z) := az+b cz+d

    It generalises the familiar affine form (c= 0 gives az+b ) by introducing an additional denominator term. Each transform is represented (up to multiplication by a nonzero scalar) by a 2×2 matrix M= a b c d and acts on a scalar z via the linear–fractional rule M(z) := az+b cz+d . Composing multiple transforms amounts to multiplying their representing matri...

  6. [14]

    pointer-like

    Figure 10.Process noise ablation on MADLAB tasks.Comparison of KLA with learnable process noise versus a variant with process noise fixed to zero (pt = 0). Removing process noise leads to severe performance degradation, with an average accuracy drop of 49.8 percentage points. E.2. Runtime Scaling Setup.We benchmark forward-only runtimes for three KLA impl...

  7. [16]

    Model # Blocksd model dstate OtherS KLA (Ours) 1 128 8 expand=1 2048 Mamba 1 128 16 expand=1 2048 GatedDeltaNet 1 128 -e k=0.5,e v=1.0,H=4 2048 GLA 1 128 -e k=0.5,e v=1.0,H=4 2048 mLSTM 1 128 -H=16, proj_factor=2.0 2048 Batch size:We use a batch size of 172 for all MAD-Lab experiments. G.4. MQAR Hyperparameters For the MQAR (hard) experiments, we evaluate...

  8. [17]

    A5 State Tracking Hyperparameters For the A5 state tracking experiments (Section 5.5), we use dmodel = 1024 and dstate = 16 for KLA

    KLA 2 256 16 expand=1 8,192 Mamba 2 256 16 expand=2 8,192 GatedDeltaNet 2 256 -e k=0.75,e v=1.5,H=8 9,216 GLA 2 256 -e k=0.25,e v=1.0,H=4 8,192 G.5. A5 State Tracking Hyperparameters For the A5 state tracking experiments (Section 5.5), we use dmodel = 1024 and dstate = 16 for KLA. We train for up to 500 epochs with early stopping (patience 50), and a lear...

  9. [1957]

    and Gu, A

    10 Parallel, Scalable and Efficient Bayesian Filters for Language Modelling Dao, T. and Gu, A. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality.arXiv preprint arXiv:2405.21060,

  10. [2005]

    Zoology: Measuring and improving recall in efficient language models.arXiv preprint arXiv:2312.04927,

    Arora, S., Eyuboglu, S., Timalsina, A., Johnson, I., Poli, M., Zou, J., Rudra, A., and Ré, C. Zoology: Measuring and improving recall in efficient language models.arXiv preprint arXiv:2312.04927,

  11. [2015]

    G., Shalit, U., and Sontag, D

    Krishnan, R. G., Shalit, U., and Sontag, D. Deep Kalman filters.arXiv preprint arXiv:1511.05121,

  12. [2016]

    and Dao, T

    Gu, A. and Dao, T. Mamba: Linear-time sequence modeling with selective state spaces.arXiv preprint arXiv:2312.00752,

  13. [2017]

    E., Bottero, A

    Luis, C. E., Bottero, A. G., Vinogradska, J., Berkenkamp, F., and Peters, J. Uncertainty representations in state- space layers for deep reinforcement learning under partial observability.arXiv preprint arXiv:2409.16824,

  14. [2019]

    KalMamba: Towards efficient probabilistic state space models for RL under uncertainty

    Becker, P., Freymuth, N., and Neumann, G. KalMamba: Towards efficient probabilistic state space models for RL under uncertainty. InICML 2024 Workshop: Aligning Reinforcement Learning Experimentalists and Theorists,

  15. [2023]

    Gated delta net- works: Improving Mamba2 with delta rule.arXiv preprint arXiv:2412.06464,

    Yang, S., Kautz, J., and Hatamizadeh, A. Gated delta net- works: Improving Mamba2 with delta rule.arXiv preprint arXiv:2412.06464,

  16. [2024]

    The illusion of state in state-space models.arXiv preprint arXiv:2404.08819,

    Merrill, W., Petty, J., and Sabharwal, A. The illusion of state in state-space models.arXiv preprint arXiv:2404.08819,

  17. [2048]

    Long-Context MQAR Table 7 lists the data configuration for the long-context MQAR benchmark task (Arora et al., 2023)

    F.2. Long-Context MQAR Table 7 lists the data configuration for the long-context MQAR benchmark task (Arora et al., 2023). Table 7.Long-Context MQAR data parameters. Long-Context Multi-Query Associative Recall (MQAR) Setting Seq. length Vocab size Training seqs Eval seqs CR (hard) 2,048 256 12,800 1,280 27 Parallel, Scalable and Efficient Bayesian Filters...

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.