Pith. sign in

REVIEW 3 major objections 5 minor 35 references

Interference and Retention in Continual Learning

T0 review · 3 major / 5 minor · reviewed 2026-07-13 · grok-4.5

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

Pith's one-line read Forgetting is interference energy between tasks; when supports are disjoint it can be removed by construction, and when they conflict a distortion floor is unavoidable.

desk verdict Clean exact identity for forgetting under frozen features, with a method that follows from it; A1 is the real hinge, not a hidden flaw. read the letter →

arxiv 2607.09202 v1 pith:O57E25DQ submitted 2026-07-10 cs.LG cs.AIcs.NE

Julius Störk This is my paper · ORCID
classification cs.LGcs.AIcs.NE
keywords continuallearningcatastrophicforgettingmodelmergingneuraltangentkernelgradientprojectionrate-distortionrepresentationgeometryinterferenceenergy
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 argues that catastrophic forgetting should be treated as interference between tasks rather than patched after the fact with replay buffers, Fisher penalties, or distillation. In the frozen-feature regime, the forgetting of an old task after a new update is exactly the interference energy of that update measured in the old task’s feature second-moment geometry. When the tasks excite disjoint feature directions, the new update can be placed entirely in the old task’s null space and forgetting is eliminated structurally; when they share conflicting directions, a positive distortion floor remains and matches the excess risk of any single shared head. The same geometry makes optimal model merging task-aware orthogonalization rather than Euclidean averaging. From this analysis the authors derive Interference-Gated Functional Allocation (IGFA), a replay-free, Fisher-free controller that shares capacity when tasks align and protects it when they conflict, relocating unavoidable cost from irreversible forgetting into deferred but recoverable plasticity.

What carries the argument

The interference functional: forgetting of A equals one-half Delta transpose Sigma_A Delta, where Sigma_A is the feature second moment of task A. This single quadratic form yields the removability dichotomy (lossless iff Delta lies in ker Sigma_A), the distortion floor under support overlap, and the optimality condition for Sigma-orthogonal model merging; path-averaged curvature recovers the same quantity under feature drift.

What would settle it

Train a linear head on frozen random features for two tasks with known, controllable support overlap: if measured forgetting of A after learning B does not equal one-half Delta transpose Sigma_A Delta to machine precision on disjoint supports and does not sit at the predicted positive floor on conflicting shared directions, the central identity fails.

Watch

Extended reading notes

Core claim

Under frozen features, realizability, and quadratic excess loss, the forgetting of task A caused by learning task B is exactly the interference energy one-half Delta transpose Sigma_A Delta. An update is lossless for A if and only if it lies in the kernel of Sigma_A; when task supports overlap on directions where the targets disagree, a non-zero distortion floor is unavoidable and equals the Bayes single-head excess risk on the shared support. Optimal merging of task solutions is mutual Sigma-orthogonalization under this same metric.

Load-bearing premise

The exact parameter-space identities assume the feature map stays fixed while the head is trained; once features drift substantially the closed-form guarantees become first-order approximations that must be tracked recursively.

Editorial extensions

If this is right

  • When task supports are disjoint, structural projection alone yields lossless retention without replay or Fisher state.
  • When supports overlap with target conflict, cost can only be relocated—from irreversible forgetting of the old task into deferred fit of the new one—not removed by any single shared head.
  • Model merging should use the task-induced Sigma metric; Euclidean or isotropic averaging leaves avoidable residual interference.
  • A similarity-gated share-or-protect rule (IGFA) recovers unconditional orthogonal projection on dissimilar streams and improves on it when transfer is worth preserving.
  • The same geometry supplies a pre-flight diagnostic: subspace overlap and target agreement decide whether structural allocation, sharing, or replay is the right tool before training begins.

Reading between the lines

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

  • The same interference ledger could serve as a controller for federated aggregation and for physics-informed constraints, treating clients or PDE residuals as additional tasks under one metric.
  • If the function-space form of the identity holds at scale, a threshold-free sign gate on measured interference rates could replace modality-specific similarity thresholds for language continual fine-tuning.
  • Capacity knees predicted by cumulative occupied rank suggest a concrete tradeoff: model width versus rehearsal budget can be set from the rate–distortion view rather than tuned post hoc.
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. The paper argues that catastrophic forgetting should be modeled as geometric interference rather than repaired post hoc. Under frozen features, realizability, and quadratic excess loss, forgetting of task A after an update from task B is exactly the interference energy ½ΔᵀΣ_AΔ (Theorem 1). This yields a removability dichotomy (lossless iff Δ ∈ ker Σ_A), a distortion floor when supports overlap with target disagreement (Theorems 4–6), and an optimal-merge characterization as Σ-orthogonalization. From the same geometry the authors derive IGFA, a replay- and Fisher-free online controller that shares aligned directions and protects conflicting ones via a similarity gate, carrying only a low-rank subspace summary. Experiments in the exact frozen-feature regime verify the identity to machine precision, lossless allocation on disjoint supports, cost relocation, the s*≈0.26 sign-change, and a capacity knee; multi-seed frozen-backbone benchmarks show IGFA matching OGD/GPM on dissimilar streams and improving when subspace overlaps straddle the gate threshold. Path-averaged curvature and a function-space formalism extend the identities beyond frozen features with quantified residuals.

Significance. If the geometric account holds, it supplies a predictive, pre-training criterion for when forgetting is avoidable versus floor-limited, unifies offline model merging and online continual allocation under one objective, and replaces heuristic projection/regularization with a derived share-or-protect rule. Strengths that raise the contribution above a pure method paper include: closed-form identities under stated assumptions (not fitted forgetting curves); machine-precision verification of Theorem 1 and the information–estimation floor identity; multi-seed tables with paired tests on Split-Digits, Rotated-Digits, and frozen-ViT streams; an explicit claim map separating exact, residual-quantified, and empirical tiers; and released code regenerating the core figures. The signed gate is a strict generalization of OGD/GPM, and the share-density diagnostic makes the gate-versus-projection choice a pre-deployment measurement rather than a post-hoc finding. These are genuine advances for the structural/replay-free branch of continual learning and for task-arithmetic merging.

major comments (3)
  1. [Abstract, §1, §5.5–5.6, Table 3, Fig. 10] Abstract and §1 claim that IGFA “improves on unconditional projection when similarity makes transfer worth preserving.” On the standard task-incremental streams of Table 3 (Split-CIFAR-100, CUB), every pairwise overlap lies below s*, so IGFA is identical to OGD on every seed; the only statistically significant gate gain at scale is ImageNet-R (+0.007 accuracy, p=0.001). The distinctive method claim therefore rests on a narrow, measurable regime (positive share density; Fig. 10). The abstract and contribution list should state this scope condition explicitly—e.g., “matches OGD on dissimilar streams; improves when share density > 0”—so that the empirical claim matches the evidence rather than implying a general accuracy advantage over structural baselines.
  2. [§3.1 A1, §3.3 Thm 2, Prop. 1, §5.4 Fig. 8, §5.7] Assumption A1 is correctly flagged as load-bearing for exact parameter-space identities, and Theorem 2 / §S15 relocate exactness to path-averaged curvature and function space. However, the online method (Algorithm 1, recursive Gauss–Newton tracker) is only an approximate surrogate of ¯H_A under joint training. Sec. 5.4 shows frozen-curvature Pearson r falling to ≈0.2–0.4 at depth before segment averaging restores r=1.00 offline; the online tracker’s residual basis error ϵ and the induced forgetting under realistic deep-network drift (Proposition 1) are not reported on the same depth/LR grid as Fig. 8. A short end-to-end ablation quantifying E[ΔL_A] versus measured ϵ (and versus stale OGD) on the depth-2/3 networks of Sec. 5.4 would close the gap between the exact-regime guarantees and the full-network deployment claim.
  3. [§6, Sec. S5 Tables S2–S4] Language-model support for the retention mechanism is important for the paper’s broader claim, but the main-text discussion (§6) leans on SI tables that are partly single-run (Tables S2–S4) or modest in margin (functional gate vs protect-all not significant at five seeds on GPT-2; active-set gate stronger but still SI-only). Either promote a multi-seed, main-text language result with clear paired tests, or temper the LLM-transfer language in §6 to “structural projection transfers; the transfer-recovering gate is validated in SI with modest margins,” so the load-bearing claim remains the frozen-feature theory rather than an under-powered scale claim.
minor comments (5)
  1. [Abstract, throughout] Notation switches between igfa / IGFA / “signed gate” and ogd/gpm; pick one capitalization and expand once in the abstract.
  2. [Fig. 1, §S4] Fig. 1 packs nine methods and four classes into one scatter; a small legend table of residual D for the three merges would make the merge claim readable without the SI.
  3. [§5.3, §5.5, §S6] s*≈0.26 is derived under unit-norm targets and finite-data bias–variance on a shared block (§5.3); the Rotated-Digits and ViT experiments use s*=0.65 / 0.6. A one-sentence note that the break-even is regime-specific (and that validation-greedy calibration removes the free parameter; §S6) would prevent readers from treating 0.26 as universal.
  4. [§6, §S13] The manuscript is long; several SI extensions are labeled conjecture or mechanism-probe (§S13). Consider moving the claim-map paragraph earlier and trimming Tier-IV material from the main discussion so the exact identities remain the focus.
  5. [§1, §S5] Typos / polish: “oblivious to past knowledge” is fine; “the herein derived” → “the derived”; “atr = 16” in S5 → “At r = 16”; consistent use of Σ-orthogonal vs Σ-orth.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: core identities are direct consequences of stated quadratic/function-space assumptions, validated externally rather than fitted-then-predicted.

full rationale

Theorem 1 is a one-line substitution of the quadratic excess loss (Eq. 1) under A1–A3, not a parameter fit to forgetting data; the same objective yields the merge optimum and distortion floor (Theorems 4–6) as independent mathematical content. Empirical checks (machine-precision identity on linear heads, Pearson r≈0.999 on Split-Digits, multi-seed frozen-ViT tables) compare the derived quantity to measured forgetting rather than reusing fitted targets as predictions. The gate threshold s* is an ordinary free hyperparameter (validation-greedy or fixed plateau), not a fitted constant renamed as a first-principles prediction. IGFA recovers OGD/GPM as the s*→∞ special case by construction of the QP constraint set—an explicit generalization, not a circular renaming of known projection methods. There is no load-bearing self-citation chain (sole author; external CL/NTK/merging literature), no uniqueness theorem imported from prior author work, and no ansatz smuggled via self-citation. Function-space and Bregman extensions (Sec. S15) are algebraic expansions of the loss, again identities under stated assumptions. Score 0 is the honest finding.

Assumptions & free parameters 3 free parameters · 5 assumptions · 3 invented entities

The central exact claims rest on three domain assumptions (frozen features, realizability, quadratic excess loss) plus standard linear algebra. The method adds one free gate threshold (or a validation-greedy surrogate) and a rank-capacity budget. Invented operational entities are the IGFA controller and the interference-ledger state; the interference functional itself is derived from the loss, not postulated. No new physical particles or forces; the main risk is treating A1 as if it were free for deep end-to-end training without the path-averaged or function-space corrections the paper itself quantifies.

free parameters (3)
  • similarity threshold s* = ≈0.26 (synthetic); 0.65 / 0.6 (benchmarks)
    Gate share-vs-protect decision; theory break-even ≈0.26 on synthetic shared block, Rotated-Digits uses s*=0.65, ViT uses s*=0.6; can be validation-greedy but remains a free operating choice that distinguishes IGFA from OGD.
  • rank cap k_max / protected subspace rank
    Capacity budget for the occupied subspace summary; when binding, least-excited directions are overwritten. Chosen relative to feature dimension and per-task effective rank.
  • Gauss–Newton decay γ (online tracker)
    Decay factor in the recursive second-moment update C ← γC + E[Jw Jwᵀ] under feature drift; affects how quickly the protected subspace tracks geometry.
assumptions (5)
  • domain assumption A1: Feature map ϕ is fixed during training (frozen backbone / PEFT / first-order NTK).
    Sec. 3.1; required for exact parameter-space identity Theorem 1 and closed-form Σ_A geometry.
  • domain assumption A2: Each task optimum w*_t attains zero excess loss (realizability).
    Sec. 3.1; mis-specification adds only a task-constant and leaves interference terms.
  • domain assumption A3: Training minimizes the quadratic excess loss L_t(w)=½(w−w*_t)ᵀΣ_t(w−w*_t) by (projected) gradient descent.
    Sec. 3.1; defines the quadratic regime in which motion is confined to range Σ_t.
  • standard math Standard linear algebra: PSD second moments, Moore–Penrose pseudoinverse for singular Σ sums, principal angles for basis error.
    Used throughout Theorems 1–6 and Proposition 1.
  • domain assumption Task boundaries known at training time for the default online loop (relaxed later by velocity detector / recursive tracker).
    Sec. 4 and Algorithm 1 process each task n; Discussion and Sec. S13 address task-free operation as extension.
invented entities (3)
  • Interference-Gated Functional Allocation (IGFA) independent evidence
    purpose: Online replay-free controller that shares aligned occupied directions and orthogonalizes conflicting ones using a low-rank subspace state.
    Derived from the interference functional; empirical entity validated on benchmarks, not an independent physical object.
  • Distortion floor D / D⋆ independent evidence
    purpose: Irreducible residual of any single shared head when task supports overlap with target disagreement; also Bayes excess risk under task confusability.
    Derived quantity (Theorems 4–6), not postulated; has falsifiable lower-bound predictions on merge residual and forgetting.
  • Interference ledger (tracked Σ̂, bases, micro-caches)
    purpose: Single state object from which overlaps, floors, OOD ratio, drift velocity, and QP constraints are read out.
    Operational packaging of the geometry for control; useful but method-specific.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Interference and Retention in Continual Learning." pith.science (2026). https://pith.science/paper/O57E25DQ

@misc{pith2026260709202,
  author       = {Pith},
  title        = {Pith review of: Interference and Retention in Continual Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O57E25DQ}},
  note         = {Machine review of arXiv:2607.09202}
}
read the original abstract

Continual learning commonly relies on post-hoc mechanisms such as replay, elastic regularization, or distillation. This work argues that forgetting should instead be modeled directly as interference between tasks. In the frozen-feature regime, forgetting from learning a new task is exactly the interference energy induced on the old task. In deep networks, the same quantity is recovered through path-averaged curvature with minimal additional forward passes. When task supports are disjoint, forgetting can be eliminated structurally and when task supports overlap in conflicting directions, a non-zero distortion floor is unavoidable. The same geometry optimally merges models through task-aware orthogonalization. From this analysis we derive Interference-Gated Functional Allocation (IGFA), a replay-free, Fisher-free method that shares directions when tasks align and protects them when they conflict. Across benchmarks, IGFA achieves lossless retention when tasks are structurally separable and moves unavoidable cost from irreversible forgetting into deferred but recoverable plasticity when they are not. It matches the strongest replay-free structural baselines on dissimilar-task streams and improves on unconditional projection when similarity makes transfer worth preserving.

Figures

Figures reproduced from arXiv: 2607.09202 by the authors.

Figure 1
Figure 1. Accuracy versus forgetting. Nine methods are tested on the exact-A1 Rotated-Digits benchmark (five rotations, frozen random features; mean over 5 seeds) and differentiated by method class: baseline (•), corrective (•), structural (•), and merging (•); our methods are marked “(ours)” and drawn as stars in their class colour (⋆, ⋆); higher accuracy (up) and lower forgetting (right) are better. Corrective (EWC, replay)… view at source ↗
Figure 2
Figure 2. The task-induced geometry in feature space motivates either full sharing, full separation, or selective parameter sharing. Left (Disjoint). When two tasks excite orthogonal feature subspaces, the update ∆ is composed of a component ∆∥ ∈ range ΣA between the iso-loss contours that incurs forgetting and ∆⊥ ∈ ker ΣA along the iso-loss contours where allocation is interference-free. Centre (Overlapping). When the suppor… view at source ↗
Figure 3
Figure 3. Geometry of the general identities. (a) Beyond frozen features the quadratic form survives with ΣA replaced by the path-averaged curvature H¯A: the local curvature of LA rotates and rescales along the segment wA→wB (ellipses, blue→red), earlier points weighted by (1 − t) (wedge), and the straight chord stands in for the drifted geodesic (dashed) (Theorem 2). (b) Optimal merging is Σ-weighted, not Euclidean: drawn fr… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: The distortion floor and its information-theoretic reading. (a) A floor appears only on a shared active direction u ∈ range ΣA ∩ range ΣB on which the targets disagree by δ, with D ≥ 1 4 σuδ 2 (Theorem 5). (b) The Bayes single-head excess risk D ⋆ factors into the geom…
Figure 5
Figure 5. Figure 5: The interference functional is the forgetting. Measured backward-forgetting of task A after learning task B against the predicted energy 1 2∆⊤ΣA∆. (a) A linear head on frozen features: the identity is exact to machine precision. (b) A nonlinear network, with the predic…
Figure 6
Figure 6. Figure 6: Allocation and the distortion floor. Left: on partially shared features with a conflict on the shared directions, naive descent destroys task A (red rises to 13.94) while allocation preserves it (< 10−17, green flat) and converts that same 13.94 into deferred under-fit…
Figure 7
Figure 7. Figure 7: Similarity and capacity. Left: total population loss of pooling (share) versus isolating (orthogonalize) two tasks on a shared block, against target similarity s. Isolation is flat in s; pooling trades variance reduction against bias and wins only when tasks are simila…
Figure 8
Figure 8. Figure 8: Curvature drift and its correction. Top: predicted-versus-measured forgetting fidelity for backprop￾trained networks, against learning rate (feature drift) and depth. The frozen-curvature identity is exact for frozen features (A1) and degrades as the feature map drifts…
Figure 9
Figure 9. Figure 9: Diagnostics on real data (Split-Digits). Left: measured versus predicted forgetting 1 2 P c ∆⊤ c ΣA∆c for all (earlier, later) task pairs—on the y = x line, Pearson r = 0.999, confirming Theorem 1 on real features. Right: the eigenvalue spectrum of a task’s feature sec…
Figure 10
Figure 10. Figure 10: The gate’s realized gain over ogd against the stream’s share density, for all six real streams in this paper (stars: statistically significant gains; grey: exact or statistical ties). Zero-density streams tie by construction—the gate has nothing to decide—and the gain…
Figure 11
Figure 11. Figure 11: Online subspace update under feature drift. Left: a stale protected subspace (ogd/gpm, red) accumulates drifted copies and its rank grows to capacity d (exhaustion); the recursive Gauss–Newton estimate (green) stays at the true rank (≈ 4). Right: consequently the stal…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 2 linked inside Pith

  1. [1]

    Kirkpatrick et al., Overcoming catastrophic forgetting in neural networks,PNAS114(13):3521–3526, 2017

    J. Kirkpatrick et al., Overcoming catastrophic forgetting in neural networks,PNAS114(13):3521–3526, 2017

    2017

  2. [2]

    Zenke, B

    F. Zenke, B. Poole, S. Ganguli, Continual learning through synaptic intelligence,ICML, 2017

    2017

  3. [3]

    Lopez-Paz, M

    D. Lopez-Paz, M. Ranzato, Gradient episodic memory for continual learning,NeurIPS, 2017

    2017

  4. [4]

    Chaudhry, M

    A. Chaudhry, M. Ranzato, M. Rohrbach, M. Elhoseiny, Efficient lifelong learning with A-GEM,ICLR, 2019

    2019

  5. [5]

    Buzzega et al., Dark experience for general continual learning: a strong, simple baseline,NeurIPS, 2020

    P. Buzzega et al., Dark experience for general continual learning: a strong, simple baseline,NeurIPS, 2020

    2020

  6. [6]

    Z. Li, D. Hoiem, Learning without forgetting,IEEE TPAMI40(12):2935–2947, 2018

    2018

  7. [7]

    Farajtabar, N

    M. Farajtabar, N. Azizan, A. Mott, A. Li, Orthogonal gradient descent for continual learning,AISTATS, 2020

    2020

  8. [8]

    G. Saha, I. Garg, K. Roy, Gradient projection memory for continual learning,ICLR, 2021

    2021

Show all 35 references
  1. [9]

    G. Zeng, Y. Chen, B. Cui, S. Yu, Continual learning of context-dependent processing in neural networks (orthogonal weight modification),Nature Machine Intelligence1:364–372, 2019

    2019

  2. [10]

    Yu et al., Gradient surgery for multi-task learning (PCGrad),NeurIPS, 2020

    T. Yu et al., Gradient surgery for multi-task learning (PCGrad),NeurIPS, 2020

    2020

  3. [11]

    Jacot, F

    A. Jacot, F. Gabriel, C. Hongler, Neural tangent kernel: convergence and generalization in neural networks, NeurIPS, 2018

    2018

  4. [12]

    Doan et al., A theoretical analysis of catastrophic forgetting through the NTK overlap matrix,AISTATS, 2021

    T. Doan et al., A theoretical analysis of catastrophic forgetting through the NTK overlap matrix,AISTATS, 2021

    2021

  5. [13]

    M. A. Bennani, T. Doan, M. Sugiyama, Generalisation guarantees for continual learning with orthogonal gradient descent,arXiv:2006.11942, 2020

    2006 arXiv

  6. [14]

    Ilharco et al., Editing models with task arithmetic,ICLR, 2023

    G. Ilharco et al., Editing models with task arithmetic,ICLR, 2023

    2023

  7. [15]

    Yadav et al., TIES-Merging: resolving interference when merging models,NeurIPS, 2023

    P. Yadav et al., TIES-Merging: resolving interference when merging models,NeurIPS, 2023

    2023

  8. [16]

    G. M. van de Ven, T. Tuytelaars, A. S. Tolias, Three types of incremental learning,Nature Machine Intelligence 4:1185–1197, 2022

    2022

  9. [17]

    V. V. Ramasesh, A. Lewkowycz, E. Dyer, Effect of scale on catastrophic forgetting in neural networks,ICLR, 2022

    2022

  10. [18]

    Wang et al., Learning to prompt for continual learning (L2P),CVPR, 2022

    Z. Wang et al., Learning to prompt for continual learning (L2P),CVPR, 2022

    2022

  11. [19]

    Wang et al., DualPrompt: complementary prompting for rehearsal-free continual learning,ECCV, 2022

    Z. Wang et al., DualPrompt: complementary prompting for rehearsal-free continual learning,ECCV, 2022

    2022

  12. [20]

    J. S. Smith et al., CODA-Prompt: continual decomposed attention-based prompting,CVPR, 2023

    2023

  13. [21]

    Evron et al., How catastrophic can catastrophic forgetting be in linear regression?,COLT, 2022

    I. Evron et al., How catastrophic can catastrophic forgetting be in linear regression?,COLT, 2022

    2022

  14. [22]

    Dohare et al., Loss of plasticity in deep continual learning,Nature632:768–774, 2024

    S. Dohare et al., Loss of plasticity in deep continual learning,Nature632:768–774, 2024

    2024

  15. [23]

    Cheng et al., Whoever started the interference should end it: guiding data-free model merging via task vectors, ICML, 2025

    2025

  16. [24]

    Gargiulo et al., Task singular vectors: reducing task interference in model merging,CVPR, 2025

    2025

  17. [25]

    Marczak et al., No task left behind: isotropic model merging with common and task-specific subspaces, 2025

    2025

  18. [26]

    Toward a holistic approach to continual model merging,arXiv:2509.23592, 2026

    2026

  19. [27]

    Geodesic-aligned gradient projection for continual task learning,CVPR, 2025

    2025

  20. [28]

    MINGLE: mixture of null-space gated low-rank experts for test-time continual model merging,NeurIPS, 2025

    2025

  21. [29]

    Jung, Cho, and Yun, Convergence and implicit bias of gradient descent on continual linear classification,arXiv, 2025

    2025

  22. [30]

    Li and Hiratani, Optimal task order for continual learning of multiple tasks,arXiv, 2025

    2025

  23. [31]

    A broad evaluation of model merging for large language models,arXiv:2511.21437, 2025

    2025

  24. [32]

    Liberty, Simple and deterministic matrix sketching (Frequent Directions),KDD, 2013

    E. Liberty, Simple and deterministic matrix sketching (Frequent Directions),KDD, 2013. 40

    2013

  25. [33]

    Gurnee, J

    W. Gurnee, J. Lindsey, et al., Verbalizable representations form a global workspace in language models, Transformer Circuits Thread, Anthropic, 2026. Code:github.com/anthropics/jacobian-lens

    2026

  26. [34]

    Friedl, N

    K. Friedl, N. Jaquier, A. Liao, D. Kragic, Learning Hamiltonian dynamics at scale: a differential-geometric approach,ICML(PMLR 306), 2026

    2026

  27. [35]

    Rosseau, R

    A. Rosseau, R. Müller, A. Nowé, Preserving plasticity in continual learning via dynamical isometry,ICML (PMLR 306), 2026.arXiv:2606.09762. 41

    2026 arXiv

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.