acceptodds
Under review as a conference paper at ICLR 2027

Global Optimization via Softmin Energy Minimization

Abstract

Non-convex optimization in high dimensions requires balancing gradient information with global exploration, a tension that Langevin dynamics and other gradient-informed methods address only partially. We introduce a gradient-based swarm method driven by a Softmin energy interaction , a smooth approximation of the minimum over the ensemble that couples particles through their relative energies, and add Brownian noise to its gradient flow. We show that at every stable equilibrium the interaction splits the ensemble into particles at local minima and particles at local maxima; that without noise the ensemble's best value converges at least as fast as gradient flow under a Polyak–Lojasiewicz condition; and that the interaction caps the Freidlin–Wentzell quasipotential barrier between adjacent basins at order , independently of the true barrier height and strictly below the barrier an overdamped Langevin particle faces. Consistently with this cap, on double-well and quadruple-well benchmarks Softmin reaches the target in every run at every barrier height tested, while Simulated Annealing, Langevin dynamics, and variants of Particle Swarm Optimization and Consensus-Based Optimization all eventually fail; counting failed runs, it is more than an order of magnitude cheaper than Simulated Annealing overall, though not on low barriers. Nor is this a mere effect of ensemble size: Softmin succeeds in every run with five particles, whereas Simulated Annealing gains from more particles only what independent restarts predict, enough on moderate barriers but not on high ones. On Lennard-Jones clusters, Softmin's energy gap to the best known minimum stays nearly flat as the cluster grows while every baseline's rises, and the method trains transformers with millions of parameters. Its limits are equally clear: Particle Swarm Optimization is faster on a high-dimensional Ackley function, and Adam on transformer training, where no method struggled.

Primary area:
optimization
Keywords:
Particle Swarm, Non-convex Optimization, Stochastic Particles Methods, Boltzmann Energy

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_nDshOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a gradient-based swarm method for continuous non-convex optimization. It uses the Gibbs-weighted average of the particles' objective values, JβJ_\beta, as an ensemble energy and adds Brownian noise to its gradient flow. This energy induces a state-dependent effective learning rate for each particle: some particles perform gradient descent, while others perform gradient ascent.

The theoretical analysis characterizes locally asymptotically stable equilibria and proves an exponential bound on the ensemble minimum under a suitable Polyak–Łojasiewicz condition for the deterministic dynamics. By constructing a one-dimensional channel in which the other particles are frozen, the paper further obtains a quasipotential upper bound of order O(1/β)O(1/\beta) that is independent of the physical barrier height. The experiments cover double-well and quadruple-well landscapes, Lennard–Jones clusters, the Ackley function, and transformers trained on modular addition. They examine reliability, computational cost, ensemble size, and the contribution of the sign-changing pre-factor.

Critical strengths
  1. A clear mathematical connection between the ensemble energy, stable equilibria, and escape channels.

Starting from a simple ensemble energy, the paper derives the explicit gradient decomposition ∇xk(nJβ)=nwk[1−β(f(xk)−Jβ)]∇f(xk),\nabla_{x_k}(nJ_\beta)=nw_k\bigl[1-\beta(f(x_k)-J_\beta)\bigr]\nabla f(x_k), and uses it to characterize the combinations of local minima and local maxima that can occur at locally asymptotically stable equilibria. The invariant channel obtained by freezing the other particles further reduces the analysis to an explicit reduced potential and yields a bound independent of the physical barrier height.

  1. The experiments report meaningful controls, failures, and limitations.

The authors report reliability and cost by difficulty, vary the ensemble size from 5 to 400 particles, and compare signed, positive-part, and absolute-value variants of the pre-factor. The Lennard–Jones experiments apply the same final local relaxation to every method, and the transformer experiments give Adam and SGD the same number of independent initializations as the Softmin ensemble.

Critical weaknesses
  1. Asymmetric hyperparameter selection weakens the principal reliability and cost comparisons.

The main concern arises in Appendix E.3. Softmin's inverse temperature β\beta is ultimately selected through direct validation of actual first-entry time, whereas the initial temperature T0T_0 of Simulated Annealing is selected using a different criterion.

The authors state that increasing T0T_0 makes escape faster in their continuous-diffusion setting. They therefore do not select it by escape time. Instead, they require $ \exp(-h/T_0)\in[0.01,0.05], $ and choose T0=hmid/4T_0=h_{\mathrm{mid}}/4, to avoid making the barrier too easy to cross at the initial temperature.

This restriction is not adequately justified for an optimization comparison whose principal metrics are first-entry time and expected running time. If increasing the temperature improves the baseline under the stated evaluation criterion, that possibility should be included in its parameter selection. Testing similar numbers of candidates does not compensate for optimizing different selection criteria.

This issue directly affects the main conclusions. The authors have already established that SA's failures at high barriers are associated with freezing under the chosen geometric cooling schedule, and these failures substantially increase the cost differences in Figure 1 and Tables 10 and 12. The current experiments therefore do not rule out the possibility that Softmin's advantage would become substantially smaller if the initial temperature and cooling rate were selected under the same objective.

The baseline selection also leaves the interaction mechanism and application-level competitiveness incompletely assessed. Existing Swarm-based Simulated Annealing uses relative objective values to transfer mass between agents and adjusts their temperatures accordingly, making it a directly relevant interacting gradient/noise method. On Lennard–Jones clusters, a comparison with basin-hopping, which the paper itself cites, would provide a more informative assessment of competitiveness on that application. These comparisons would respectively help assess the benefit of particle interaction and the strength of the application results.

  1. The specific benefit of JβJ_\beta is not sufficiently isolated; a barrier cap of the same order does not require sign reversal.

The paper emphasizes sign reversal and barrier flattening as its principal mechanism. However, the simpler log-sum-exp energy already discussed in the related work also admits a channel bound of the same order. The following calculation uses the paper's definitions and the same channel construction.

Let $ F_\beta(\mathbf{x})=-\frac1\beta\log\sum_{i=1}^{n}e^{-\beta f(x_i)}, $ and apply the same factor of nn used in the proposed dynamics. Then ∇xi(nFβ)=nwi∇f(xi).\nabla_{x_i}(nF_\beta)=nw_i\nabla f(x_i).

Every pre-factor is positive, so there is no gradient ascent. Along the same channel with the other particles frozen, let r=f(x)−f(z−)r=f(x)-f(z_-). The potential relative to the initial configuration is v(r)=−nβlog⁡(n−1+e−βrn).v(r)=-\frac n\beta\log\left(\frac{n-1+e^{-\beta r}}n\right).

For r>0r>0, 0<v′(r)=ne−βrn−1+e−βr<1.0<v'(r)=\frac{ne^{-\beta r}}{n-1+e^{-\beta r}}<1.

Therefore, for every h>0h>0, $v(h)<\min ${ h,nβlog⁡nn−1h, \frac{n}{\beta}\log \frac{n}{n-1}}.

The same channel-action argument thus gives an upper bound that is independent of hh and of order O(1/β)O(1/\beta). Consequently, the existence of an O(1/β)O(1/\beta) barrier cap does not by itself distinguish the proposed sign-changing mechanism from Gibbs-weighted diffusion whose drift always follows the descent direction.

This observation does not eliminate the value of JβJ_\beta. The proposed energy can provide a smaller initial channel-action constant and different post-saddle dynamics, and those differences may be consequential in the small-noise regime. The paper nevertheless needs to establish quantitatively when they translate into an advantage.

The reported ablations reinforce this concern. Appendix E.5.4 reports mean first-entry times of 32.69, 30.70, and 10.82 for the signed, positive-part, and absolute-value variants, respectively, on the quadruple well at A=30A=30 under the main setting β=10\beta=10. The distributions for the three variants substantially overlap on Ackley and Lennard–Jones. The additional β=2\beta=2 ablation does provide some evidence in favor of sign reversal, but it does not explain the advantage under the main settings.

A direct comparison with diffusion driven by nFβnF_\beta, together with a pure-noise control that removes the drift, would assess the specific benefit of the proposed energy more directly than changing only the sign treatment of AkA_k.

Questions for authors
  1. Do the principal conclusions persist when SA is calibrated under the same objective? If the initial temperature and cooling rate are jointly selected using the same reliability, first-entry-time, or ERT criterion used for Softmin, without imposing a low Arrhenius escape factor, do the high-barrier reliability and cost advantages remain?

  2. What does JβJ_\beta add beyond diffusion driven by the log-sum-exp energy? Under comparable parameter-selection ranges, ensemble sizes, noise levels, and computational budgets, can the authors directly compare against drift −nwk∇f(xk)-nw_k\nabla f(x_k) and include a pure-noise control? In particular, under what conditions do the smaller action constant and altered post-saddle dynamics yield a measurable benefit?

  3. Which exit theorem is intended in Corollary 3.5? The point Γ(xm)\Gamma(x_m) is necessarily an equilibrium, whereas Appendix B excludes equilibria on the boundary. Can the authors state a compatible theorem and its hypotheses, and at least explain how those hypotheses apply to a concrete class of potentials considered in the paper?

  4. Are unsuccessful runs stopped early, and what cost does Table 12 record? Table 1 specifies n=100n=100 and a budget of 10710^7 steps, while Table 12 reports an SA success rate of approximately 72.572.5%. If the remaining runs exhaust the full budget, as stated in the caption, the mean number of per-particle gradient evaluations should be at least approximately (1−0.725)×100×107=2.75×108,(1-0.725)\times100\times10^7=2.75\times10^8, whereas the table reports 1.414933×1071.414933\times10^7. Early stopping could explain this discrepancy, but the current description is insufficient to establish that.

1 comment

We thank the reviewer for the detailed review. We checked the calculations in it and agree with most of them.

  • W1/Q1 The reviewer is right that β\beta and T0T_0 were selected under different criteria. The escape time of Simulated Annealing decreases monotonically in T0T_0, so selecting it by escape time would pick the largest value in any grid; we excluded this with the Arrhenius criterion, but we agree that this does not justify a different objective. We have not re-run Simulated Annealing during the discussion, so we cannot say whether the high-barrier advantage survives: under geometric cooling a larger T0T_0 should shift the barrier at which it freezes rather than remove it, but this has to be checked. In a revised version we will select T0T_0 and the cooling rate jointly by ERT at the same point used for β\beta, without the Arrhenius restriction, extend the sweep until the schedule fails, and report the outcome either way. Until then, our conclusions on Simulated Annealing are conditional on the calibrated schedule.

  • W2/Q2 The reviewer's calculation is correct: the O(1/β)O(1/\beta) cap comes from the Gibbs weighting, which log-sum-exp shares, and we will say so. What the sign change adds is a smaller constant. From q(rc)=0q(r_c)=0 one gets exactly uc=n(rc−1/β)u_c=n(r_c-1/\beta), so uc/v∞=(βrc−1)/log⁡nn−1u_c/v_\infty=(\beta r_c-1)/\log\frac{n}{n-1}, which is 0.400.40 at n=2n=2 and decreases to 1/e1/e. The post-saddle structure also changes: for h>rch>r_c, JβJ_\beta creates a shallow well at Γ(s)\Gamma(s), but even adding this second crossing the channel cost 2uc−u(h)2u_c-u(h) stays below v(h)v(h) for every n≥2n\ge2. The gain is therefore a constant factor in the bound on the exponent, not a different order in β\beta. Empirically, only the β=2\beta=2 ablation (App. E.5.4) shows the sign mattering; we will add the drift −nwk∇f(xk)-nw_k\nabla f(x_k) and a pure-noise control on the wells, with the same seeds, ensemble size, noise and budget.

  • Q3 The reviewer is right: the point Γ(xm)\Gamma(x_m) where the pre-factor flips is an index-one saddle of nJβnJ_\beta on ∂D−\partial D_- (for h>rch>r_c the Hessian there is diagonal, with entries −nβw1f′(xm)2<0-n\beta w_1f'(x_m)^2<0 and nn−1f′′(z−)>0\frac{n}{n-1}f''(z_-)>0), so the exit theorem recalled in Appendix B does not apply as stated. We will restate Corollary 3.5 with an exit theorem that allows a characteristic boundary, restricted to the exit from D−D_-; we are still checking which result fits our setting and what its exact hypotheses are. For a gradient drift, such a theorem gives lim⁡σ→0σlog⁡EτD−=min⁡∂D−nJβ−nJβ(Γ(z−))≤uh\lim_{\sigma\to0}\sigma\log\mathbb E\tau_{D_-}=\min_{\partial D_-}nJ_\beta-nJ_\beta(\Gamma(z_-))\le u_h. For our double wells, Γ(z−)\Gamma(z_-) and Γ(xm)\Gamma(x_m) are nondegenerate, while the boundedness of D−D_- and the equilibria on its boundary remain to be checked. The quasipotential bounds of Theorem 3.4 are not affected.

  • Q4 Yes, unsuccessful runs are stopped early, for every method, when the ensemble collapses (largest per-coordinate standard deviation below 10−610^{-6}) or the best value does not improve for 5⋅1055\cdot10^5 steps. All failed Simulated Annealing runs stop by collapse at about 5.1⋅1055.1\cdot10^5 steps, which gives the value in Table 12 (0.275×100×512,900≈1.41×1070.275\times100\times512{,}900\approx1.41\times10^7), and all failed Langevin runs stop by stagnation at 5.49⋅1055.49\cdot10^5 steps. The caption of Table 12 is wrong: a failed run contributes the steps it ran, not the full budget. We will fix it and describe the rule in App. E.3. Since failed runs are charged less than the full budget, the cost ratios are conservative. Stopping does not change the outcomes: the collapsed Simulated Annealing ensemble is frozen, and for Langevin at the smallest failing barrier Kramers' formula gives about 10810^8 time units for the first of 100 particles to escape, against a full budget of 10410^4.

Reviewer_q1kzOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary
  • This work introduces a new global particle optimization methods.
  • It shows that this method has appealing mathematical properties using large deviations theory.
  • It validates the method on several benchmarks.
Critical strengths
  • The method seems new
  • The theory is interesting and very promising: if the main result (height independent exit time) actually holds, this is a significant result.
  • Experiments are diverse and show the method's advantage on standaard benchmarks.
Critical weaknesses

Theory

Barrier height dependence in the prefactors

The FW type analysis here neglects the prefactors when looking at the regime where σ→0\sigma \to 0. However, this may not be informative in the large barrier height regime. Indeed, the current proof relies on r<<erβr << e^{r\beta} (proof of lemma 3.3). However, the Eyring–Kramers formula suggests that the expected exit time will have prefactors of order erβe^{r\beta} which is the quantity that the current paper assumes is large. This introduces some potentially tricky interactions between r,β,σr, \beta, \sigma which should be properly laid out for the presented result to support the main claim of "independence on the barrier height".

Actual exit time

Cor. 3.5 cannot be applied to analyze the exit time from the effective basin: it requires no attractor of the flow of JJ while ss is a local min of ss so it cannot be applied to also analyze the jump above ξ+\xi_+ which separates ss from z+z_+. This should be clarified and renders the main claim less supported.

Presentation

The main text is very unstructured and the problem is worse for the appendix. It reads more like research notes that an actual paper, making me believe that this work is not all ready for publications.

Experiments

As the paper itself notes, the transformer experiment is inconclusive. The x axis also indicates "steps", which renders the comparison between SGD and particle methods unfair, since they are much more expensive per step.

Questions for authors

Address the weaknesses above

2 comments

We thank the reviewer for a careful reading of the theory. We agree with part of the criticism and want to be precise about what the paper claims.

  • In the paper we limit our main theoretical result to the computation of the Freidlin–Wentzell quasipotential, which is shown to be capped at order 1/β1/\beta uniformly in the barrier height. However, the reviewer is right in saying that the σ→0\sigma\to0 limit is not shown to be uniform in hh, since it is taken at fixed hh and sub-exponential factors can depend on it. The main point is that the comparison is between exponents in 1/σ1/\sigma at fixed hh: exactly hh for independent Langevin, and a quasipotential bound of order 1/β1/\beta for Softmin. In the σ→0\sigma\to0 limit, the exponential factors are what determine the exit times, and these are correctly captured by the FW quasipotential. A prefactor of order eβhe^{\beta h} would not change these exponents, but it could matter at finite noise, and the paper does not control it; we will state this explicitly. On the double well, with βh\beta h up to about 2,2502{,}250, Softmin reaches the target in every run within 10510^5 steps, so no growth of that order appears in practice.

Computing the exact exit time poses difficulties beyond the computation of the Eyring–Kramers prefactors, and the reviewer is right that this is not laid out in the paper.

On the specific step, we could not find an approximation r≪eβrr\ll e^{\beta r} in Lemma 3.3 or its proof. The cap follows from the exact inequality u(r)=n r e−βrn−1+e−βr<nn−1 r e−βr≤ne(n−1)β(r>0).u(r)=\frac{n\,r\,e^{-\beta r}}{n-1+e^{-\beta r}}<\frac{n}{n-1}\,r\,e^{-\beta r}\le\frac{n}{e(n-1)\beta}\qquad(r>0). If the reviewer had another line in mind, we would be grateful for the pointer.

  • About Corollary 3.5, the reviewer is right: it only concerns the exit from the initial basin D−D_-, and it cannot describe the jump beyond ξ+\xi_+, since Γ(s)\Gamma(s) is itself a local minimum of nJβnJ_\beta when rc<hr_c<h. For this second crossing the paper only gives a quasipotential bound, uc−u(h)<ucu_c-u(h)<u_c (Theorem 3.4, part 3). Moreover, the boundary of D−D_- contains the point where the sign of the pre-factor flips, which is a saddle of nJβnJ_\beta, so the exit theorem recalled in Appendix B does not apply as stated; we will restate the corollary with a theorem that allows a characteristic boundary. Corollary 3.5 alone is therefore not enough to make a statement about exit times, and we do not make an unconditional one in the paper. This should be made clear in a remark following the corollary, and we thank the reviewer for pointing it out.

  • We are sorry that the reviewer found the presentation of the paper insufficient. We will take this into account and improve the readability of both the main body and the appendix.

  • About the transformer experiment: as the reviewer points out, it is not meant to further support reliability, but to show that our method scales to very high dimensions. We need to make this scope clearer, and we thank the reviewer for this comment. As for the x-axis, Adam and SGD also run as 10 independent chains, so every method performs the same 10 forward and backward passes per step; we will nonetheless plot gradient evaluations and wall-clock time instead of steps.

Thank you for your answers, they clarify my first earlier concern as well as my question on the x-axis. My concerns about Cor. 3.5 and presentation as well as W2/Q2 from Reviewer nDsh remain.

Reviewer_MmRZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a gradient-based swarm optimization method, based on a sotftmin energy functional JβJ_\beta. The softmin energy couples the evolution of each particle by scaling their respective gradient, while keeping the direction of the gradient unaffected. The contribution is mostly theoretical and establishes the connection between the stability of the ensemble dynamics and the local minimizer of the objective function (Theorem 2.1), the exponential convergence of the best-performing particle in the swarm for objective functions satisfying the PL condition, and improved exit times from local minima compared to independent Langevin. Overall, the paper is mostly well-written, delivers its stated objective, and the method seems novel. My main criticism is the effectiveness of the proposed approach in actual optimization problems, given its similar overall properties to independent gradient descent-type methods in terms of convergence.

Critical strengths
  1. The method seems novel.
  2. The main advantage of the method seems to be improved escape from local minima, which is studied theoretically in 1D and as well as experimentally.
Critical weaknesses
  1. It seems that, except in special cases, such as the double-well scenario, the method has similar convergence properties to independent Langevin, while being more expensive to run in practice due to the need to track a cluster of candidate solutions, as well as their interactions.
  2. The interaction among the population affects only the magnitude of the gradient of each particle. This seems rather restrictive compared to other swarm-based approaches.
Questions for authors
  1. Remark 3.6, which seems to be the main selling point of the method, could merit some revision to improve readability. Can the authors more explicitly state how the 1/β\beta factor helps with exit times in the proposed approach? Is there an explicit statement about how the method compares with independent Langevin?

  2. In the experimental evaluation, in Example 1, it seems that the cost of the Langevin method in Figure 1 scales much worse than the softmin method, which scales the best when well depth is large. While this constitutes some empirical proof that the coupled softmin method performs better than independent Langevin, it would be interesting to see a comparison for a fixed iteration budget.

  3. Please provide the corresponding reference for the PL condition in Definition 2.3.

1 comment

We thank the reviewer for the positive comments and the precise questions. About the weaknesses:

  • W1 For both our method and Langevin, we compute nn gradients and nn objective evaluations per step. The quantities that need to be computed specifically for the Softmin method are the Gibbs weights, which use values already evaluated, resulting in O(n)O(n) extra scalar operations. We will report wall-clock times to show that computation times are comparable: on the transformer at depth 50, a Softmin step takes 2.5 s against 2.0 s for Adam (App. E.4.6).

  • W2 The interaction term crucially changes the sign of the gradient, as well as its magnitude. This lets suboptimal particles perform gradient ascent and leads to the barrier-reduction result.

About the questions:

  • Q1 The remark fails to make clear the comparison between Softmin and independent Langevin. In the remark we will state the expected exit time for independent Langevin, which in the small-noise regime satisfies lim⁡σ→0σlog⁡Eτ=h\lim_{\sigma\to0}\sigma\log\mathbb E\tau=h. The 1/β1/\beta factor results from the quasipotential computation and represents the action that a particle has to pay to go uphill out of the basin, until it reaches the point where the sign reverses (due to the interaction term) and it is pushed by the gradient up to the saddle point ss. This computation (Theorem 3.4, parts 1–2) is what shows that the barrier hh is capped uniformly at order 1/β1/\beta at the quasipotential level, while for independent Langevin it is exactly hh.

  • Q2 The iteration budget is actually fixed, and the same for all methods. We let each method run with 100100 particles up to a limit of 10710^7 steps, with hyperparameters calibrated at the midpoint of the difficulty range (the x-axis). Figure 1 shows the success rate and the expected cost of each method. A run is considered failed when it does not reach the lower minimum within the limit; runs whose ensemble collapses or that stop improving for 5⋅1055\cdot 10^5 steps are stopped early and counted as failed. The expected running cost grows when a method loses reliability, which happens for all methods but Softmin as the potential barrier grows. This loss of reliability is what makes the Langevin method scale badly compared to Softmin. We will make all this more explicit in the main body, rather than leaving it to the appendix alone.

  • Q3 Thank you for pointing this out; we will cite Polyak (1963), Łojasiewicz (1963), and Karimi, Nutini & Schmidt (2016) in Definition 2.3.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.