Auditable Routing for Mixtures of Concept Experts

Some interpretable models predict in two stages. A network first detects human-readable concepts such as "bone spurs" or "wing color," and then a small set of readable formulas, called experts, maps those concepts to the answer. A separate component, the router, decides which formula applies to each input. The weak point of these models is the router. In the most capable recent designs it is an ordinary opaque neural network, and in some it sees the raw image rather than only the concepts. A user can therefore read the formula that was applied but cannot see why that formula was chosen. A router like this could quietly carry most of the predictive work, which would make the readable formulas a veneer.

We propose to make the whole post-concept stage conditionally Gaussian, so that it can be fitted with exact, simple Bayesian updates. The router becomes a sparse linear function of the concepts. It is joined by a single "residual" channel for information the concepts miss, and a shrinkage prior decides whether that channel is needed at all. The posterior size of the channel then directly measures how much routing depends on things a human cannot see. The same machinery prunes redundant experts and absorbs noise in the concept predictions. The method rests on one old trick, the scale-mixture representation: build heavy-tailed or sparsity-inducing distributions out of Gaussians whose variance is itself random. That trick keeps all of these pieces tractable.

Core Claim

In a mixture of concept experts, routing can be restricted to a sparse, concept-level rule plus one explicitly measured escape hatch without losing the accuracy of opaque routing. The posterior on that escape hatch is a faithful estimate of routing leakage, meaning prediction-relevant routing information that bypasses the concepts. Neither current architecture provides this estimate. Mixture of Concept Bottleneck Experts (M-CBE) routes concepts to linear or symbolic expert formulas through a dense selector network, and chooses the number of experts by sweeping and picking a "Pareto knee." Interpretable Mixture of Experts (IME) forces a choice between an interpretable assignment module, which it reports as having lower capacity, and a deep one, which is opaque. It prevents expert collapse with hand-weighted diversity and utilization losses. Our proposal replaces the selector sweep and the heuristic losses with one probabilistic model whose sparsity pattern is the explanation of the routing.

The Model

We work on top of frozen concept predictions $c \in \mathbb{R}^k$, trained sequentially as in Concept Bottleneck Models. We also use a low-dimensional residual feature $h \in \mathbb{R}^d$: a projection of the backbone embedding, orthogonalized against $c$, with $d$ around 16. Input $i$ is routed to expert $m$ among $M$ candidates with probability

$$P(z_i = m \mid c_i, h_i) \propto \exp\big(a_m^\top c_i + \gamma\, b_m^\top h_i\big),$$

where $a_m$ are concept-level gate weights, $b_m$ are residual gate weights, and $\gamma \ge 0$ is a single global scale on the residual channel. Given the route, the output follows

$$y_i \mid z_i = m \;\sim\; \mathcal{N}\big((w_0 + \delta_m)^\top c_i,\; \lambda_i \sigma^2\big),$$

where $w_0$ is a shared expert applied to every input and $\delta_m$ is expert $m$'s correction to it. $\lambda_i$ is a per-input variance with an inverse-gamma prior, which makes the noise Student-$t$ and lets an input with badly predicted concepts inflate its own variance instead of dragging the experts. $\sigma^2$ is the base noise level.

Horseshoe priors are sparsity priors that zero out small coefficients while leaving large ones nearly untouched. We place them on $a_m$, $\delta_m$ and $\gamma$. They are sampled with the auxiliary inverse-gamma representation of A Simple Sampler for the Horseshoe Estimator, which makes every conditional update a standard Gaussian or inverse-gamma draw. The gate uses the Pólya-Gamma augmentation that the same authors apply to sparse logistic regression. This augmentation is itself a scale mixture of normals, and extending it to multi-class gating means updating one expert's gate at a time.

Each prior does a specific job. The shared-plus-correction structure borrows the division of labor found empirically by Decoding Knowledge Attribution in Mixture-of-Experts, where shared experts carry invariant logic and routed experts refine it. We turn that observation into a prior. It has a statistical motive. On Minimax Estimation of Parameters in Softmax-Contaminated Mixture of Experts proves that when experts nearly coincide, gating estimates degrade from $\tilde O(n^{-1/2})$ to $\tilde O(n^{-1/2}\|\Delta\|^{-2})$. Shrinking $\delta_m$ to exactly zero merges near-duplicate experts into the shared one before they make the router unidentifiable. The expert count is then read off the posterior rather than swept.

Bayesian Regression Analysis with scale mixtures of normals constrains the design. It shows that under flat priors a sampler can appear to converge while the posterior mean does not exist, when the data per regressor is thin ($n$ near $k$). Experts with few assigned inputs are exactly that regime. We therefore use only proper priors and report posterior medians alongside means.

Every conditional is Gaussian or inverse-gamma, so the same updates give a closed-form mean-field variational version. We use that version when a Gibbs sampler is too slow.

Research Program

The work has three parts. First, we test whether leakage is real in existing designs. We train M-CBE's linear variant with its image-fed selector, then replace every expert with a constant (intercept only) and measure how much accuracy survives. If a large fraction survives, the selector is doing the task and the expert formulas are not the explanation. This test is cheap, and a positive result would reframe how mixtures of concept experts are evaluated.

Second, we build the model above on the concept-level heads of existing concept predictors. We compare Gibbs and variational fitting for speed and for agreement of the fitted model.

Third, we validate $\gamma$ as a leakage meter on planted problems where the truth is known. We construct variants of dSprites-Exp and Pendulum in which the regime that selects the formula is either an annotated concept or an unannotated generative factor. The posterior of $\gamma$ should concentrate near zero in the first case and away from zero in the second. Its magnitude should track the accuracy lost when $h$ is removed.

Baselines

The naive reference points come from Concept Bottleneck Models, which fits a single linear head on concepts, and from the end-to-end black box. On CUB-200 that paper reports a 19.9% error for the joint CBM and 17.5% for the black box. Auxiliary Losses for Learning Generalizable Concept-based Models (coop-CBM with an orthogonality loss) reports 84.1% CUB accuracy against 82.3% for its black box. We quote these numbers but rerun the baselines, because our heads sit on frozen concepts and the backbone must be identical across every method.

The strongest competitor is Linear M-CBE. It is the closest existing design and has the highest reported accuracy among readable-expert models. The file reports only that it beats logic-based concept models by "up to 65%" on CUB. There is no absolute accuracy for Linear M-CBE on Lacuna for CUB, OAI, or the synthetic tasks, so we will rerun it. We rerun it in two forms, once with its selector fed the image and once fed concepts only. Both forms receive the same concept predictor, the same range of $M$, and the same hyperparameter-search budget in GPU hours as our model.

We also run IME's interpretable-router variant with its diversity and utilization losses. On tabular tasks we run Covariate-Guided Clusterwise Linear Regression, whose paper reports RMSE competitive with XGBoost and CatBoost on BIKE and ELECTRICAL. The file gives no specific RMSE values for it on Lacuna, so those numbers come from our reruns.

To isolate the key ingredients we run three ablations. The first sets $\gamma=0$, giving a concept-only router. The second replaces the horseshoe on $a_m$ and $\delta_m$ with an L1 penalty tuned by cross-validation. The third replaces Student-$t$ noise with Gaussian noise.

We say we have beaten M-CBE if our method matches the image-fed M-CBE's accuracy within its seed-to-seed spread, using a router with far fewer nonzero coefficients. The router must also stay faithful under the constant-expert ablation.

Why not just…?

Why not just use an L1-sparse softmax router over concepts and sweep the number of experts? This is the most serious threat. On clean data it may well match our accuracy, because the concepts are fixed and a lasso-gated mixture is a strong model. The horseshoe's advantage, that it leaves large coefficients unshrunk, is documented for regression by A Simple Sampler for the Horseshoe Estimator. Whether that advantage matters for gates is untested. We therefore run the L1 router as a named ablation. If it ties on accuracy and router size, the contribution narrows to what L1 cannot provide: the posterior leakage estimate from $\gamma$, calibrated uncertainty over which expert applies, and automatic expert merging. We would then state the contribution that narrowly.

Why not just feed M-CBE's selector the concepts instead of the image? That change removes most of the leakage, but the router stays an opaque network over concepts, and the user still cannot state the rule that picks an expert. We include this variant. If it beats our sparse linear gate by a clear margin, that margin is the measured price of a readable router, which is itself a useful result.

Why not just explain the selector afterward with SHAP or a surrogate tree? In the user study of Interpretable Mixture of Experts, participants given post-hoc SHAP were 27% less accurate at predicting model behavior than those given built-in explanations. RouterInterp finds that learned routers pack 7 to 8 unrelated semantic clusters into each expert, which is exactly what a post-hoc summary would have to untangle. We will also fit a depth-limited surrogate tree to M-CBE's selector and report how often it disagrees with the actual routing on held-out inputs.

Why not use Bernoulli concept indicators, as in Sparse Linear Concept Discovery Models? That model sparsifies a single global linear layer using relaxed binary indicators. It reaches 95.30% on CIFAR-10 and keeps 6.96% of concepts on ImageNet. It has no routing, so it cannot represent multiple regimes. Its discrete indicators also need gradient relaxations, whereas scale mixtures give exact conditional updates. We compare it to our model with $M=1$ on its label-free CIFAR-10 setup, using its released code.

Proposed Experiments and Decisive Checks

We use four families of benchmarks. The first is the M-CBE synthetic tasks: Pendulum, where M-CBE recovered $8.00\cdot\sin(\theta)+10.00$, plus MNIST-Arith and dSprites-Exp, with our planted-regime variants added. The second is OAI knee radiographs, where Concept Bottleneck Models reports that correcting two concepts cuts error from over 0.4 to about 0.3. The third is CUB-200 and CUB+Places, where the black box's reported error rises to 62.7% under background shift. The fourth is the CG-CLR tabular suite and Rossmann.

We measure task error, router size in nonzero gate coefficients, the constant-expert ablation score, and the posterior of $\gamma$. We also measure the error reduction when concepts are corrected by hand, whether a correction changes the chosen expert in the intended way, and accuracy under injected concept noise.

Five results would decide the proposal. The first is whether the constant-expert ablation exposes leakage in image-fed M-CBE; if it does not, the leakage motivation fails. The second is whether the posterior of $\gamma$ separates planted annotated regimes from unannotated ones. The third is whether our model matches M-CBE's accuracy with a readable router and no leakage. The fourth is whether it beats the L1 ablation on anything beyond uncertainty estimates. The fifth is whether Student-$t$ noise preserves accuracy under concept noise better than Gaussian noise does.