Interpretability & Mechanistic Interp

The internal kNN graph of UMAP can reveal critical insights about high-dimensional data that traditional embedding approaches overlook.

Models trained with explanation-guided constraints not only outperform traditional approaches but also produce interpretations that align with expert knowledge, bridging the gap between model performance and interpretability.

Cross-lingual type representations can be extracted from untyped code, revealing hidden structures in state-of-the-art code models that challenge our understanding of their internal workings.

Memorized knowledge in LLMs can exist without being effectively utilized, leading to a surprising 58–75% recovery in generalization performance through targeted internal adjustments.

Achieving a 6.06% boost in semantic alignment while maintaining over 99% reconstruction fidelity, $S^2AE$ redefines how we approach concept consistency in vision-language models.

State-of-the-art audio encoders may obscure critical frequency-localized features, but a simple post-hoc intervention can recover access and improve interpretability.

Compressing prompts into a single activation vector can cut computational costs while maintaining nearly full accuracy in LLM responses.

CIF reveals that many purported causal claims in mechanistic interpretability lack statistical support, challenging the reliability of existing evaluation methods.

GTRC intelligently adjusts the number of k-means++ restarts based on dataset difficulty, optimizing performance and efficiency without sacrificing clustering quality.

DeepPySR achieves superior performance in symbolic regression, yielding interpretable models that significantly outperform traditional methods in real-world scientific applications.

Achieving state-of-the-art parsing performance with 99.94% fewer rule-scoring parameters, Hol-PCFG transforms how we interpret grammar rules in unsupervised parsing.

Internal representations in LLMs can serve as powerful lie detectors, revealing hidden shifts in forecasts that chain-of-thought reasoning fails to capture.

Phoneme-level insights reveal the hidden linguistic cues that differentiate real speech from deepfakes, enhancing trust in detection systems.

Achieving over 70% correlation in cross-seed feature universality reveals that BERT models can be aligned to uncover interpretable sociolinguistic insights.

Moderately expressive neural networks outperform more complex models in recovering mechanistic operators from sparse data, revealing the critical balance needed in architecture and optimization.

Unraveling the black box of neural networks reveals how internal circuits and sparse features can be harnessed for safer, more interpretable AI systems.

Sparse interventions can activate complex task behaviors in language models by targeting just 0.01% of neurons, revealing hidden nonlinearities in model dynamics.

Transformers can learn modular multiplication by partitioning input space into local algebraic regions, revealing surprising new insights into their reasoning capabilities.

IAIML reveals that leveraging feature interactions can dramatically enhance interpretability without sacrificing predictive power, outperforming traditional methods in complex datasets.

Deep ReLU networks can create intricate piecewise linear partitions of input space, fundamentally altering our understanding of their training dynamics.

LAD uniquely combines interpretability and fidelity, delivering human-readable explanations that are grounded in the model's own feature geometry without any retraining.

Uncovering the hidden semantic vocabulary of deepfake detectors transforms our understanding of how these models differentiate between real and fake content.

ReMoDEx reveals that deep learning classifiers often rely on shortcut associations, exposing critical blind spots in conventional evaluation metrics.

Entity familiarity and factual reliability in LLMs are distinct phenomena, with models showing high awareness of known entities yet rarely abstaining from incorrect answers.

Riemannian Mean Pooling reveals that leveraging geometric properties of embeddings can significantly enhance classification performance while avoiding pitfalls of annotation-driven biases.

Different LLMs encode sycophancy in strikingly diverse ways, revealing a complex interplay between factual agreement and subjective belief.

Achieving 99.60% accuracy with a model that requires only 2,370 FLOPS could redefine the landscape of IoT security solutions.

Extracting interpretable policies from deep RL agents can boost performance while simplifying complex decision-making processes.

Instruction leakage can lead to misleadingly high accuracy in spatial relation tasks, revealing a critical flaw in goal-conditioned models that could misguide future research.

By grounding adaptive retrieval in interpretable uncertainty signals, this framework transforms how LLMs handle knowledge gaps and ambiguities in real-time.

Adversarial attacks can fundamentally alter LLM internal reasoning, revealing hidden vulnerabilities that can be directly addressed through causal interventions.

Modular task decomposition in AI-generated analyses boosts transparency and reliability, enabling smaller models to outperform larger counterparts.

Distinct training objectives in self-supervised speech models create unique acoustic compression regimes that impact downstream task performance.

Latent reasoning methods may seem unfaithful at convergence, but their causal contributions to answers decay over training, revealing hidden complexities in model behavior.

Compact evidence masks can achieve up to 96% repair rates in MLLM tasks, proving that less can be more in visual attribution.

Token-level explanations reveal how FEMRs leverage patient history, bridging the gap between black-box models and clinical trust.

ExplAIner can express a diverse array of explanation types while ensuring efficient evaluation, transforming how we approach interpretability in machine learning models.

Early detection of failure in LLM agents can save over 47% of inference compute by leveraging internal representations rather than observable behavior.

Uncovering the causal influence of driving behavior concepts can significantly enhance the safety and performance of autonomous vehicles.

Annotation noise in vascular CT scans can be detected with a novel method that reveals systematic biases, improving training robustness dramatically.

Injection-driven, weakly supervised training can achieve reliable Real-Bogus classification without human labels, providing calibrated uncertainties and robust performance under class contamination.

Pathway Activity Autoencoders reveal that integrating multi-omics data can significantly enhance cancer risk stratification and survival predictions while maintaining interpretability.

Second-order scattering coefficients reveal that temporal amplitude modulation is the key electrophysiological signature of schizophrenia, outperforming traditional EEG analysis methods.

SSL representations of satellite imagery can reveal hidden environmental associations that significantly impact downstream task performance.

Observability in representation learning is fundamentally constrained by the geometry of latent states, revealing critical gaps in current interpretability methods.

Verbalization, not knowledge, drives fabrication in LLMs, revealing a critical leverage point for improving mathematical reasoning accuracy.

Training a neural network with $L^4$ loss enables it to compute more functions than neurons, revealing a surprising efficiency in representation.

Fine-grained taxonomic representations can boost recall in antisemitism detection, but at the cost of precision, revealing a critical trade-off in LLM performance.

A novel Circadian Rhythm Score reveals that nearly all predictive power can be captured from multi-domain behaviors, transforming depression screening and intervention strategies.

Pre-generation signals can predict failure sources in vision-language models, enabling targeted interventions before answers are generated.