- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper studies how contextual feedback can reconfigure a recurrent neural system while preserving its intrinsic computation. In an MEC-inspired attractor model, the authors derive a state-dependent tangent direction for structure-preserving translation and distinguish high-dimensional connectivity from low-dimensional effective control. They examine learned feedback geometry and intervene on tangent components during transport and subsequent integration. A second recurrent system and an external waypoint task investigate whether reaching a target representation or preserving instantaneous phase is sufficient to maintain computation and behavior.
The geometric distinctions and MEC interventions are useful. However, the supplementary materials raise substantial concerns about the training provenance of the learned geometry and the validity of target-manifold attainment in the second system.
Critical strengths
-
The MEC interventions provide useful evidence that local control direction matters beyond membership in the predicted control plane. Retaining the tangent component reportedly reduces median maximum distortion from 0.5036 to 0.2514, whereas removing it increases distortion to 0.6464 (p. 5, lines 232–237). The interpretation is limited by the selection of 16 high-distortion starting states and their checkpoint cohort. Nevertheless, this is meaningful functional evidence, and the teacher-initialization concern does not by itself invalidate it.
-
The study makes a substantial effort to connect internal dynamics to sustained behavior. It examines transport, post-capture integration, and external performance, with seed-level comparisons and matched-phase/external-state restarts. This is a worthwhile experimental approach. The distinction among representation, state, and computation provides a useful organizing idea, although the second-system results require further validation before establishing its full empirical scope.
Critical weaknesses
- The training provenance substantially limits the interpretation of the learned geometry, and the manuscript’s own limitation statement does not adequately address it.
Appendix B.1 explicitly states: “This architectural statement does not establish independence from geometry in the loss or supervision: a structure-preservation loss may favor manifold-compatible directions” (p. 13, lines 664–666). This is an appropriate caveat. However, the supplementary materials document a much more direct source of geometry than a possible preference induced by the shape-preservation loss.
In stage4_2a/dagger_control_v1/run.py, lines 27–30, the teacher current is a scalar multiple of (0.8\sin(\beta-\theta)), which is exactly the tangent direction (u_+(\beta;\theta)) derived in Section 2.2, including the 0.8 coefficient. The supervision therefore supplies the specific analytical direction whose appearance is subsequently presented as evidence for the predicted geometry. In stage4_2a/hc_coverage_task1/run_grid.py, lines 110–121, the interface is trained to match these teacher currents and saved as BC.pt. The continuation pipeline then selects this checkpoint in stage4_2a/full_shape_repair_v1/common.py, line 16.
The connection to the reported geometry is specific. stage4_2a/analysis/teacher_near_d2_geometry_v1/DEEP_FINDINGS.md, lines 17–26, contains the same numerical ranges as the manuscript, including Jacobian plane alignment of 0.999954–0.999994. In the same directory, FINDINGS.md, line 31, states that these properties already existed at the start of D2 continuation, that (W) subsequently changed by only approximately 0.036%–0.091%, and that the experiment demonstrates preservation of teacher-established geometry rather than discovery from random initialization.
The manuscript therefore recognizes the relevant limitation in principle but does not apply it to the actual training procedure. A general warning about geometry in the loss or supervision does not substitute for disclosing explicit tangent-teacher pretraining. The results can support implementation and retention of teacher-supplied geometry; they do not independently establish discovery from functional objectives. The authors should disclose the full training history and either provide a teacher-free comparison or narrow this interpretation. This concern addresses the origin of the geometry, not the separate intervention evidence that using the tangent direction can improve transport.
- The second system’s distance metrics do not reliably establish arrival at the specified target manifold, and the behavioral results require clearer model attribution.
In stage5/code/core.py, transverse removes both the oscillator and additional harmonic subspaces (lines 160–168). latent_distance2 consequently omits deviations in the latter subspace (lines 295–307). The purported exact metric also reconstructs harmonic coordinates from the latent state rather than using their actual values in the neural state (lines 331–335), so it shares this blind spot.
A diagnostic using the supplied default core illustrates the consequence. Starting from context 0 at phase 0.37 and adding (0.5) times the first Gextra basis vector gives a surrogate squared distance of approximately (2.39\times10^{-25}) and a purported exact distance of approximately 0.000024. An independent search in the original state space gives a distance of approximately 0.4591, with a conservative continuous-distance lower bound of 0.4589. The evaluator calls this purported exact metric in stage5/code/evaluate.py, lines 123–127, and uses the resulting terminal distance to define reconfigured at line 182.
The concern is measurement validity. This counterexample does not establish how frequently trained trajectories encounter the problem, but it shows that the current metric cannot reliably support the interpretation of target attainment. Existing trajectories should be reevaluated with a corrected distance measure.
Separately, a corresponding offset remains unchanged during a two-second integration of the uncontrolled default core. This is an observation about the supplied intrinsic dynamics, not evidence that a trained closed-loop controller cannot correct the offset. My assessment rests primarily on the metric defect.
The waypoint results also combine distinct controller versions. stage5_full_v2/docs/RESULTS_V2_A2_SAMPLE_SIZE_ALIGNMENT.md, lines 53–64, reports 0.51484 for the original full controller and 0.74686 for full_v2_naive, retrained with a sixfold larger release penalty; both use 24 seeds. The original controller supplies the temporal and matched-restart analyses in stage5_1/STAGE5_1_RESULTS.md, lines 106–115 and 159–173. These are author-reported values, not independently recomputed here. Their aggregation is defined in stage5_full_v2/code/stats_a2.py, lines 25–30 and 69–85: average external_success within each seed, then take the median across seeds. The original controller still outperforms the phase-only condition, but the manuscript needs to identify which version supports each result.
Questions for authors
-
Were the models underlying Figure 2 initialized through the tangent-teacher supervision documented above? Please identify the training sequence and checkpoints, and explain how this relates to the limitation acknowledged in Appendix B.1. Is there teacher-free evidence supporting discovery from functional objectives, or should the interpretation be narrowed?
-
How often do the trained trajectories contain harmonic-subspace deviations missed by the current distance metrics? Please evaluate those trajectories using a corrected state-space distance and report whether target-attainment rates and the principal comparisons change.
-
Which code version and outcome fields generated Figure 4? In stage5/code/evaluate.py, lines 182–184 and 288–289, the representation, phase, and computation fractions are computed independently of first_persistent; I therefore do not attribute those fractions to its error. However, first_persistent initializes its terminal recursion value to False at lines 63–70, causing even an all-true input to return NaN, and its outputs feed the event-time and release-related fields at lines 174–187. Please identify any reported results that use those fields. Please also reconcile the manuscript’s persistent (d_{\mathrm{surr}}<0.1) criterion with the terminal-distance threshold of 0.25 implemented at line 182.
-
Which controller version and release-loss coefficient were used in each panel of Figures 4–5? Please provide the corresponding checkpoints and raw evaluation outputs so that the overall, temporal, and matched-restart results can be checked consistently.