Search papers, labs, and topics across Lattice.
96 papers published across 4 labs.
Fuzzy rules reveal how CLIP encodes domain-specific features in clinical reports and film reviews, offering a peek inside the black box of multimodal embeddings.
Forget static embeddings: this paper shows how modeling scientific concepts as evolving complex networks reveals surprising connections between conceptual change and network topology.
Locomotion policies, often considered black boxes, can autonomously learn interpretable phase structures and branching logic, revealing a hidden order in their decision-making.
Video diffusion transformers exhibit a hidden "magnitude hierarchy" in their activations that can be exploited for training-free quality improvements via a simple steering method.
LLMs don't just regurgitate token probabilities when expressing confidence; they perform a more sophisticated, cached self-evaluation of answer quality.
Forget static embeddings: this paper shows how modeling scientific concepts as evolving complex networks reveals surprising connections between conceptual change and network topology.
Locomotion policies, often considered black boxes, can autonomously learn interpretable phase structures and branching logic, revealing a hidden order in their decision-making.
Video diffusion transformers exhibit a hidden "magnitude hierarchy" in their activations that can be exploited for training-free quality improvements via a simple steering method.
LLMs don't just regurgitate token probabilities when expressing confidence; they perform a more sophisticated, cached self-evaluation of answer quality.
LLMs encode hierarchical semantic relations asymmetrically, with hypernymy being far more robust and redundantly represented than hyponymy.
Attention sinks aren't just a forward-pass phenomenon; they actively warp the training landscape by creating "gradient sinks" that drive massive activations.
People prefer XAI explanations that tell them *why* a feature change doesn't alter the outcome, not just *that* it doesn't.
MLLMs' image segmentation prowess isn't a given: a critical adapter layer actually *hurts* performance, with the LLM having to recover via attention-mediated refinement.
Anomaly detection gets a dose of interpretability: SYRAN learns human-readable equations that flag anomalies by violating learned invariants.
Pinpointing the training data behind an LLM's behavior is now possible without retraining, opening the door to precise debugging and targeted interventions.
Acoustic and phonetic NACs encode accent in fundamentally different ways, with implications for how we interpret and manipulate these representations.
Control the emotional tone of generated speech without any training by directly manipulating specific neurons within large audio-language models.
Image editing models leak fascinating hints about their world knowledge through "edit spillover"—unintended changes to semantically related regions—and this paper turns that leakage into a probe.
CLIP struggles with fine-grained details in cross-domain few-shot learning, but a cycle-consistency method can fix its vision-language alignment and boost performance.
You can now audit multi-agent LLM systems and trace responsibility for harmful outputs even without access to internal execution logs, thanks to a clever "self-describing text" technique.
An AI model can estimate legal age from clavicle CT scans with higher accuracy than human experts, potentially revolutionizing forensic age assessment.
Unlock explainable outlier detection in foundation models with FoMo-X, a modular framework that adds negligible inference overhead while revealing interpretable risk tiers and calibrated confidence measures.
Standard PCA, despite its widespread use in CAD, struggles to directly reveal the original design parameters of a geometry, but this paper identifies conditions for accurate parameter estimation.
LLMs aren't monolithic black boxes: they contain spatially organized, functionally specialized modules that can be automatically discovered.
Forget one-hot encodings: conditioning timbre VAEs on continuous perceptual features unlocks more compact and controllable latent spaces.
Achieve expert-level accuracy in wasp identification with a YOLO-based model that also shows *why* it makes its classifications, thanks to integrated HiResCAM explainability.
Transformers have a hidden symmetry: depth-wise residuals are secretly doing the same thing as sequence-wise sliding window attention, unlocking new architectural insights.
LLMs often fail to update their final predictions after interventions on intermediate reasoning steps, suggesting that these structures function more as influential context than stable causal mediators.
Fuzzy logic and deep learning join forces to make radio astronomy ML pipelines less black-box.
Object hallucinations in LVLMs aren't just a language problem—abnormal visual attention patterns are also to blame, and can be fixed without retraining.
By strategically exploiting LLMs' inconsistent cross-lingual performance, this work offers a surprisingly scalable way to pinpoint the specific experts responsible for storing and retrieving factual knowledge.
Transformers trained on a simple grid-world learn hidden representations that directly reflect the underlying predictive geometry, offering a glimpse into how neural networks internalize structural constraints.
LRMs can often recover from injected errors in their reasoning steps, revealing a hidden "critique" ability that can be harnessed to improve performance without additional training.
An interpretable machine learning framework leveraging XGBoost and DeepSeek reveals key genetic factors driving drug response in lung cancer, offering a path towards personalized treatment strategies.
Escape the flatland of traditional recommender systems: RecBundle uses differential geometry to disentangle user interactions from preferences, opening the door to understanding and mitigating systemic biases.
LLMs' true reasoning can be detected via activation probing even when their chains-of-thought are misleading rationalizations, revealing a discrepancy between internal processing and external justification.
A tabular LLM, TAP-GPT, rivals state-of-the-art general-purpose LLMs in few-shot Alzheimer's prediction while offering interpretable reasoning and robustness to missing data, opening the door to more transparent and reliable clinical AI.
Distributional counterfactual explanations are now possible for black-box tabular models, thanks to a novel sparse search algorithm that sidesteps the need for gradients.
Adversarial representation learning can improve the out-of-distribution generalization of age predictors, but don't mistake correlation for causation.
Injecting data-derived spectral priors into neural network initialization can dramatically accelerate convergence and improve the efficiency of function parameterizing architectures.
Standard upsampling methods in XAI systematically corrupt attribution signals, but a novel semantic-aware redistribution approach provably preserves attribution mass and improves explanation faithfulness.
LLMs can now write better quantitative trading algorithms than humans, thanks to a new framework that turns unstructured financial reports into executable code.
Even when multimodal LLMs get face verification right, their explanations are often wrong, relying on hallucinated facial attributes.
Forget blindly pruning LLMs: this work shows you can use Sparse Autoencoders to identify and protect the most functionally important components during compression, leading to more robust models.
Forget fine-tuning: steer a 35B MoE's agency on the fly with SAE-decoded vectors, revealing a surprisingly simple, one-dimensional control knob.
LLMs' "Aha!" moments aren't about magic tokens, but about explicitly verbalizing and managing uncertainty during reasoning, which drives performance.
Unlock the secrets hidden within LoRA weights: a novel method reveals that these weights already encode adapter behavior and performance, enabling accurate predictions without running the base model or accessing training data.
LLMs dissect tables in three distinct attention phases: broad scanning, cell localization, and contribution amplification.
Forget holdout data for feature effect estimation: training data's larger sample size usually wins, and cross-validation can further reduce model variance.
Forget hand-crafted features: this system uses an LLM to automatically discover features from event sequences that outperform even state-of-the-art embeddings by up to 5.8%.
Uncover hidden biases and track evolving viewpoints: POLAR reveals individual-level associations in text data that are masked by traditional aggregate analyses.
LLMs exhibit a surprising degree of moral indifference, compressing distinct moral concepts into uniform probability distributions, a problem that persists across model scales, architectures, and alignment techniques.
Questioning the common practice of interpreting data through a single model class, this work reveals the existence of alternative well-performing models across multiple model classes and their hyperparameters.
Forget persistent homology's computational cost: Euler Characteristic Surfaces unlock 98% accuracy in ECG classification with linear complexity, rivaling deep learning while staying interpretable.
Forget iterative optimization – this method synthesizes adversarial patches for facial re-ID in a single forward pass, dropping mAP from 90% to near zero.
Infant motor learning reveals a sharp phase transition in control strategy arbitration, governed by context window size and predictable via a closed-form exponential moving average.
Stylometric features, combined with modern multilingual language models, significantly boost the performance of machine-generated text detection, often surpassing language-specific models.
Debugging multi-agent systems just got easier: AgentTrace pinpoints root causes of failures with high accuracy and speed, without needing costly LLM inference during debugging.
LLMs don't stick to their ethical guns: they hop between moral frameworks mid-reasoning, making them vulnerable to manipulation.
Algorithmic metrics for counterfactual explanations? Turns out humans don't really agree with them.
Unlock robust feature importance analysis with `xplainfi`, an R package that fills critical gaps by offering conditional importance methods and statistical inference for diverse ML models.
KANs become far more robust and interpretable with in-context symbolic regression, achieving near-perfect error reduction in hyperparameter sweeps.
By blending counterfactual and feature attribution methods, GradCFA generates more realistic and diverse explanations, offering a richer understanding of neural network decisions than either approach alone.
XGBoost models can be debiased for gender fairness in critical healthcare settings with minimal performance loss using a novel multi-metric Bayesian optimization approach.
LLMs' true power lies in the "unexplainable" – capabilities that exceed rule-based systems, challenging the pursuit of full interpretability.
LLMs can withstand 3,000 sequential knowledge edits without catastrophic forgetting, thanks to a new sparse editing framework that surgically manipulates knowledge circuits.
VLMs' hallucinations aren't just errors, but traceable pathologies in their "cognitive trajectory," diagnosable via geometric anomalies in a learned state space.
Forget prompt engineering – surgically altering a model's internal activations can jailbreak it, exposing vulnerabilities even when the input looks harmless.
LLMs can now offer globally contestable decision support by systematically mapping decision spaces into argumentation frameworks, allowing users to challenge the underlying logic, not just individual outputs.
Uncover the hidden dynamics of your RL agent with a new visualization framework that reveals how TD errors sculpt the optimization landscape and drive policy updates.
LLMs can automatically translate complex access control rules into plain English, making security policies understandable to non-experts.
Visualizing the critic's loss landscape reveals distinct characteristics linked to stable vs. unstable learning in online RL, offering a new window into algorithm dynamics.
Surprisingly, video diffusion models contain recoverable physics-related cues in their intermediate denoising representations, enabling more physically plausible video generation with reduced computational cost.
ResNet50 is shown to leak semantic attributes into its null space, while DinoViT better preserves class semantics, revealing critical differences in how these architectures handle semantic invariants.
Even with weaker assumptions, ICA post-processing can unlock state-of-the-art disentanglement from vanilla autoencoders and foundation model-scale MAEs.
Single-cell foundation models exhibit surprising annotation bias, with 40% of highly connected features lacking biological annotation, suggesting current interpretability methods may be systematically skewed.
Unlocking interpretable AI just got easier: HyperExpress disentangles image concepts into composable parts using hyperbolic space, letting you reconstruct visuals from their semantic building blocks.
Despite the intuition that noisy environments should make models rely more on visual cues, AVSR models stubbornly cling to audio, even when it's heavily degraded.
XAI can boost trust in fake news detection by revealing which words sway the model, but choosing the right XAI method (SHAP, LIME, or Integrated Gradients) matters for performance and interpretability.
Forget black-box anomaly detection: this neuro-symbolic VLM agent uses natural language descriptions and visual grounding to explain *why* an event occurred in multivariate time series data, even with little training.
Neuromodulation offers a way to disentangle global contextual parameters from local manifold representations in constrained autoencoders, enabling context-aware dimensionality reduction.
Fine-tuning unlocks LLMs' surprising ability to predict how memorable a sentence is and how long it takes to read, exceeding traditional methods.
Fuzzy rules reveal how CLIP encodes domain-specific features in clinical reports and film reviews, offering a peek inside the black box of multimodal embeddings.
Ditch the concatenation: a new neural dependence estimator sidesteps MINE's computational baggage, offering a more stable and efficient way to analyze autoencoder features.
Uncover the surprising locations of demographic biases within CLIP's vision encoder by pinpointing specific attention heads responsible for encoding gender and age stereotypes.
Unlock precise, training-free color control in text-to-image models by directly manipulating the latent space's emergent Hue, Saturation, and Lightness structure.
Stop wrestling with opaque expression models: ELISA lets you directly translate single-cell RNA sequencing data into mechanistic biological hypotheses using an interpretable hybrid generative AI agent.
Concept erasure in text-to-image models no longer needs to be a blunt instrument: OrthoEraser precisely removes harmful content while preserving image quality by analytically orthogonalizing the erasure process.
Softmax attention's normalization creates unavoidable "attention sinks" when implementing trigger-conditional logic, but ReLU attention offers a sink-free alternative.
Robots can boost their perceived competence by 83% simply by tweaking navigation behaviors suggested by a causal Bayesian network.
Achieving fairness doesn't just mean equal outcomes—this work shows how to enforce consistent reasoning across groups by penalizing disparities in counterfactual explanations.
Uncover hidden backdoors in your neural networks by tracing the active paths that malicious triggers exploit.
Forget subjective human evaluations: this paper uses a clever knowledge distillation trick to objectively rank XAI methods for NMT, revealing that attention-based attributions beat gradient-based ones.
Uncover the hidden causal chains inside your LLM with Causal Concept Graphs, which outperform existing methods for reasoning by explicitly modeling concept dependencies.
Speech deepfake detection gets a reasoning upgrade: HIR-SDD uses chain-of-thought prompting with Large Audio Language Models to not only detect fakes but also explain *why* it thinks they're fake.
Clinicians using HeartAgent, a cardiology-specific agent system, improved diagnostic accuracy by 26.9% and explanatory quality by 22.7% compared to unaided experts.
Forget fine-tuning: surprisingly, single neuron activations in VLMs can be directly probed to create classifiers that outperform the full model, with 5x speedups.
Chinese metaphor identification is highly sensitive to the choice of protocol, dwarfing the impact of model-level variations, yet can be tackled with fully transparent, LLM-assisted rule scripts.
Prompt highlighting in LLMs gets a serious upgrade: PRISM-$\Delta$ steers models to focus on relevant text spans with better accuracy and fluency, even in long contexts.
Fair-Gate disentangles speaker identity and sex in voice biometrics, boosting fairness without sacrificing accuracy by explicitly routing features through identity and sex-specific pathways.
LLMs possess a "word recovery" mechanism that allows them to reconstruct canonical word-level tokens from character-level inputs, explaining their surprising robustness to non-canonical tokenization.
