Interpretability & Mechanistic Interp - Weekly Roundup

Instruction tuning degrades the latent moral geometry of LLMs, even though distribution-driven steering vectors can naturally mirror human value topologies to enable coherent cross-value generalization.

Chain-of-thought is far more than surface imitation: LLMs carve out discrete, context-dependent geometric subspaces for functional operations like deduction and decomposition, with middle layers encoding the abstract reasoning step rather than the literal tokens.

DIFFINT reveals the specific feature ranges that contribute to anomalies, making it the only interpretable detector among the top-performing methods in its class.

Autoencoder parameters can act as powerful, compact representations of data, revealing deep connections between model architecture and statistical properties.

Instruction tuning degrades the latent moral geometry of LLMs, even though distribution-driven steering vectors can naturally mirror human value topologies to enable coherent cross-value generalization.

Chain-of-thought is far more than surface imitation: LLMs carve out discrete, context-dependent geometric subspaces for functional operations like deduction and decomposition, with middle layers encoding the abstract reasoning step rather than the literal tokens.

DIFFINT reveals the specific feature ranges that contribute to anomalies, making it the only interpretable detector among the top-performing methods in its class.

Autoencoder parameters can act as powerful, compact representations of data, revealing deep connections between model architecture and statistical properties.

Human corrections can directly reshape AI inference in real-time, eliminating the need for separate instructions or records.

LLMs not only understand plural references but also mirror human preferences in pronoun usage based on entity similarity and conjunctions.

LLM judges may recognize important reasoning steps, but they still miss critical nuances, highlighting a gap between legibility and true interpretability in AI reasoning.

Explanation fidelity in Federated Learning can be enhanced by dynamically calibrating Differential Privacy noise, leading to significant gains in both accuracy and interpretability.

WAND achieves anomaly detection with built-in, accurate explanations at no additional computational cost, revolutionizing interpretability in unsupervised learning.

Causal explanations just got scalable—Probabilistic Causal Impact bridges the gap between rigorous causality and practical application, enabling nuanced insights from large datasets.

Beyond Intuition outperformed other methods in explaining model behavior, yet its correlation with heart rate accuracy was surprisingly weak, challenging conventional assumptions about explainability in AI.

LLMs can accurately predict typological features while providing interpretable rationales, even for low-resource languages, when given the right contextual evidence.

Joint audio-video diffusion models hallucinate not from diffuse attention noise, but because bidirectional audio-video cross-attention systematically overrides text conditioning in favor of learned canonical priors.

Security mechanisms based on LLM linguistic outputs are fundamentally unreliable due to inherent "linguistic illegibility," necessitating new isolation techniques for model safety.

Logical validity can be accurately decoded from hidden states of language models, even when their outputs are incorrect, revealing a critical gap between representation and expression.

TDA can dramatically streamline the computation of maximum likelihood thresholds in colored Gaussian graphical models, revealing insights previously obscured by traditional methods.

Disentangled representations can reduce unlearning collateral damage by up to 4x, challenging the assumption that entanglement is merely a theoretical concern.

Stacked, a novel supervised method, significantly enhances hallucination detection in LLMs, outperforming alternatives in nearly half of the tested scenarios.

Contrastive explanations can reveal nuanced differences between competing arguments, significantly enhancing interpretability in AI-driven decision-making.

Saliency maps can mislead when images are rotated, but EquiGrad-CAM reveals how to achieve consistent explanations without retraining models.

Semantic structures can serve as a reliable fingerprint for LLM ownership, achieving perfect detection without sacrificing model performance.

Counterfactual medical images can now be generated directly from classifiers, providing clearer insights into decision-making without the biases of generative models.

LLM-brain alignment might suggest shared computational principles, but it fails to confirm a common underlying mechanism, revealing deeper issues of underdetermination in AI models.

Trust in AI can be systematically cultivated through a structured educational framework that demystifies machine learning.

Tracing the origins of synthetic media is now possible without altering generator architectures, thanks to a novel self-referential framework that ensures round-trip consistency.

NeuSOGA transforms complex geometric observations into interpretable symbolic representations, bridging the gap between perception and reasoning in AI.

Steering language models can shift their reliance on context versus memory, but this authority is surprisingly task-specific.

Achieving competitive inventory decisions with significantly fewer features could revolutionize how businesses approach demand forecasting and inventory management.

Quantum machine learning enhances Q-matrix estimation by providing greater stability and robustness, even when classical methods achieve higher average accuracy.

TREVIS is proposed, an approach for learning decision trees with respect to complex objectives, based on the exploration of the latent space of a Tree Transformer Variational Auto-Encoder (TTVAE), that replaces the discrete search space with a continuous one, enabling gradient-based optimization via a differentiable surrogate model.

The truth direction in linear probes is robust to writing style and medical specialty but falters dramatically with specific corpus shifts, revealing hidden vulnerabilities in medical QA systems.

Sparse autoencoders not only reduce bias in causal effect estimation but also enhance interpretability, paving the way for more robust analyses in text-based research.

Storing wider latent representations together can slash reconstruction errors by over half while boosting classification accuracy, challenging conventional wisdom on latent size constraints.

More heads in attention layers not only enhance parameter identification but also reveal that some biases can be entirely irrelevant to model performance.

NSR consistently outperforms traditional symbolic regression methods, achieving superior noise robustness and interpretability in mathematical modeling.

A pre-carved modular structure in LLMs emerges before training, with task dominance established through sharp gradient disparities that defy conventional learning-rate expectations.

CNNs struggle with unseen concepts, while LLMs show surprising clustering of related topics, challenging our understanding of internal representations in neural architectures.

HiLRP achieves 0.97 Pointing accuracy in class activation mapping, far surpassing the 0.55 of existing methods, revealing the critical need for unified attribution in diverse ViT architectures.

Multi-agent LLM systems can misfire due to interconnected errors, and EDGE reveals how understanding these dependencies dramatically enhances error attribution accuracy.

LEAP reveals that separating evidence evaluation can drastically enhance the accuracy and interpretability of probabilistic forecasts in LLM applications.

Emotional information in LLMs is more interconnected than previously thought, with optimal probing layers varying significantly across different text sources.

Swapping internal activations can dramatically improve agreement in model predictions across different question framings.

QILP-0 reconstructs quantum circuit behaviors with perfect accuracy, setting a new standard for observational declarative modeling in quantum computing.

Internal representations in large language models can be read with high accuracy long before they become causally effective, revealing a critical disconnect in model interpretability.

Behaviorally effective LoRA writes are not only sparse but also structured, revealing a surprising concentration of useful signals that could redefine model adaptation strategies.

Translation in multilingual LLMs is more modular than previously thought, with syntax and surface language production operating as distinct processes.

ICLFS achieves superior clustering accuracy by treating features as instances, challenging traditional UFS methods that rely on indirect structural criteria.

Fine-tuning LLMs for PR-issue alignment can boost classification accuracy by over 14%, revealing critical insights into how software artifacts influence model predictions.

Camera motion is geometrically organized in video foundation models, enabling smoother video interpolation that outperforms traditional linear methods.

Visual attention in VLMs is not uniformly reliable; it varies across three distinct modes, revealing a surprising gap between model predictions and human intuition.

Intra-atomic supervision is crucial for achieving accurate energy decomposition in machine-learning interatomic potentials, revealing hidden allocation failures.

A single depression vector in a large language model can accurately distinguish between depressive and non-depressive speech, paving the way for more interpretable mental health assessments.

GAPS slashes toxicity rates in language models by intelligently steering only the neurons that matter, achieving a remarkable reduction from 6.52% to 0.48%.

A novel version-space method reveals that circuit analysis problems can be solved in polynomial time, challenging the exponential assumptions of prior approaches.

Traditional lens methods can mislead due to corpus dependence, but Sparse Readout Prism reveals stable readout features that enhance interpretability and accuracy in language model analysis.

Roleplay scenarios can trick language models into complying with harmful requests, revealing a critical vulnerability in their refusal mechanisms.

VisLens achieves a remarkable speedup in visual search for MLLMs, completing the task in a single pass while maintaining high accuracy.

Competitive dynamics among neurons can autonomously induce specialized modules that meaningfully capture semantic categories in neural networks.

Transforming representational similarity into statistical distinguishability reveals how perturbations propagate through transformer architectures, offering fresh insights into model behavior.

MolLedger achieves exact interpretability in ADME predictions without compromising performance, revealing more chemically relevant insights than traditional methods.

Trustworthiness of LLM-generated explanations varies dramatically, with a reasoning model outperforming others by 15% in fidelity tests.

Tensor methods could unlock significant efficiency gains in LLMs, bridging the gap between theoretical compression and real-world performance.

Many LLMs may possess intact reasoning capabilities hidden beneath output bottlenecks, with simple corrections yielding substantial accuracy gains.

LLMs struggle with self-awareness, making systematic errors on simple counterfactuals, but reinforcement learning can boost their self-modeling skills.

The first token of a multi-token concept can unlock a staggering 88.3% accuracy in predicting subsequent tokens, revolutionizing how we interpret language models.

A lightweight linear probe can unlock effective preference adaptation in LLMs with minimal labeled data, outperforming traditional methods.

Achieving a weighted F1 score of 0.9615 while also doubling pulmonary attribution containment metrics highlights a critical shift in how we assess CXR models.

LLMs may not lead in ranking accuracy for repurchase recommendations, but they excel at providing interpretable rationales that enhance user understanding.

Fragmented defenses against LLM backdoor attacks stem from fundamental differences in how clean and dirty-label attacks manipulate model features.

XVAE-WMT achieves superior sound separation without requiring paired clean recordings, setting a new standard for interpretability in biomedical signal processing.

Models already know when a question is fundamentally unanswerable—they hallucinate anyway because their internal "impossibility" representation is completely orthogonal to their safety-refusal circuitry.