Pith. sign in

REVIEW 2 major objections 4 minor 12 references

Direction of the weights decides which circuit a grokking run will adopt; their magnitude only decides how easily that choice can be overwritten.

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-4.5

2026-07-11 01:07 UTC pith:5DDNAH4E

load-bearing objection Clean causal dissociation: direction portably carries donor-specific circuit identity (40/40), and the flip threshold is predicted by recipient norm with perfect class separation. the 2 major comments →

arxiv 2607.06628 v1 pith:5DDNAH4E submitted 2026-07-07 cs.LG cs.AI

Cross-Trajectory Chimera Interventions Reveal Dissociable Roles of Weight Magnitude and Direction in Grokking

Truong Xuan Khanh This is my paper
classification cs.LG cs.AI
keywords grokkingchimera interventionsweight normweight directioncircuit identitymodular arithmeticcross-trajectory transferbasin selection
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.

When two neural networks train on the same modular-arithmetic task from different random seeds, they often settle into different structured solutions after a long delay known as grokking. This paper asks which pieces of a partially trained network can be transplanted into a second, independently trained network and still determine its outcome. By splitting every weight vector into a size (norm) and a pure direction, then recombining one run's size with the other's direction, the authors build chimeras that continue training. Across forty independent recombinations the final circuit always follows the direction donor, not the size donor; a control that matches the angular distance but uses a random direction produces no such transfer. The switch itself is threshold-like rather than a smooth blend, and the location of that threshold is predicted by how large the recipient's weights already are: high-norm recipients flip early, low-norm ones resist. Size itself carries only a weak, whole-network effect on how long generalization takes. Direction therefore indexes which solution a trajectory approaches, while magnitude governs how susceptible that identity is to being overwritten.

Core claim

Cross-trajectory chimera interventions dissociate the portable roles of weight magnitude and direction in grokking. Implanting a donor's unit direction at the recipient's norm drives the continued run to the donor's eventual circuit identity in 40/40 independent recombinations on two modular-arithmetic tasks; the transfer is donor-content-specific and threshold-like. The interpolation threshold at which identity flips is predicted by the recipient's weight norm, separating perfectly by norm class over all 20 pairs. Norm carries only a modest distributed delay effect and no identity signal.

What carries the argument

Cross-trajectory chimera interventions: given two independently trained runs, decompose each weight vector into norm r and unit direction u, recombine one run's norm with the other's direction (or interpolate directions on the geodesic at fixed recipient norm), continue training, and measure whether circuit identity (cosine similarity of embedding power spectra) follows the direction donor and delay follows the norm donor.

Load-bearing premise

The claim treats the power spectrum of the token embeddings as a faithful stand-in for the whole circuit's identity, so that spectral similarity correctly reports which solution a chimera has entered.

What would settle it

Find a pair of modular-arithmetic runs whose embedding spectra look highly similar yet whose full circuits (attention patterns, MLP features, unembedding) differ, or a chimera that lands in the donor's spectrum but not the donor's actual circuit; either would break the identity metric that drives the 40/40 result.

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

If this is right

  • Direction alone can be used as a portable control knob to select among known circuits without re-training from scratch.
  • Recipient weight norm becomes a measurable state variable that predicts how easily a network's identity can be overwritten by a foreign direction.
  • The adaptive bisection procedure supplies a cheap, reusable way to localize stability thresholds for any intervention that is stable under its training protocol.
  • Norm and direction play non-interchangeable causal roles, so single-trajectory rescaling or freezing experiments that mix the two axes cannot isolate portability.
  • Identity transfer is a basin switch, not a continuous blend, so intermediate directions will typically collapse to one parent circuit or the other.

Where Pith is reading between the lines

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

  • The same norm-versus-direction dissociation may appear in other delayed-generalization regimes once a reliable circuit fingerprint exists.
  • If high-norm networks sit in shallower landscape regions, early training interventions that keep norms large could enlarge the set of reachable circuits.
  • Non-additivity of the identity signal across layers hints that circuit membership is a collective property of the weight configuration rather than a sum of layer-wise votes.
  • The chimera construction itself is architecture-agnostic and could test portability of other geometric or spectral features beyond modular arithmetic.

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 / 4 minor

Summary. The paper introduces cross-trajectory chimera interventions that recombine the weight norm of one independently trained run with the unit direction of another, then continue training under AdamW. On modular addition and multiplication (p=59, one-layer transformer), direction carries a transferable, donor-specific circuit identity: reverse-radial chimeras adopt the donor circuit in 40/40 recombinations (CFS lean sign-consistent), while an angle-matched random control produces no lean. Identity transfer is threshold-like under geodesic (slerp) interpolation of direction at fixed recipient norm; the flip location t* is predicted by recipient norm, separating perfectly by high/low-norm class over all 20 pairs (joint exact permutation p=1.9e-4). Norm itself produces only a modest, non-localizable delay effect (~30% fractional displacement) and no identity signal. An adaptive bisection localizes t* to ±1/64; an optimizer-state ablation (reset / recipient / donor Adam moments) leaves both the identity transfer and the threshold–norm separation intact.

Significance. If the results hold, the work supplies a clean causal dissociation between portable circuit identity (direction) and state-dependent susceptibility (norm) that single-trajectory interventions cannot establish. The chimera construction, matched-random control, and reusable bisection procedure are concrete methodological contributions for probing basin membership across runs. Strengths include fully independent (disjoint-seed) pairs, exact sign tests, perfect class separation under a pure ordinal claim, and an optimizer-state ablation that rules out moment history as the driver. The findings are limited to two cyclic-group tasks and a single architecture, but within that scope they give a falsifiable geometric division of labour that is of clear interest to the grokking and mechanistic-interpretability communities.

major comments (2)
  1. The circuit-identity metric (normalized power spectrum of token-embedding rows, with discrete-log reordering for multiplication) is only a proxy for full circuit equivalence (Limitations; §7 / Appendix A). While the paper correctly flags circularity for embedding-swap cells and applies the metric consistently to parents and chimeras, the central 40/40 and threshold claims rest on this proxy. A modest additional check—e.g., attention-pattern or logit-lens agreement on a subset of pairs—would substantially strengthen the claim that CFS lean reports basin membership rather than embedding-spectrum coincidence alone.
  2. Pair selection deliberately favours large recipient-norm differences (§5, “Scope of the norm–threshold relationship”), producing two well-separated classes rather than a continuum. The perfect 20/20 separation and joint p=1.9e-4 are therefore an ordinal claim under that design. The manuscript should either (a) add a denser sweep of intermediate-norm recipients or (b) more explicitly bound the claim to class separation, so that readers do not over-read a continuous law t*(r).
minor comments (4)
  1. Figure 4 shaded bands are described as visual aids only; a short caption note that they are not a fitted relationship would prevent misreading.
  2. The non-additivity of layer-group identity signals (Appendix A) is left as an open puzzle; a sentence in the Discussion noting that combined swaps still re-grok to full accuracy (already stated) would help readers rule out instability.
  3. Sparse-parity exploration is mentioned only in Limitations; a brief footnote on the hyperparameter mismatch would make the scope decision fully transparent.
  4. Table 1 and Figure 6 are clear; ensuring that the ±1/64 half-width is printed on every t* panel would aid reproducibility.

Circularity Check

1 steps flagged

No load-bearing circularity in the main claims; the only self-definitional issue is the embedding-swap localization the paper itself flags and excludes.

specific steps
  1. self definitional [Appendix A / Section 7 (layer-localization of identity signal)]
    "because the circuit metric is computed on the embedding spectrum, directly swapping the embedding’s direction manipulates the very object being measured, so the embedding and “all-but-embedding” results are circular and we do not interpret them; the non-circular hidden-layer results are the basis for the redundancy claim above"

    Swapping the embedding direction and then scoring identity via the embedding power spectrum makes the measured lean partly tautological for those cells. The paper correctly identifies this and excludes those cells from interpretation; the main full-vector and non-embedding-group results are not affected by this construction.

full rationale

The central results—40/40 direction-driven identity transfer, donor-specificity vs angle-matched random, threshold-like geodesic response, and perfect 20/20 recipient-norm class separation of t*—are empirical outcomes of continued training under cross-trajectory recombinations, not quantities forced by definition or by a fitted parameter renamed as a prediction. CFS lean is measured after continued training against the parents’ final spectra; the matched-random and mid-norm controls, the non-embedding layer-group swaps (attention/MLP/unembedding), and the optimizer-state ablation all provide independent checks that the outcome is not automatic from the implant. Prior single-trajectory work by the same author is cited only for consistency of a weak delay effect and is explicitly disclaimed as non-load-bearing (‘neither an extension nor a re-analysis… draw no stronger link’). The sole self-definitional step is the embedding-direction swap in Appendix A, which the paper correctly labels circular (metric computed on the embedding spectrum) and does not interpret; that supporting analysis is not used for the main dissociation. Score 1 reflects that minor, already-excluded issue only.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 3 invented entities

The central claims rest on standard modular-arithmetic circuit analysis, a geometric decomposition of weights, and an embedding-spectrum proxy for circuit identity. No free parameters are fitted to produce the identity-transfer or threshold-separation results; the only free choices are experimental design decisions (checkpoint selection favoring large norm differences, bisection resolution). The invented measurement entities (chimera, CFS lean, adaptive bisection) are operational definitions with direct experimental handles.

free parameters (2)
  • pair-selection bias toward large recipient-norm differences
    Pairs are chosen so that slow/fast (high/low norm) groups are well separated by design; this makes the continuous law t*(r) untestable and turns the reported relationship into a two-class separation. Stated in Section 5 'Scope of the norm–threshold relationship'.
  • bisection resolution ±1/64
    Threshold localization stops at three bisection steps after a coarse 0.25 grid; the reported t* values inherit this resolution bound rather than sampling noise.
axioms (4)
  • domain assumption Normalized power spectrum of token-embedding rows (index-domain FFT for addition; discrete-log reordering for multiplication) is a valid proxy for full circuit identity.
    Inherited from Nanda et al. 2023 Fourier account of modular-arithmetic circuits; used throughout as the CFS lean metric. Paper itself notes it is only a proxy (Limitations).
  • domain assumption Full-batch AdamW training with the stated hyperparameters produces sufficiently stable continuations that repeated runs differ only by measurement resolution, not sampling noise.
    Required for the adaptive bisection procedure to treat the sign-change of CFS lean as a sharp threshold rather than a stochastic boundary (Section 3).
  • standard math Spherical linear interpolation (slerp) on the unit sphere isolates pure angular change without confounding magnitude.
    Standard geometric fact; linear interpolation of unit vectors would shrink norm by up to ~29% and confound the two axes the paper seeks to dissociate.
  • standard math Disjoint seed matching (no seed reused across pairs) yields independent observations for exact sign and permutation tests.
    Standard statistical design to avoid pseudoreplication (Hurlbert 1984, cited).
invented entities (3)
  • cross-trajectory chimera (r_B u_A or r_A u_B) independent evidence
    purpose: Causal probe that recombines magnitude and direction across independently trained runs to test portability.
    Operational construction; not a new physical entity. Independent evidence is the continued-training outcome itself.
  • CFS lean (circuit-fingerprint similarity lean) independent evidence
    purpose: Scalar that reports whether a chimera's final spectrum is closer to donor or recipient final circuit.
    Derived quantity from cosine similarity of normalized power spectra; inherits the proxy status of the spectrum metric.
  • adaptive bisection localization of interpolation threshold t* independent evidence
    purpose: Reusable measurement tool that localizes the identity-flip threshold to ±1/64 with ~3 evaluations per pair.
    Methodological contribution; independent evidence is the perfect norm-class separation it reveals.

pith-pipeline@v1.1.0-grok45 · 15384 in / 3434 out tokens · 32066 ms · 2026-07-11T01:07:37.864283+00:00 · methodology

0 comments
read the original abstract

Which properties of a partially trained network are causally portable to a different, independently trained network? Single-trajectory interventions show necessity within one run, not portability across runs. We introduce cross-trajectory chimera interventions: given two runs from different seeds, we split each weight vector into a norm and a unit direction, recombine one run's norm with the other's direction, and continue training. On two modular-arithmetic tasks that grok, the components dissociate. Direction carries a transferable, donor-specific circuit identity: implanting a donor's direction at the recipient's norm drives the run to the donor's circuit in 40/40 cases, while an angle-matched random control yields no shift. The transfer is threshold-like, and its location is predicted by the recipient's norm, separating perfectly by norm class over all 20 pairs (joint permutation probability 1.9e-4). Norm carries only a modest, distributed delay effect and no identity signal. An adaptive bisection procedure localizes the threshold to +/-1/64. Direction indexes which solution a trajectory approaches; norm governs how susceptible that identity is to being overwritten.

Figures

Figures reproduced from arXiv: 2607.06628 by Truong Xuan Khanh.

Figure 1
Figure 1. Figure 1: Cross-trajectory chimera intervention. Two networks are trained from different seeds on the same task. Each weight vector is split into a norm r and a unit direction u. A chimera recombines one run’s norm with the other’s direction (here rB uA) and is trained onward; we ask whether its grokking delay follows the norm donor and its circuit identity follows the direction donor. The directional dose-response … view at source ↗
Figure 2
Figure 2. Figure 2: The angular component carries transferable circuit identity. Per-pair CFS lean under the endpoint swaps, for both tasks. The radial variant (keep recipient direction uA; blue) yields an A-like final circuit in every pair (negative), while reverse radial (keep donor direction uB; orange) yields a B-like circuit (positive). Pairs are fully independent (disjoint seeds). Sign is consistent in 40/40 recombinati… view at source ↗
Figure 3
Figure 3. Figure 3: Identity transfer is threshold-like, not a continuous blend. CFS lean as the implanted direction is interpolated along the geodesic from uA (t=0) to uB (t=1) at fixed recipient norm. Grey curves are all pairs on the coarse grid; coloured curves are the representative pairs measured on a 16-point fine grid. Each pair stays near one plateau, then switches over a narrow range—a graded intervention with a shar… view at source ↗
Figure 4
Figure 4. Figure 4: The transfer threshold is predicted by recipient norm. Bisection-localized threshold t ⋆ (midpoint ± half-width, ±1/64) against the recipient’s weight norm rA, for all 20 pairs across both tasks (circles: modular addition; squares: modular multiplication). High-norm (slow) recipients flip early; low￾norm (fast) recipients flip late. The two norm classes separate without overlap; the green band marks the se… view at source ↗
Figure 5
Figure 5. Figure 5: The radial component carries a modest, distributed delay effect. Per-pair delay donor￾following score under the radial variant. The mean is reliably positive (dashed) but no pair reaches the pre-registered “clean” threshold of 0.7 (dotted): the norm shifts delay toward its donor, but only partially, and the effect vanishes when the swap is restricted to individual layer groups (not shown). Task Variant res… view at source ↗
Figure 6
Figure 6. Figure 6: Optimizer-state ablation for the threshold–norm relationship (C5). Bisection-localized t ⋆ against recipient norm rA, repeated under three sources for the continuation optimizer’s Adam moments: freshly reset (left, as reported in [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Division of labour between the two geometric components. Direction carries a transferable component associated with circuit identity, i.e. which solution a chimera approaches; this is portable across independent trajectories regardless of recipient state. Norm does not carry identity, but predicts the inter￾polation threshold at which a donor’s identity overwrites the recipient’s—i.e. how susceptible the r… view at source ↗
Figure 8
Figure 8. Figure 8: The identity component is carried by multiple hidden groups, non-additively. Angular CFS lean when the directional swap is restricted to one layer group (mean over three pairs). Attention, MLP, and unembedding each individually carry the correct-sign signal, but combining them (“attn+mlp”) gives a weaker lean than the strongest single group—the contributions are non-additive. The embedding-based groups (ma… 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

12 extracted references · 12 canonical work pages · 7 internal anchors

  1. [1]

    Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets

    Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets , author =. arXiv preprint arXiv:2201.02177 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    International Conference on Learning Representations (ICLR) , year =

    Progress Measures for Grokking via Mechanistic Interpretability , author =. International Conference on Learning Representations (ICLR) , year =

    work page

  3. [3]

    Towards Understanding Grokking: An Effective Theory of Representation Learning

    Towards Understanding Grokking: An Effective Theory of Representation Learning , author =. arXiv preprint arXiv:2205.10343 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    Explaining grokking through circuit efficiency

    Explaining Grokking Through Circuit Efficiency , author =. arXiv preprint arXiv:2309.02390 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  5. [5]

    Let Me Grok for You: Accelerating Grokking via Embedding Transfer from a Weaker Model

    GrokTransfer: Transferring Features Accelerates Grokking in Modular Arithmetic , author =. arXiv preprint arXiv:2504.13292 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    International Conference on Learning Representations (ICLR) , year =

    Git Re-Basin: Merging Models Modulo Permutation Symmetries , author =. International Conference on Learning Representations (ICLR) , year =

    work page

  7. [7]

    The Role of Permutation Invariance in Linear Mode Connectivity of Neural Networks

    The Role of Permutation Invariance in Linear Mode Connectivity of Neural Networks , author =. arXiv preprint arXiv:2110.06296 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    International Conference on Machine Learning (ICML) , pages =

    Linear Mode Connectivity and the Lottery Ticket Hypothesis , author =. International Conference on Machine Learning (ICML) , pages =. 2020 , note =

    work page 2020

  9. [9]

    Deep Ensembles: A Loss Landscape Perspective

    Deep Ensembles: A Loss Landscape Perspective , author =. arXiv preprint arXiv:1912.02757 , year =

    work page internal anchor Pith review Pith/arXiv arXiv 1912

  10. [10]

    A Basin-Selection Perspective on Grokking via Singular Learning Theory

    A Basin-Selection Perspective on Grokking via Singular Learning Theory , author =. arXiv preprint arXiv:2603.01192 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    Ecological Monographs , volume =

    Pseudoreplication and the Design of Ecological Field Experiments , author =. Ecological Monographs , volume =. 1984 , doi =

    work page 1984

  12. [12]

    Journal of Human Resources , volume =

    A Practitioner's Guide to Cluster-Robust Inference , author =. Journal of Human Resources , volume =. 2015 , doi =

    work page 2015