wikillm

Paper · 2026-10-30 · stat.ML, cs.LG

WASABI: Wasserstein Sampling via Bridge Interpolants

WASABI trains a diffusion sampler for a Boltzmann target known only through its energy and gradient. Each step turns a Wasserstein gradient-flow update of the KL divergence into a regression on the denoiser, using score-free mean and tilted-noise covariance estimators. Author-reported: alanine dipeptide JS 0.0254 (best baseline 0.068) with about 3.1M energy calls against 10.24M for the baselines.

Diffusion modelsLangevin dynamicsOptimal transportdiffusionsamplingboltzmann-distributionwasserstein-gradient-flowmolecular-simulation

arXivPDFMarkdown17 min read

AI-compiled by claude-haiku-5-5 on 2026-11-21, not yet reviewed by a human. Every claim cites its source; check it when a detail matters.

In plain words

Some scientific problems ask for typical configurations of a molecule or a cluster of particles, where how likely a configuration is depends on an energy score that is cheap to compute for a single configuration but expensive to average over many. Classic simulators walk through configurations one small step at a time, which is slow. WASABI trains a generative model instead, so that after training it can produce new configurations directly, without asking the energy function again.

The training idea borrows from Langevin dynamics, which nudges a configuration downhill on the energy while adding noise. WASABI applies the same kind of nudge to the whole distribution of the model's outputs, and then rewrites that nudge as a regression target the model can learn from. The difficult part is that the nudge needs the gradient of the model's own density, which is hard to get from a diffusion model. The authors show how to build the training target from samples the model has already produced and from the energy's gradient, so that the model's own density gradient never has to be computed.

TL;DR

  • WASABI trains a diffusion model to sample a Boltzmann distribution from its energy function and the energy's gradient, without samples from the target §1 §2.
  • Each outer iteration takes a Wasserstein gradient-flow step on KL(q‖π) and writes it as a regression for a correction to the denoiser, using score-free mean and covariance estimators §3.1 §3.2 §3.3.
  • Author-reported on alanine dipeptide: Jensen–Shannon divergence 0.0254 ± 0.0033 against 0.068 ± 0.009 for the best baseline, with a budget of about 3.13 M energy calls against 10.24 M for ASBS and BMS tab. 3.
  • On the particle benchmarks the results are mixed: lowest mean energy W2 on DW-4, but behind BMS on LJ-13 tab. 1 tab. 2.
  • No official code was found. Several choices that matter for reproduction (noise scale, time schedules, learning rate, optimiser) are not stated.

Problem

The target is a Boltzmann distribution π ∝ exp(−U), where U is an energy function and only U and ∇U can be queried, with no samples from π §1. MCMC and molecular dynamics need many energy or gradient evaluations to produce representative samples, and the sampled chains can take a long time to move between likely regions §1. Amortised samplers spread the training cost over samples they later produce without further energy calls §1.

The method targets reverse KL, KL(q‖π), where q is the sampler's endpoint distribution §1 §3. The Wasserstein gradient of this objective is ∇log q + ∇U §2, and a diffusion sampler's own score ∇log q is generally unavailable §1 §3.1. The paper's contribution is an algorithm that never evaluates that score §3.2.

Method

Background used by the method

A sampler is defined by the SDE dX_t = (m_t(X_t) − X_t)/(1 − t) dt + B dW_t, started from X_0 ~ p_0, for 0 ≤ t < 1 §2. Here p_0 is a source distribution that is easy to sample from. The coefficient B, which must be positive, scales the noise, and W is a Wiener process §2.

For endpoint samples X̃_1 ~ p_1 and independent source X̃_0 ~ p_0 and noise Z ~ N(0, I), the bridge observation is Y_t = (1 − t) X̃_0 + t X̃_1 + h_t Z, with h_t = B √(t(1 − t)) §2. The denoising representation of p_1 is the conditional mean m̄_t(y) = E[X̃_1 | Y_t = y] §2. A neural denoiser m_ω(y, t) is fitted by regressing X̃_1 on Y_t §2.

Denoiser update induced by a Wasserstein step

Let q_{n,η} be the endpoint distribution after moving each point by −η∇_W2 F[q_n], where F = KL(·‖π) §3.1. A first-order expansion in η gives the change of the denoiser §3.1, Theorem on denoiser change:

  • Mean term: M_tq = E[−∇U(X̃_1) − ∇log q(X̃_1) | Y_t = y], the average movement of the endpoints §3.1 §3.2.
  • Covariance term: C_tq = (t/h_t) Cov(X̃_1, −Z·∇_W2Fq | Y_t = y), which accounts for endpoints becoming more or less likely to produce the same observation §3.1.

The full first-order change is M_t + C_t. The paper states the update T[m̄]_t = m̄_t + η_t (M_t + λ_t C_t) §3.1. With λ_t = 1 and a constant η_t it matches the desired transport to first order; for any finite λ_t, π is a fixed point because both terms vanish at q = π §3.1.

Score-free estimators

  • Mean estimator: Stein's identity turns M_t into a conditional mean of −∇U(X̃_1) + (t/h_t) Z, so the model score is not needed §3.2 app. A.3. Regressing this target on (Y_t, t) learns M_t §3.2.
  • Covariance estimator: the inner product E[φ(Y_t)·C_t] is estimated with a tilted noise draw Z̃ §3.2 app. A.4. With F_φ(y, x) = φ(y)·(x − m̄(y)) and a(x) = (h_t / 2t) s_π(x), where s_π = −∇U, the estimator is Ĉ_t[φ] = (t/h_t) φ(Y)·Z + (t²/h_t²)(d + ‖a(X̃_1)‖²)[F_φ(Ỹ, X̃_1) − F_φ(Y, X̃_1)], with Ỹ built from Z̃ instead of Z §3.2 app. A.4. It is unbiased when the denoiser m̄ is the true conditional mean and boundary terms vanish §3.2.
  • Tilted noise: the density φ̃(z | x) = ‖z + a(x)‖² φ(z) / (d + ‖a(x)‖²) is normalised §3.2 app. A.4. The appendix gives an exact sampler for it: draw G ~ N(0, I) and H_1, H_2 ~ N(0, 1); set V = G with probability ‖a‖²/(d + ‖a‖²), otherwise V = √(1 + (H_1² + H_2²)/‖G‖²) G; then keep V with probability 1/2 + V·a/(‖V‖² + ‖a‖²) and output −V otherwise app. A.4.

Training objective

The regression target is r_t = M_t + λ_t C_t §3.3. Expanding the squared-error loss and dropping constants gives a computable objective in the correction f_θ §3.3 app. A.6:

\mathcal L_n(\theta) = \mathbb E\left[\tfrac12\|f_\theta(Y_t,t)\|^2 - f_\theta(Y_t,t)\cdot\left(-\nabla U(\tilde X_1)+\tfrac{t}{h_t}Z\right) - \lambda_t\hat{\mathcal C}_t[f_\theta(\cdot,t)]\right] §3.3 app. A.6

The expectation is over bridge time, source and endpoint samples, noise and the tilted noise Z̃ §3.3. The denoiser m_{ω_n} is held fixed while the correction is fitted §3.3.

WASABI algorithm

The paper's main listing is Algorithm 1 alg. 1. Its outer loop is written below in our own pseudocode, which follows the same steps.

Pseudocode
# WASABI outer loop  [alg. 1] [§3.3]
# Inputs: grad_U (target energy gradient), sample_prior (draws from p0),
#         B (diffusion coefficient), N (outer iterations),
#         D0 (initial endpoint samples), eta(t), lam(t) (schedules)
def wasabi(grad_U, sample_prior, B, N, D0, eta, lam):
    # Fit the initial denoiser on D0 by denoising regression  [§2] [§3.3]
    denoiser = fit_denoiser(D0)
    for n in range(N):
        # Freeze the denoiser and fit the correction f on the loss  [§3.3]
        correction = fit_correction(denoiser, grad_U, B, lam, sample_prior)
        # Corrected denoiser: m~ = m_w + eta_t * f  [§3.3] [alg. 1]
        corrected = lambda y, t: denoiser(y, t) + eta(t) * correction(y, t)
        # Canonicalise: draw fresh endpoints from the corrected sampler  [§3.3]
        fresh = simulate_sampler(corrected, sample_prior, B)
        # Refit the denoiser on the fresh endpoints for the next iteration  [§3.3]
        denoiser = fit_denoiser(fresh)
    return denoiser

The sampler itself is simulated with the SDE of §2. The discretisation scheme is not stated; the experiments use 128 sampling steps for LJ-13 §5.2 and 256 integration steps for ALA2 app. B.4.

Pseudocode
# One training sample for the correction loss  [§3.2] [§3.3] [app. A.4]
# x0 ~ p0, x1 ~ replay buffer, t ~ bridge-time distribution, z ~ N(0, I)
def correction_loss_sample(f, denoiser, x0, x1, grad_U1, t, z, B, lam):
    h = B * sqrt(t * (1 - t))                    # h_t  [§2]
    y = (1 - t) * x0 + t * x1 + h * z            # bridge observation  [§2]
    # Mean regression target  [§3.2]
    mean_target = -grad_U1 + (t / h) * z
    # Tilt vector a = (h / 2t) * s_pi(x1), with s_pi = -grad U  [§3.2]
    a = -(h / (2 * t)) * grad_U1
    z_tilde = sample_tilted_noise(a)             # Lemma, [app. A.4]
    y_tilde = (1 - t) * x0 + t * x1 + h * z_tilde
    # Score-free covariance estimate, denoiser held fixed  [§3.2] [app. A.4]
    residual_tilde = x1 - denoiser(y_tilde, t)
    residual = x1 - denoiser(y, t)
    weight = (t / h) ** 2 * (d + dot(a, a))
    cov_estimate = (t / h) * dot(f(y, t), z) + weight * (
        dot(f(y_tilde, t), residual_tilde) - dot(f(y, t), residual))
    # Per-sample loss; minimised over the correction f  [§3.3]
    return (0.5 * dot(f(y, t), f(y, t))
            - dot(f(y, t), mean_target)
            - lam * cov_estimate)
Pseudocode
# Tilted noise with density proportional to ||z + a||^2 phi(z)  [app. A.4]
# a has shape (d,); d is the dimension
def sample_tilted_noise(a):
    g = standard_normal(d)                       # G ~ N(0, I)
    h1 = standard_normal()                       # H_1 ~ N(0, 1)
    h2 = standard_normal()                       # H_2 ~ N(0, 1)
    # Choose between the plain and the radially stretched draw  [app. A.4]
    if uniform() < dot(a, a) / (d + dot(a, a)):
        v = g
    else:
        v = sqrt(1 + (h1 * h1 + h2 * h2) / dot(g, g)) * g
    # Sign choice, probability 1/2 + v.a / (|v|^2 + |a|^2)  [app. A.4]
    p_keep = 0.5 + dot(v, a) / (dot(v, v) + dot(a, a))
    if uniform() < p_keep:
        return v
    return -v

The ALA2 variant from the appendix uses an extended listing, Algorithm 2 alg. 2, with three changes app. A.7:

  • The source is coupled to the endpoint, X_0 = A(G, X_1), where A is the Kabsch alignment of the prior sample G onto X_1 app. A.2 §5.3. This changes the mean target to −∇U + J_t^T Z / h_t and the covariance estimator to the general coupling form app. A.3 app. A.5.
  • The target score is −β_n ∇U, with β_n set by the annealing schedule app. A.7 app. B.3.
  • Replay: endpoints are kept in a ring buffer with their cached ∇U values; a partial refresh each iteration adds fresh endpoints app. A.7 §5.3.

For the Kabsch case the Jacobian is J_t = tI + (1 − t) T H^{-1} T^T, with H = Σ_i [(a_i·x_i) I − a_i x_i^T] and T ω = (ω × a_i)_i app. A.2 app. A.3. The covariance estimator then uses the symmetric part S_t of J_t, which is positive definite under the stated invertibility assumption app. A.5 app. A.2, Corollary.

BASH readout for the ALA2 correction

For ALA2 the correction vector field is not a plain network output. It is split into bond-length and bond-angle directions using a regularised projector built from the Jacobian of the chosen bond lengths and cosines app. B.3.

Pseudocode
# BASH readout for the ALA2 correction  [app. B.3]
# x: (N, 3) zero-centroid coordinates; J: Jacobian of bonds and cosines
def bash_readout(x, t):
    # Jacobian of the bond lengths and bond-angle cosines at x  [app. B.3]
    J = bond_angle_jacobian(x)                   # (|E| + |A|, 3N)
    # Regularised projector onto the bond-angle directions, delta = 1e-4  [app. B.3]
    P = J.T @ inverse(J @ J.T + 1e-4 * I) @ J    # (3N, 3N)
    # Two separate PaiNN heads, combined through their projections  [app. B.3]
    stiff = sigma_stiff * (P @ head_stiff(x, t))
    soft = sigma_soft * ((I - P) @ head_soft(x, t))
    return stiff + soft

The values of sigma_stiff and sigma_soft are not stated in the text the card is built from. The paper notes that the split is not an exact orthogonal decomposition and that the chirality restraint acts only through the target energy and score app. B.3.

Hyperparameters

Values as stated in the paper. "Not stated" means the text the card is built from does not give the value.

NameValueWhereNotes
Diffusion coefficient Bnot stated§2 §3.3Symbol only in the experiments
Bridge-time distributionnot stated§3.3The range and sampling of t are not given
Step schedule η_tnot stated§3.3Experiments report only a scalar step scale s_η
Covariance weight schedule λ_tnot stated for main runsapp. B.4Constant λ_cov in the sweeps
Covariance weight λ_cov, ALA2sweep 0 to 1; lowest mean at 0.4; 0.25 in step sweepapp. B.4 tab. 5Main runs: see note below
Step scale s_η, ALA2sweep 0.10 to 1.00; lowest mean at 0.10app. B.4 tab. 5Relation to η_t not defined in the text read
Outer iterations, DW-41,000 rounds§5.1 app. B.1Same for all methods
Network, DW-416×4 EGNNtab. 1The paper does not define the two numbers
Seeds, DW-46 (damped BMS: 5 after excluding one divergent run)tab. 1 app. B.1
Endpoints per seed, DW-416 checkpoints × 128 endpoints, every 25 rounds from 625 to 1,000app. B.12,048 endpoints per seed
Potential, DW-42D, 4 particles, pairwise double wellapp. B.2Pair minima at about 2.51 and 5.49 §5.1
Network, LJ-13centre and correction PaiNN, 4 layers, 64 channels§5.2
Prior, LJ-13Gaussian, σ_0 = 1.75§5.2
Outer iterations, LJ-137,000§5.2
Sampling steps, LJ-13128§5.2
Seeds, LJ-133§5.2
Potential, LJ-1313 particles in 3D, Lennard-Jones plus quadratic confinementapp. B.2
Outer iterations, ALA24,000§5.3 app. B.3
Annealing, ALA2β_n interpolates 1,000 K to 300 K over the first 1,500 iterations, then fixedapp. B.3Replay samples reweighted as β changes
Replay buffer, ALA2ring of 4,096 endpoints; 256 refreshed per iteration (6.25%)§5.3 app. B.3
Larger replay (comparison)8,192 capacity; 512 refreshed per iteration§5.3 app. B.4Matched seeds 0–2 at iteration 2,500
Prior bank, ALA220,000 configurations at 1,000 K from a seven-temperature REMD runapp. B.3Used for the reported training runs
Alternative prior, ALA220,000 configurations from local MD at 1,000 Kapp. B.3Tested; agreement with REMD bank reported
Network, ALA2one PaiNN per head, 4 layers, 64 channels, BASH readoutapp. B.3 app. B.4
Projector regulariser δ, ALA21e-4app. B.3
Batch size, ALA264app. B.4
Updates per outer iteration, ALA2100 correction, 200 denoiserapp. B.4
Fresh endpoints per iteration, ALA2256, each produced with 256 integration stepsapp. B.4
Denoiser warm-up, ALA210,000 updatesapp. B.4Footnote to the cost count
Learning rate, optimisernot stated—Not given for any system
Evaluation samples, ALA21,024,000 per seed (table); 20,000 per seed (sweeps)tab. 3 app. B.4

Note on the main ALA2 settings: the paper does not state the λ_cov and s_η used for the headline numbers. The headline JS 0.0312 ± 0.0023 at 20,000 samples, which the paper says falls to 0.0254 ± 0.0033 with 1,024,000 samples, matches exactly the s_η = 0.10 row of the step sweep, where λ_cov = 0.25 app. B.4 tab. 5. This match is our inference, not a stated setting.

Results

Author-reported. Baselines are the paper's own reruns on DW-4 and quotes from Blessing et al. on LJ-13 and ALA2 §5 app. B.1. Lower is better for all metrics.

DW-4 (2D, four particles), six seeds, 1,000 rounds tab. 1 fig. 2

MethodConfig. W2Energy W2
ASBS0.94 ± 0.162.35 ± 0.73
ASBS (damped)0.97 ± 0.302.31 ± 0.34
BMS1.27 ± 0.192.87 ± 1.23
BMS (damped), five seeds1.48 ± 0.212.72 ± 0.58
WASABI1.13 ± 0.391.89 ± 0.43

The table excludes one divergent damped-BMS run (energy W2 2,449). If that run is kept, the damped-BMS row changes: its energy W2 becomes a mean of 410 with SD 999 app. B.1.

LJ-13 (13 particles in 3D), three seeds, 7,000 outer iterations tab. 2 fig. 3

MethodConfig. W2Energy W2
AS1.67 ± 0.012.40 ± 1.25
ASBS1.59 ± 0.031.99 ± 1.01
BMS1.54 ± 0.000.53 ± 0.04
WASABI1.597 ± 0.0031.677 ± 0.224

Alanine dipeptide at 300 K, three seeds tab. 3 fig. 4

MethodTrain (TFLOPs / outer iter.)Energy calls (millions)JS divergence (nats)
ASBSnot reported10.24 (inferred)0.232 ± 0.018
ASBS (damped)not reported10.24 (inferred)0.111 ± 0.017
BMS112.9 (inferred)10.24 (inferred)0.311 ± 0.050
BMS (damped)190.2 (inferred)10.24 (inferred)0.068 ± 0.009
WASABI31.73.130.0254 ± 0.0033
REMD (reference protocol)not reported73.5not reported

Baseline costs are inferred by the authors from an assumed eight-GPU setup and the released BMS code tab. 3 footnote app. B.4. WASABI's energy count is a budget for a locally generated prior bank, not a measured cost of the reported REMD-prior runs app. B.4.

ALA2 hyperparameter sweeps, 20,000 samples per seed, three seeds tab. 5 fig. 5

λ_cov (s_η = 1.00)JSs_η (λ_cov = 0.25)JS
0.00000.0595 ± 0.02120.100.0312 ± 0.0023
0.06670.0520 ± 0.01250.160.0387 ± 0.0218
0.13330.0479 ± 0.01370.220.0408 ± 0.0173
0.20000.0377 ± 0.00610.280.0512 ± 0.0167
0.26670.0426 ± 0.02370.340.0332 ± 0.0020
0.33330.0416 ± 0.00470.400.0370 ± 0.0135
0.40000.0365 ± 0.00610.460.0422 ± 0.0180
0.46670.0475 ± 0.02160.520.0479 ± 0.0139
0.53330.0613 ± 0.01350.580.0400 ± 0.0173
0.60000.0782 ± 0.02240.640.0409 ± 0.0145
0.66670.0681 ± 0.00790.700.0485 ± 0.0299
0.73330.0822 ± 0.01940.760.0342 ± 0.0111
0.80000.0808 ± 0.01020.820.0456 ± 0.0089
0.86670.1117 ± 0.00730.880.0384 ± 0.0179
0.93330.0918 ± 0.00620.940.0449 ± 0.0185
1.00000.1562 ± 0.04351.000.0441 ± 0.0095

The paper's lowest means are λ_cov = 0.4 and s_η = 0.10 app. B.4. The replay comparison at s_η = 0.20 and λ_cov = 0.25 is shown in fig. 5 and is not tabulated app. B.4.

Figures, in words. fig. 1 is a schematic of the procedure: endpoint distributions moving toward π, and the denoiser being updated at each step. fig. 2 shows pooled pair distances for DW-4, with the two minima of the pair potential marked, next to the DW-4 table. fig. 3 shows energy and pair-distance densities for LJ-13, with WASABI underrepresenting the low-energy shoulder §5.2. fig. 4 shows Ramachandran free-energy maps for 20,000 WASABI and 20,000 REMD samples, next to the ALA2 table. fig. 5 shows the sweeps and the replay comparison, where increasing replay capacity recovers a smaller central peak in one highlighted seed app. B.4.

Paper vs code

No official code.

The paper links only the released Bridge Matching Sampler codebase, used for the baseline cost inference app. B.4, and the FAB and Zenodo datasets app. B.4. No project page is linked. GitHub searches for "wasabi", "WASABI wasserstein", "bridge interpolants diffusion", "wasserstein boltzmann diffusion" and the organisation blei-lab returned no matching repository. Searches for two author user names returned errors. Nothing from the BMS or ASBS repositories is used for WASABI.

Pitfalls

  • Score sign. The tilt vector is a(x) = (h/2t) s_π(x) with s_π = −∇U §3.2 app. A.4. A flipped sign changes which noise draws are favoured in the covariance estimator. The annealed target is −β_n ∇U app. A.7.
  • Denoiser must be the conditional mean. The covariance estimator is unbiased only if the frozen denoiser equals E[X̃_1 | Y_t] §3.2. The fitted denoiser is an approximation, so the loss has an extra approximation error. The paper says so explicitly app. A.6.
  • Time-dependent scale factors. The factors t/h_t and t²/h_t² equal √t/(B√(1−t)) and t/(B²(1−t)) §2 §3.2. They grow without bound as t → 1 and the drift in the SDE has a 1/(1 − t) term §2. The bridge-time range and the sampler's last step are not stated, so the implementation must choose them.
  • Division by the Gaussian norm. The tilted-noise sampler divides by ‖G‖ app. A.4. Its factor becomes large when ‖G‖ is small. The paper gives no guard. This is our reading of the formula; the effect in the low-dimensional DW-4 case is untested in the text.
  • Replay makes the update approximate. Stored endpoints come from earlier samplers, so the training distribution only approximates q_n app. A.7. The paper says this reuse follows ASBS and BMS and that samples match the targets empirically app. A.7.
  • Prior and replay mismatch. The reported ALA2 runs use the REMD-derived prior bank app. B.3. A local MD bank gives backbone-angle JS of 0.0309 and 0.0307 to the two REMD banks, while the two REMD banks differ from each other by 0.0296 app. B.3. Other systems' priors are not described beyond the Gaussian for LJ-13 §5.2.
  • Sample count changes the JS. Sweeps use 20,000 samples per seed; the headline uses 1,024,000 app. B.4. The text says "one million" in one place §5.3 and 1,024,000 in another app. B.4. The lower value is partly sampling noise app. B.4.
  • Low JS does not mean full mode coverage. On ALA2 a smaller backbone-angle peak can be underrepresented while JS is low. Larger replay and refresh counts help in the evaluated runs but do not guarantee coverage §5.3 §6.
  • Energy-call budget. The 3.13 M count includes 2 M production and 0.1 M equilibration steps for a local prior bank and 1.03 M training calls tab. 3 footnote. The reported REMD-prior runs are not charged for their prior app. B.4. The abstract's "3.3 times fewer" follows from 10.24 M over 3.13 M app. B.4 §1. The introduction states the same gain as "30%" of baseline evaluations §1.
  • Baseline costs are inferred. The BMS and ASBS costs come from an assumed eight-GPU configuration and a released default of 5,000 iterations; with the released default of 7,000 iterations the energy count would be 14.336 M app. B.4.
  • Excluded run on DW-4. One damped-BMS seed diverged (energy W2 2,449). The reported damped-BMS row drops it; with all six seeds the energy W2 becomes 410 ± 999 instead of 2.72 ± 0.58 app. B.1 tab. 1.
  • Unstated optimiser and noise schedule. The learning rate, optimiser, bridge-time distribution, diffusion coefficient and the η_t and λ_t schedules are not stated for any system. An implementation has to choose them and should report the choice.
  • Covariance weight. The paper reports that covariance weights below 0.5 generally work better on ALA2, and that the mean is lowest at 0.4 app. B.4 §5.3. The main setting is not stated; see the note under Hyperparameters.
  • Mode coverage in LJ-13. The energy histogram puts too little mass on the low-energy shoulder, and BMS is better on both metrics §5.2 §6.
  • First-order update. The derivation is first order in the step size, and finite fitting and replay add further approximations. The paper's fixed-point claim is for the ideal update §3.1 §6.

Lineage

  • Builds on: Wasserstein gradient flows and Langevin dynamics for the target-KL flow §2 §1; denoising-representation bridges and denoiser regression §2; the replay and annealing practices of ASBS and BMS app. A.7.
  • Compared against: Adjoint Sampling (AS), Adjoint Schrödinger Bridge Sampler (ASBS), Bridge Matching Sampler (BMS), and REMD as a reference protocol §5 tab. 3. These cards are not in this wiki and are named without content from them.
  • Superseded by: not known from content already in this wiki.

Source and license

Adapted from “WASABI: Wasserstein Sampling via Bridge Interpolants” by Sebastian Salazar, Dilum Fernando, Michal Kucer, Emily Casleton, Luhuan Wu, David M. Blei (arXiv:2610.41958), used under CC BY 4.0. Changes: condensed and restructured for implementation, with tables, equations and algorithms taken from the paper and its code, each cited to its source. Not affiliated with or endorsed by the authors.