- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes a landscape theory of association. A density over a representation space is mapped to a potential ; non-degenerate modes of are taken as associative states, and association is modeled as overdamped gradient diffusion on with isotropic noise of strength .
The central claim is that three quantities derived from the same landscape must be kept distinct: (i) fixed-duration path likelihood, ranked by the full Onsager-Machlup (OM) action including its divergence term; (ii) the weak-noise scale of escape from a basin, set by Freidlin-Wentzell direct-exit barriers ; (iii) next-state choice, obtained by normalizing competing Eyring-Kramers exit rates over saddle-adjacent basins, whose log-odds slopes in equal barrier differences.
Validation consists of three simulation studies on hand-designed two-dimensional landscapes. On a scaled Müller-Brown potential, fitted Arrhenius slopes match the two barriers within 5.45% and 0.14%. On an equal-barrier two-channel landscape, a moving-tube backward-Kolmogorov PDE that does not use any action functional reproduces the full-OM ordering of six prespecified paths, while the quadratic-only truncation selects a different winner. On a five-well Gaussian mixture, 160,000 first-exit trajectories recover all four barrier-difference log-odds slopes within 4.5%; multiplying one mixture weight by 1.5 (centers and covariances fixed) raises the targeted exit probability by 21.8 points on average and reverses the dominant exit, and a non-adjacent well is never reached directly.
Critical strengths
1.A correct and clearly stated decomposition of three quantities that applied work frequently conflates. The statements are accurate: the free-time infimum of the full OM action diverges to , quadratic-only or Euler-discretized objectives drop the divergence correction and can select a different path, and sourcewell depth cancels exactly in same-source log-odds. The paper demonstrates concretely, via the equal-barrier construction, that dropping the divergence term can reverse which route is selected. Applied communities that compute "most probable paths" from learned energy landscapes (for example transition-path sampling with score-based models) can genuinely benefit from this clarification, so the paper has value as a careful reference independent of its cognitive framing.
2.Exemplary numerical discipline. The experiments use frozen protocols, prespecified candidate sets, complete event accounting ( 160,000 releases with 17 censored and zero failed), whole-batch bootstrap intervals, time-step, grid, and initialization audits, and an action-independent PDE adjudication in Section 4.2. Deviations from asymptotics are reported honestly. I could rebuild the equal-barrier landscape from Equations (34) and (35) alone; my critical points and Hessians match Table 7, and my minimized OM actions reproduce Table 11 to within about 0.05 nats (for example 144.48 versus the reported 144.43 for c00). This level of reproducibility is rare and deserves recognition.
Critical weaknesses
1.The cognitive claim, which is the paper's reason for existing in this primary area, is never tested, and the controls are weaker than they look. By the authors' own account the theory is classical (Kramers and Arrhenius escape, the OM small-tube theorem, Eyring-Kramers competing rates), so the contribution must rest on the application to association. Yet there is no association data: the USF and SWOW norms are cited but never used, there is no learned density, no embedding space, and no comparison with association or memory-search models. The non-circular test the authors themselves describe in Section 2 (learn the landscape without association responses, then predict held-out behavior) is not run. All three experiments generate data from the same SDE that the theory analyzes, so agreement is expected by construction.
Moreover: (a) "no direct exits" is guaranteed by basin topology for continuous paths, so it tests nothing; (b) the penalty of the topology control equals under the frozen distance score and is a data-independent constant, which I verified from Table 13; (c) the center-distance baseline is a strawman, since it ignores mixture weights by construction, and a trivial weight-aware score also flips the dominant exit (E: 0.346 to 0.297, U: 0.327 to 0.422); (d) for well-separated isotropic components the barrier reduces, to leading order, to a quadratic distance term plus a log-weight bias, so the choice law approximates a bias-weighted Gaussian-similarity Luce rule restricted to neighbors (cf. Shepard 1987; Nosofsky 1986). What the landscape adds empirically over this classical family is never shown.
2.The interpretation of the path layer is contradicted by the paper's own benchmark. The OM small-tube ranking describes the mode of the path measure, not where probability mass flows. Reproducing the setup of Section 4.2: the full-OM winner spends about 7.7 of the time units within 0.1 of the narrow upper saddle, where the Laplacian of is about 36, because confinement there maximizes small-tube probability. Direct simulation of the same SDE (Euler-Maruyama, and , same ) tells the opposite story: the narrow channel carries only about one third of actual transitions, both for free-time crossings (34 to 35% across runs of 4,000 to 8,000 trajectories) and for fixed- bridges ending near (34 to 37% across radii and time steps). Trajectories that do use the narrow channel linger near the saddle for only about 0.1 time units, roughly 1/70 of what the OM-optimal path does. The Eyring-Kramers asymptotic split is 26%/74%, so the paper's own third layer predicts the opposite channel preference from its first layer, a tension the paper never acknowledges. Meanwhile the tube log-probabilities are around -1600 , so these "most probable routes" are events of no observable relevance. Claims such as "local curvature shapes probable routes" therefore hold only in a technical sense that has no counterpart in association behavior. Route-level questions call for ensemble quantities such as reactive currents in transition path theory, which are not discussed.
Questions for authors
1.Real data. Can you demonstrate the framework on association data? For instance, learn or an energy model on word embeddings without association responses, identify modes and saddles, and predict SWOW or USF response distributions for held-out cues, against baselines such as embedding-similarity Luce choice with frequency bias and random walks on semantic networks. This is the result most likely to change my rating.
2.Meaning of the path layer. Which association observable corresponds to a fixed- small-tube probability, and how should be chosen? In your own benchmark the narrow channel carries about one third of simulated transitions while the OM winner sits at the saddle for about of . Do you agree that the OM ranking does not describe where trajectories go, and would reactive-flux channel probabilities be the appropriate route observable?
3.Reversibility. Gradient diffusion with isotropic noise satisfies detailed balance, so the reduced jump chain obeys Kolmogorov's cycle condition and asymmetry can enter only through a scalar potential. Is this consistent with free-association norms, and how would the framework capture directional effects not explained by state depth, such as forward asymmetry in sequential recall (Kahana, 1996)?
4.Distinctive predictions. Which quantitative predictions distinguish the landscape model from a biasweighted similarity choice rule restricted to neighbors (see Weakness 1 )?
5.Clarification. For c00, both zero-radius extrapolations (140.28 and 142.70) miss the full-OM limit of 144.43 by 1.7 to 4.2 nats, far outside the stated -nat deterministic bounds. Please report extrapolation uncertainty separately from grid uncertainty.
Suggestions for future work: relate the framework to sampling accounts of memory search (for example Zhu, Sanborn and Chater, NeurIPS 2018); enlarge Figure 2.
6 comments
Response to Reviewer a5a9 [Part 2/3: Q3–Q4]
Response to Q3: Reversibility and directional effects
The gradient diffusion in our model is reversible. In the metastable approximation, the jump chain has reversible weights , where is basin occupancy and . Opposite cue-conditional probabilities can therefore differ, while the forward and reverse probability products around a cycle remain equal.
Your question about directional retrieval points to a useful extension. Persistent net circulation at stationarity, measured on a common set of states, would require a nonreversible model, for example one with a nonconservative drift. The direction and strength of that drift require information beyond a static representation density. We use gradient dynamics here to isolate predictions from the density-derived landscape; adding such a field would be a separate modeling and validation task. Forward effects tied to study order also require assumptions about encoding or retrieval context.
Response to Q4: Weighted similarity controls and topology
We agree that weighting can reverse the winner. We therefore compared full probability predictions using four calibrated similarity-choice controls [3]: shared and inverse-noise scores, each with component-center or actual-minimum distances. The score is , optionally multiplied by . We fitted two nonnegative shared coefficients in five leave-one-noise-level-out folds, excluding both G0/G1 conditions at the test noise. All methods use the same E/U/L support. FW and EK keep their original frozen geometric predictions.
| Model | Mean held-condition NLL ↓ | Mean JS ↓ |
|---|---|---|
| Original fixed weighted Luce (centers) | 1.05923 | 0.050915 |
| Calibrated: centers, shared | 0.93102 | 0.014025 |
| Calibrated: centers, inverse-noise (primary) | 0.91749 | 0.009960 |
| Calibrated: actual minima, shared | 0.92732 | 0.013041 |
| Calibrated: actual minima, inverse-noise | 0.91347 | 0.008922 |
| FW | 0.88247 | 0.001082 |
| EK | 0.87816 | 0.000053 |
Calibration closes about 80% of the original fixed-control gap to FW. Against the primary control, FW and EK reduce NLL by 0.03502 [0.03335, 0.03718] and 0.03934 [0.03813, 0.04091], respectively. Against the best listed sensitivity control, the gains are 0.03100 [0.02940, 0.03307] for FW and 0.03532 [0.03418, 0.03681] for EK. For each of the four calibrated controls, FW has lower NLL in 9/10 conditions and EK in 10/10. Both also have positive mean gains in all five held-noise folds. The intervals are nominal 95% matched-batch bootstrap intervals from five batches, with training parameters refitted in every resample. This comparison is posthoc: G1 enters training at the other noise levels.
At the middle noise level in G0, U and L have equal weights and center distances but measured probabilities of 0.308 and 0.086. The primary control predicts 0.180 for each. Using actual-minimum distances gives only a small asymmetry, 0.185 and 0.173, whereas EK predicts 0.304 and 0.091. A center-based score cannot separate two destinations with identical weight and distance inputs, regardless of the shared coefficients. Across five noise levels, the mean absolute error in the intervention-induced increase in is 12.37 percentage points for the primary control and 11.95 for the best sensitivity, compared with 1.19 for FW and 1.26 for EK. These comparisons show that the landscape-based models predict the full choice distribution and the magnitude of the intervention effect more accurately, on average, than the calibrated endpoint controls tested here.
Adding R incurs the penalty because it diverts probability mass from E/U/L. This follows from normalization. The predictive comparisons above all use the same E/U/L support.
Response to Reviewer a5a9 [Part 3/3: Q5 and References]
Response to Q5: Extrapolation and presentation
For versus , linear and quadratic zero-radius fits give 140.283 and 142.703, compared with the full-OM limit of 144.434. We will report three quantities separately: propagated PDE numerical sensitivities (0.060 and 0.062 nats), extrapolation residuals (−4.152 and −1.731 nats), and sensitivity to the extrapolation specification (2.420 nats). The PDE figures quantify numerical sensitivity, not total extrapolation error. At the smallest tested relative radius (0.03), the independent PDE reproduces the full-OM ordering of all six predeclared candidates under grid refinement and tighter solver tolerances.
References
[1] Zhu, J.-Q., Sanborn, A. N., & Chater, N. (2018). Mental Sampling in Multimodal Representations. Advances in Neural Information Processing Systems, 31.
[2] E, W., & Vanden-Eijnden, E. (2006). Towards a theory of transition paths. Journal of Statistical Physics, 123, 503–523.
[3] Nosofsky, R. M. (1986). Attention, similarity, and the identification–categorization relationship. Journal of Experimental Psychology: General, 115(1), 39–57.
Response to Reviewer a5a9 [Part 1/3: Q1–Q2]
Thank you for the detailed review and for independently reproducing our geometry and action calculations. Your comments prompted us to carry out additional channel measurements and calibrate the similarity controls. Below, we report these results and address your questions.
Response to Q1: Data-derived landscapes and human association
We agree that testing the model against human association data is an important next step. We plan to pursue the experiment you suggest in our next paper. We have already carried out part of the computational work on real image representations and can report those results here.
We used THINGS images and frozen CLIP ViT-B/32 embeddings to construct a spherical energy field over 1,854 concepts, retaining all 512 coordinates. Tangent-Hessian checks certified 1,847 minima, and a bounded search identified 55 concept-pair connections through index-one saddles. We did not use SWOW responses to construct the landscape or find these connections.
Starting at the certified shrimp minimum (D=0.13905), we used 768 new importance-weighted paths to estimate first-grid-exit event masses, averaged over a fixed preparation-age window. Within the finite observation window, the scallop-to-seafood log mass ratio was 0.663 (nominal family interval [0.222, 1.105]). We checked sensitivity to the grid with a separate 256-path refinement panel. We will include this high-dimensional example and its local geometry diagram in the appendix.
These results show that we can construct a landscape from learned representations and measure local transition events on it. They provide a computational basis for the proposed human-association study.
Response to Q2: Reference tubes, channel use, and duration
Thank you for pointing this out. The OM result compares shrinking tubes around specified time-parameterized paths. Channel fractions sum over the trajectories using each channel. An empirical test of the path observable would need repeated, time-resolved latent trajectories with a specified readout and time alignment. Zhu et al. [1] offer a related example, connecting sampling in multimodal representations to memory-search behavior through retrieval distances and temporal correlations. We examined both probability objects on the original equal-barrier two-channel landscape.
For the six predeclared reference paths, the existing action-independent absorbing-tube PDE gives the full-OM ordering under both spatial refinement and tighter solver tolerances. At the smallest tested relative radius, ε/∣b−a∣=0.03, logP(c00)−logP(c03)=108.085 nats. Full OM selects c00, whereas omitting the divergence term selects c03 and ranks c00 last. The PDE calculation independently tests the effect of drift divergence on these reference-tube probabilities.
We also ran SDE simulations at the original noise level and target radius 0.20, with observation horizons T and H=64T:
The first-arrival experiment has 205 censored trajectories among 8,192 releases. The endpoint experiment uses 16,384 releases. The paired time steps give identical primary fixed-time classifications. First-arrival upper shares stay between 34.15% and 34.44% across the tested radii, crossing rules, and time steps. Within the same [0,T] window and radius-0.10 saddle neighborhood, c00 spends 7.707 time units, compared with a median of 0.082 among upper-channel paths conditioned on the terminal event.
These results distinguish reference-tube preference from ensemble channel use. Each candidate tube ends inside the same target disk, so conditioning on that terminal event preserves the tube-probability ratios. For equilibrium channel use, committor-based reactive flux is the appropriate observable [2]. We will put these results alongside the tube rankings in the equal-barrier experiment and clarify the abstract and Figure 2 caption. We chose T=3τ≈8.57, where τ=1/λmin(∇2Φ(a)) is the source well's slow relaxation time. Relating this duration to human response timing requires independent calibration.