Interpretability & Mechanistic Interp - Weekly Roundup

Audio embeddings capture physical damping properties far better than single visual frames, and this gap is partially bridgeable through audio-to-vision distillation.

Language model states are not static points but possess dynamic, context-adaptive reference frames that can be approximated to achieve near-perfect interpretability and steerability.

Neural networks preferentially learn low-frequency sinusoidal patterns due to the inherent graded structure of their gradient descent dynamics, explained by an orthogonal polynomial framework.

Concept interventions can actively worsen model predictions, and existing "benign leakage" criteria fail to prevent this, necessitating new training objectives.

Specialized experts in MoE LLMs, not dense model neurons, are the key to interpretable and controllable personalization.

Audio embeddings capture physical damping properties far better than single visual frames, and this gap is partially bridgeable through audio-to-vision distillation.

Language model states are not static points but possess dynamic, context-adaptive reference frames that can be approximated to achieve near-perfect interpretability and steerability.

Neural networks preferentially learn low-frequency sinusoidal patterns due to the inherent graded structure of their gradient descent dynamics, explained by an orthogonal polynomial framework.

Concept interventions can actively worsen model predictions, and existing "benign leakage" criteria fail to prevent this, necessitating new training objectives.

Specialized experts in MoE LLMs, not dense model neurons, are the key to interpretable and controllable personalization.

Prioritizing upstream causes over downstream effects in feature attribution is now computationally tractable, even with complex causal dependencies.

LLM reasoning failures are often misdiagnosed by standard probes, with even specialized probes failing to pinpoint the exact faulty step more reliably than trace length.

A single forward pass now reveals which input variables matter for PFN predictions, achieving exact FDR control and outperforming lasso, BART, and other methods on sparse nonlinear regression.

SAE features are more compositional than random projections, but pairwise interactions are still essential for accurate behavioral steering.

Humanoid robot control policies can be generated directly from their neural weights using diffusion models, unlocking new avenues for policy synthesis and understanding.

A single reward for multiple retrieval actions is a bottleneck; CASI unlocks precise within-set credit assignment, boosting agent performance by 3.51% in production.

Actively repelling uncertain features during representation learning slashes extreme segmentation errors in ambiguous anatomical boundaries by 22x, even without sacrificing overall accuracy.

For star-free languages, transformers learn best when attention weights lag behind output layer updates, defying intuitive feature learning theories.

LLMs can learn to manipulate their own internal activations, but this control is limited and often mimicked by external generation.

Foundation models trained on synthetic order flow perform amortized Bayesian inference, decoding latent market regimes with surprising accuracy and outperforming traditional methods for simulator calibration.

Globally optimized Boolean rule sets are now achievable with explicit control over accuracy-sparsity trade-offs, thanks to a novel MILP formulation that drastically cuts down on integer variables.

Visualize KANs like never before: directly manipulate B-spline control points and see gradient descent rewire activation functions in real-time.

Conditional flow models hallucinate non-additive scientific interactions even when achieving high aggregate fidelity, failing to generalize interaction structure when condition patterns diverge.

SAEs can achieve near-optimal reconstruction with orders of magnitude less data than typically used, but recovering specific, reproducible features demands significantly more.

Language models encode task information in hidden states *before* their unembedding matrix can express it, suggesting current developmental timelines are misleading.

Transformer weight-space learning is now invariant to the full, rich symmetry group of attention heads and token embeddings.

Prompt adaptation methods don't just change what LLMs say, they fundamentally rewire *how* they compute, with distinct mechanisms for text optimization, prompt tuning, and prefix tuning.

Explanations from Natural Language Autoencoders can identify their target activation but fail to discriminate it from plausible rivals, suggesting current training paradigms may not yield robust oversight tools.

A grid-based world model can achieve 89% out-of-view object recall and outperform a Kalman tracker in view, but a node-based approach offers drastically superior memory efficiency and robustness to unmodeled velocities.

Pruning LLMs while preserving their influence functions is now over 28x faster and more effective, unlocking better data selection for downstream tasks.

A simple post-hoc geometric calibration method recovers lost out-of-domain capabilities and boosts in-domain performance after SFT, even on large, diverse models.

Weak interactive feedback can dynamically steer LLMs to the correct similarity criterion for test-time decisions, unlocking significant performance gains beyond static mixtures.

Sparse autoencoders recover known factors 80% better by training in the data's generative space, not its scaled readout space.

LLM judges can be recalibrated to align with human preferences by probing their sensitivity to evaluative biases, even without explicit rubric criteria.

Adam's implicit bias offers a practical defense against training data reconstruction attacks, unlike SGD and Muon which leave reconstructible fingerprints.

Multilingual LMs encode a universal color-language geometry, but not a universal color-perception geometry, with perceptual grounding fading in later layers.