Mechanistic interpretability has largely focused on language models and deterministic toy tasks. Much less is known about how sequence models internally represent latent stochastic dynamics under noisy, partially observed observations. We study this question in a controlled multivariate stochastic volatility setting, where models observe only returns while the ground-truth latent volatility state is known to the researcher. This setting provides a useful benchmark for mechanistic interpretability under partial observability: the latent state is hidden from the model but directly available for evaluation. Across architectures, losses, and output heads, we find evidence for a two-stage computation. Hidden representations encode substantial information about the next latent volatility state, and the output head maps this representation to squared return forecasts. Furthermore, in Transformers, latent-state decodability emerges at identifiable architectural stages whose location depends on the volatility period. In long-cycle regimes, this computation simplifies into an explicit latent-state filter consisting of a learned linear projection followed by $\ell^2$ normalization. Output-head replacement further shows that part of the degradation under noisy MSE training arises from readout misalignment rather than representation failure. These results suggest that stochastic volatility models provide a useful benchmark for mechanistic interpretability under noisy latent dynamics and partial observability.
We study robust generalization under spurious correlations: tasks where a shortcut feature is correlated with the true label in training but anti-correlated in an adversarial held-out split. Varying the spurious ratio $r$ (the fraction of training examples where shortcut = true label) and model capacity, we find a counterintuitive result: data imbalance promotes generalization in sufficiently capable models. On a synthetic task where the true label is sum parity of an integer sequence and the shortcut is the parity of the maximum-valued element, a 2-layer, 2-head transformer generalized (reached $100\%$ adversarial accuracy) in 0% of seeds at $r{=}0.50$ but 77% of seeds at $r{=}0.90$. The effect is absent in 1-layer models, where imbalance instead traps the model on the shortcut. Through mechanistic analysis -- gradient conflict dynamics, circuit evolution, and QK/OV circuit ablations -- we characterize a mechanistic pathway consistent with imbalance promoting generalization.
Mechanistic interpretability often evaluates explanations by intervening on a model: swapping hidden states, patching activations, ablating components, or comparing a compressed model to the original one. These experiments are usually summarized by a point estimate, even though the evaluation may be monitored while it runs or adapted toward suspected failures. This makes it hard to tell whether a reported fidelity or patching effect is a stable causal claim or a consequence of finite sampling and evaluation choices. We introduce Certified Interventional Fidelity (CIF), a statistical layer for interventional interpretability evaluations. CIF first writes the quantity being reported as a causal estimand: an expectation of a bounded score over a stated input distribution and a stated intervention distribution. It then provides confidence intervals and anytime-valid confidence sequences for this estimand, including under adaptive intervention sampling via bounded mixture importance weighting. We instantiate CIF with Hoeffding-style sequences and variance-adaptive betting sequences, the latter reducing certification cost by 10-30x in our experiments. On MNIST abstractions and GPT-2 Small IOI circuits, CIF certifies high-fidelity claims, shows when apparent method differences are not statistically supported, and makes sensitivity to the intervention distribution explicit.
Mriganka Basu Roy Chowdhury, Eric McLaughlin Weinercs.LG stat.ML
One of the major difficulties in the mechanistic interpretability of neural networks is the occurrence of polysemanticity, which suggests that each neuron is typically responsible for multiple different tasks, impeding a clean interpretation of their function. The seminal paper of Elhage et al. (2022) argues that this occurs due to superposition, a phenomenon where the neural network represents distinct features as non-orthogonal directions in a lower-dimensional space, a strategy that allows much greater compression of the data without sacrificing fidelity due to the feature sparsity of input vectors. Elhage et al. (2022) empirically validates these hypotheses in a rather natural and simple autoencoder with sparse inputs. The contribution of the present work is to analyze the mathematical basis for the occurrence and optimality of superposition, while rigorously corroborating some of their findings. In particular, we provide upper and lower bounds for the L2 reconstruction loss, tight in the very sparse regime, for power activation functions. A short list of interesting open problems are also included at the end.
We present the first causal mechanistic analysis of a tabular foundation model, investigating how TabPFN 2.5's feature wise attention heads distribute computation across layers. Using activation patching, ablation, and attention entropy across two synthetic regression datasets, we find clear temporal specialisation: one head's causal necessity dominates that of the others by 2 to 5 times at peak layer, with its dominant layer shifting across tasks of different complexity, while the remaining heads exhibit symmetric late layer profiles. Attention entropy and patching provide convergent evidence for the computationally active layers of the dominant head. We additionally investigate inference time steerability via contrastive activation steering, which fails to transfer across samples. We attribute this result to TabPFN's in context learning mechanism, which encodes task structure through context dependent attention rather than the stable parametric directions that make steering tractable in language models.
Marin Biloš, James T. Wilson, Anderson Schneider +1cs.LG
Tabular foundation models with different architectures converge in accuracy across a range of classification and regression tasks. This raises questions a leaderboard cannot answer: (i) whether the models execute the same in-context algorithm, (ii) where row, column, and class-permutation invariances originate, and (iii) how robust they are under perturbations engineered against the inferred mechanism. We characterize all three. The model families realize qualitatively distinct similarity-based readouts: from an attention-weighted vote over context labels to a class-conditional mean readout, each confirmed by causal intervention. We find that the representation collapse highlighted in prior work is not a practical concern for them. Each model's permutation invariances trace to specific positional parameters whose removal preserves accuracy and makes approximate invariance exact. Perturbations engineered against each readout reproduce predicted failure modes; hub and rank attacks isolate them from refit baselines. Together these results give a mechanistic account of contemporary tabular foundation models and identify which inductive biases govern both their accuracy and characteristic failures.