Search papers, labs, and topics across Lattice.
69 papers published across 4 labs.
First-token broadcasting heads in multilingual models can dictate language output, revealing a critical mechanism behind language generation errors.
Safety-aligned LLMs can be manipulated through a steerable "safety axis," exposing critical vulnerabilities while also offering a pathway for robust defenses.
Runtime safety detection in coding agents can be achieved through a novel intervention in hidden representations, drastically reducing harmful actions during multi-turn interactions.
DiffusionGemma's reasoning may seem opaque, but by interpreting its intermediate states, we can dramatically enhance transparency without sacrificing performance.
EFIQA achieves superior image quality assessment by leveraging anatomical knowledge, offering spatial insights without the need for labeled data.
First-token broadcasting heads in multilingual models can dictate language output, revealing a critical mechanism behind language generation errors.
Safety-aligned LLMs can be manipulated through a steerable "safety axis," exposing critical vulnerabilities while also offering a pathway for robust defenses.
Runtime safety detection in coding agents can be achieved through a novel intervention in hidden representations, drastically reducing harmful actions during multi-turn interactions.
DiffusionGemma's reasoning may seem opaque, but by interpreting its intermediate states, we can dramatically enhance transparency without sacrificing performance.
EFIQA achieves superior image quality assessment by leveraging anatomical knowledge, offering spatial insights without the need for labeled data.
Ground-truth latent variables from a new synthetic data model can be decoded from neural network activations, revealing insights into interpretability like never before.
GEMS enables LLMs to maintain high accuracy even when injecting multiple semantic directions, defying the collapse typically observed in such scenarios.
Tri-Info achieves 83% accuracy in real-world failure detection for VLA models without retraining, revealing critical insights into model behavior across diverse tasks.
Attention-guided deep learning can achieve over 90% accuracy in sperm morphology classification while providing critical interpretability for clinical applications.
Identifying a shared activation-space direction across language model families could revolutionize how we detect and mitigate emergent misalignment in AI systems.
Single-neuron interventions can control language model behaviors without collapsing outputs, but only if they respect a defined coherence budget.
CUPID outperforms existing deepfake detection methods by leveraging 3D facial representations for unmatched robustness and interpretability, even without prior exposure to specific individuals.
Users can now access complex global explanations of vision models through simple natural language queries, transforming how we interact with black-box classifiers.
Transforming 2D visual features into interpretable 3D representations unlocks a new level of spatial intelligence in MLLMs, offering unprecedented insights into their internal workings.
Superposition in continual learning doesn't always lead to forgetting; instead, strong representations can mitigate memory loss even in sparse settings.
Essay quality is encoded in LLMs as linearly accessible representations, revealing a surprising robustness across prompts and a systematic reliance on deeper layers for longer essays.
Semantic clustering transforms the Tsetlin Machine into a powerful yet interpretable model, rivaling BERT's performance without sacrificing transparency.
Achieving over 51% accuracy improvement, MM-CBM redefines interpretability in multimodal deep learning by aligning image and text features with natural concepts.
HydraHead achieves a remarkable 69% performance boost in long-context tasks by smartly hybridizing attention mechanisms at the head level, challenging conventional layer-wise designs.
Executable programs can now replace attention heads in transformers with minimal performance loss, achieving over 75% similarity to original patterns.
Achieving faster and more accurate neurosymbolic learning, NeSyCat Torch bridges the gap between classical and neural systems with a unified framework.
OrthoReg prevents neural networks from absorbing symbolic structures, leading to clearer and more interpretable hybrid models in dynamical systems.
Renewable energy sources, particularly solar, are revealed to be critical drivers of electricity prices, challenging conventional perceptions of their market impact.
KPP reveals a unified geometric structure for tree ensembles that could transform how we interpret model predictions and their robustness.
SAERec achieves a breakthrough in intent-based recommendations by automatically constructing a nuanced intent space that significantly enhances both accuracy and interpretability.
ThinkDeception not only detects deception but also provides interpretable reasoning paths, setting a new standard in multimodal analysis.
Code-Augur reveals hidden vulnerabilities in software by transforming agentic assumptions into explicit security specifications, leading to unprecedented detection rates.
Output vector editing can suppress up to 87.9% of memorized sequences in large language models, significantly outperforming traditional neuron-level methods.
PYPILINE's innovative use of a suspicious API knowledge base allows for a dramatic leap in malicious package detection accuracy, setting a new standard in open-source security.
SMART redefines brain atlas construction by achieving state-of-the-art forecasting accuracy while maintaining interpretability and scalability in high-dimensional medical imaging.
BrainFusionNet achieves 98% accuracy in brain tumor detection by uniquely combining CNNs, ViTs, and GRUs, revealing that pixel intensity distributions significantly influence deep learning performance.
GRIDEX transforms deepfake detection by providing clear, structured explanations of audio anomalies, making forensic analysis more efficient and reliable.
Social reasoning in language models is rooted in distinct training data, with targeted unlearning revealing its vulnerability to data removal.
Verified safety properties for multi-agent communication policies can be reliably transferred from decision trees to neural networks, ensuring safer deployment in critical robotic systems.
Independent control over musical attributes like pitch and duration is now achievable without retraining, thanks to a novel activation steering framework.
Text dominance in Audio LLMs can be mitigated through a novel back-patching technique that enhances audio representations, challenging the status quo of multimodal processing.
Sign patterns in transformer hidden states can predict next tokens with up to 93% accuracy, revealing a rich semantic structure without any training.
Self-distillation can be transformed from mere imitation of a privileged distribution to a powerful tool for diagnosing and correcting specific reasoning failures in large language models.
MCS can predict OOD performance with astonishing accuracy, revealing a hidden linearity that traditional methods miss.
Topological regularization in NMF not only improves interpretability but also unifies the modeling of diverse data structures, from images to time-series.
Counterfactual Time Series Explanations (ConTex) can generate actionable insights for time series forecasting with a staggering 12-36x reduction in computational costs compared to traditional methods.
KANLib achieves competitive efficiency and flexibility in KAN architectures, enabling researchers to explore innovative designs with minimal performance trade-offs.
Temporal credit dilution can lead models to misattribute importance, but CREST effectively re-anchors event credit, enhancing robustness in global readouts.
BBMF uncovers interpretable patterns in cancer genomics that traditional methods miss, linking patient subsets to critical chromosomal alterations.
STATEWITNESS not only identifies deception in LLMs but also offers detailed insights into the reasoning behind suspicious responses, transforming how we audit AI behavior.
PhaseWin slashes the computational cost of visual attribution from quadratic to linear while preserving high faithfulness, revolutionizing how we interpret model decisions.
Clamping harmful features in models may only mask misbehavior, allowing for a surprising 95.8% recovery of suppressed actions even under intervention.
SAE-based explanations can be certified for faithfulness, revealing that later layers of language models are significantly easier to interpret without sacrificing predictive accuracy.
Image classifiers rely more on phase than magnitude for identity recognition, challenging conventional views on feature importance in neural networks.
Explanation Cards transform opaque algorithmic explanations into actionable insights, ensuring users understand their limitations and applications.
Uncovering hidden environmental factors can transform anti-poverty strategies, revealing actionable insights that were previously overlooked.
A novel hybrid framework allows for the recovery of unknown ion channel dynamics directly from voltage recordings, bridging gaps in biophysical neuron modeling.
Transformers can adaptively infer complex latent contexts, revealing a surprising link between model depth and in-context learning efficiency.
CircuitLasso reveals how human-interpretable semantic features propagate through LLMs, achieving high accuracy with drastically lower computational costs.
Circuit discovery in LLMs is fraught with variability, and a new method reveals that this complexity may be an inherent challenge rather than a flaw in existing techniques.
Swapping concept detectors in CBMs reveals that high task accuracy can mask significant reliability issues, but a novel training strategy can nearly eliminate this leakage.
REFLEX achieves remarkable sample efficiency, solving complex tasks with fewer than 10 LLM calls while ensuring transparent policy evolution.
LLMs often attend to the right tools but still fail to select them, with readout errors accounting for the majority of miscalls.
Multimodal fusion of text and audio reveals that deep learning embeddings perfectly encode the geometric structure of human emotions, challenging the notion that such structures are merely artifacts of human labeling.
Language models can internally track their trajectory's value, influencing their confidence and decision-making in real-time.
Pronoun fidelity in LLMs is driven by a competition between group binding, recency, and stereotype biases, revealing a complex internal landscape that challenges simplistic behavioral interpretations.
Grounded explanations for speech deepfake detection can boost accuracy by over 45%, transforming how we interpret AI decisions in this critical domain.
Interpolating between opposing directorial personas not only reveals surprising coherence improvements but also uncovers a shared moral-tone substrate in transformer models.
LLMs can effectively evaluate the clinical relevance of AI-generated visual explanations in skin disease diagnosis, revealing critical insights into model trustworthiness.
Class-conditioned perturbations reveal that in-distribution samples remain stable while OOD samples diverge, offering a new lens on model uncertainty.
Hierarchical visual concepts learned through cascaded sparse autoencoders could revolutionize how we interpret and manipulate MLLM outputs.
Moderate geometric convergence in LLMs can coexist with near-perfect functional transfer, challenging assumptions about structural alignment across architectures.
Mapping nanopore signals into a latent space cuts computational costs by three orders of magnitude while enhancing molecular identification accuracy.
Executable reasoning graphs reveal every step of AI's thought process, making it auditable and reproducible in ways traditional methods cannot match.
