Search papers, labs, and topics across Lattice.
100 papers published across 6 labs.
Class-discriminative signals can be preserved in GNNs even after extensive propagation, challenging the oversmoothing narrative.
Traditional one-point feedback in bandit problems misses the mark, but new algorithms can exploit action similarities to achieve significantly lower regret.
Rejecting high-risk predictions can significantly boost forecasting accuracy, especially for challenging time series.
Lower bounds for minimal risk can be confidently established without dependence on model complexity, revolutionizing how we assess learning machine performance.
Trajectories in mirror flows can converge to a limiting behavior that transforms incremental learning into a dynamic optimization problem.
Class-discriminative signals can be preserved in GNNs even after extensive propagation, challenging the oversmoothing narrative.
Traditional one-point feedback in bandit problems misses the mark, but new algorithms can exploit action similarities to achieve significantly lower regret.
Rejecting high-risk predictions can significantly boost forecasting accuracy, especially for challenging time series.
Lower bounds for minimal risk can be confidently established without dependence on model complexity, revolutionizing how we assess learning machine performance.
Trajectories in mirror flows can converge to a limiting behavior that transforms incremental learning into a dynamic optimization problem.
Surrogate models using self-attention can dramatically reduce the computational cost of simulating complex free-surface flows while maintaining high accuracy.
Neural networks can now learn the hidden parameters of mean field games from population dynamics, unlocking new avenues for modeling complex strategic interactions.
AI agents can yield correct results while relying on fundamentally flawed reasoning, raising critical questions about their reliability as scientific co-authors.
Rare failures in decision-making can be quantified and bounded, revealing new insights into the trade-offs between privacy and performance in interactive settings.
Persistent homology reveals a structured approach to understanding and steering LLM responses to ambiguous questions, boosting accuracy and acceptable response rates significantly.
Relying on proxy utility functions can lead to harmful decision-making outcomes that undermine the very goals they aim to achieve.
LLMs may not reason as often as humans, but when they do, their internal representations mirror the abstract geometric structures of the human brain.
Agency emerges not as a static property but as a dynamic interplay of time-dependent processes that can reshape our understanding of life itself.
The study uncovers unexpected limitations in the expressivity of argumentation frameworks, suggesting a critical threshold for uncertain preferences that could redefine our understanding of argumentation theory.
Adding written definitions to political axes can improve rater agreement by nearly 36% and enhance the reliability of political position measurements in underrepresented regions.
A novel Stackelberg game approach cuts LLM token costs by 17.4% without sacrificing quality, challenging traditional static resource management methods.
Human evaluative claims can significantly enhance the quality of AI-generated peer reviews, striking a crucial balance between automation and accountability.
VCT guarantees the integrity and verifiability of LLM conversation records, even in complex, non-linear interactions.
Prompt optimization can significantly enhance multi-agent LLM systems, but its effectiveness hinges on the specific configuration of the agents and their interactions.
Adaptive estimation of slowly-varying sequences can cut costs by leveraging local variations, achieving a new bound that outperforms traditional methods.
Language models face irreducible error floors that prevent them from being universal solvers, revealing the critical limits of prompt-based learning.
Training RL models on beneficial behaviors can lead to over 80% improvement in alignment across diverse, out-of-distribution tasks.
The interplay between digital humanism and evolutionary design reveals that optimizing for sustainability may inadvertently hinder open technology development.
Behavioral mappings in LLMs can lose substantial predictive power when transferred across environments, even when those environments share identical incentive structures.
Verification can geometrically amplify reliability in problem-solving systems, achieving up to 99.9% accuracy with strategic orchestration of models.
Agents should enhance causal discovery workflows without compromising the integrity of causal claims by relying solely on data and expert knowledge.
A roadmap for bridging the gap between academic techniques and practical applications in autonomous system engineering reveals both immediate solutions and pressing research needs.
Formal verification of robot safety can now be achieved without compromising the expressive power of foundation models, thanks to a novel modular architecture.
Trustworthy AI can now be certified with cryptographic proofs that don't require trusting the agent or re-executing its computations.
Evolving prompts and verifying answers can boost visual reasoning model accuracy by over 19%—a game changer for scaling reliable data in AI.
Mastering agentic AI requires a holistic understanding of the entire system, from foundational models to inter-agent communication protocols.
Clients can now independently verify the integrity of LLM interactions, reducing the risk of manipulation by third-party gateways.
PORTICO eliminates unauthorized actions in coding agents, achieving perfect compliance while a traditional system fails to do so.
Reviewers approve AI-generated code more often while actually engaging less, revealing a troubling trend of habituation that could compromise code quality.
WebCQ achieves 33.3% more state exploration and 42.2% more unique actions than previous MARL methods, revolutionizing web GUI testing scalability.
Bypassing final-layer perturbations can significantly enhance reasoning capabilities in aligned LLMs, achieving better performance with zero memory overhead.
LLM critiques can be systematically evaluated for alignment with human judgment, revealing that better models significantly enhance evaluation reliability.
Autotelic AI reveals that the real challenge lies in how agents construct and understand their own identities, not just in generating goals.
DiffusionGemma's reasoning may seem opaque, but by interpreting its intermediate states, we can dramatically enhance transparency without sacrificing performance.
Deterministic predictors can now achieve optimal multicalibration sample complexity, challenging the long-held belief that randomization was necessary.
Achieving oracle-level Bayesian predictions with a multi-task framework that adapts seamlessly to new priors, all while being orders of magnitude faster.
MAA not only outperforms traditional batch-level distillation methods but also slashes optimization costs by 75%, redefining efficiency in memory-driven agent evolution.
The traditional complexity of leverage-score algorithms is misleading; the real challenge lies in identification, not accuracy, allowing for a dramatic reduction in query complexity.
Optimal coarse correlated equilibria can be achieved in mean field games, enabling performance optimization beyond individual player goals.
U-Net achieves a remarkable 5.38x speed-up over traditional solvers while maintaining 3% prediction accuracy for battery internal states, highlighting the power of spatial inductive bias.
LLMs can effectively coordinate multi-agent strategies, achieving human-like adaptability without manual rule crafting.
Bounded reward noise can lead to a dramatic reduction in regret bounds, achieving \(O(\log T)\) versus the standard \(\tilde{O}(\sqrt{T})\) for sub-Gaussian conditions.
LLMs can now resolve knowledge conflicts in real-time, leading to more reliable and interpretable outputs.
Intelligence can be quantified through a thermodynamic lens, revealing that recursive self-simulation is crucial for amplifying rare but valid futures.
Scale, not task complexity, is the real challenge for multi-agent orchestration in enterprise AI, with significant implications for system design and performance.
Identifying a shared activation-space direction across language model families could revolutionize how we detect and mitigate emergent misalignment in AI systems.
Token-oriented inference optimizations can cut production costs and boost efficiency, transforming large model services from merely callable to fully operable.
Auditable financial chart QA is now achievable on-premise without sacrificing accuracy, revealing critical insights into model failures and trustworthiness.
Honest heterogeneous peers can drastically reduce harmful revisions in LLM debates, but adversarial peers can completely undermine these benefits.
The interaction between AI utilization and cognitive convergence capacity explains 86% of productivity variance, revealing a critical oversight in traditional economic models.
PRDiT achieves unprecedented detail in 3D CT volume generation while simplifying the optimization process, outperforming leading models in the field.
Achieving sound probabilistic verification for AI agents could redefine how we secure complex systems against policy violations in uncertain environments.
Rigid geometric compression can collapse reasoning space, while generative reconstruction preserves semantic integrity and enhances reasoning accuracy.
Expert calibration alone isn't enough for soft-routed MoE models; an innovative adversarial reweighting approach is needed to tackle distribution shifts effectively.
Implicit feedback from user interactions can boost LLM alignment accuracy by nearly 20%, revealing a hidden layer of user preference that explicit feedback misses.
Editorial alignment can empower knowledge institutions to reclaim authority in LLM-driven information services, ensuring that AI reflects their values and standards.
Frontier language models are now central to cyber operations, yet Africa is entirely excluded from their development and access, raising urgent security concerns.
The framework identifies critical anti-patterns in AI integration that could save engineers from costly pitfalls while paving the way for the next generation of generative models.
Risk-controlled model updates can now be certified, ensuring that only safe improvements are accepted, which could transform how we approach model training.
Merging models for multi-task learning can be done without training, preserving task knowledge while minimizing interference through a focus on essential subspaces.
Transitioning from probabilistic to deterministic predictors incurs a quantifiable cost that can be precisely controlled, revealing new insights into generalization bounds.
Type I error inflation in conditional independence testing can be drastically reduced without sacrificing power, thanks to a novel betting-based approach.
Eliciting only the essential statistics from LLMs can dramatically enhance feature acquisition in complex clinical scenarios, outperforming traditional methods even in the toughest cases.
Learning without explicit rewards can yield high-action accuracy and nuanced value inference, challenging traditional reward-based paradigms in reinforcement learning.
Excluding features based on manipulability can lead to suboptimal predictions, revealing a critical flaw in standard feature selection practices.
Scaling AEB systems with massive unlabeled data can lead to a 35% increase in accident-free driving mileage while maintaining a positive-to-false activation ratio over 100:1.
A noisy predictor that barely beats random guessing can dramatically enhance the efficiency of exact exponential algorithms for NP-hard problems.
Evolving Meta-Skills enables automatic Multi-Agent Systems to achieve superior performance without sacrificing experience retention or scalability.
Achieving sublinear cumulative Wasserstein regret in online distributional prediction without any parametric model for drift or corruption is a game-changer for adaptive learning systems.
Explicitly modeling the asymmetry between information growth and action expiration can drastically improve decision-making efficiency in complex environments.
Closing the supervision gap in GUI agents boosts success rates from the low-30% range to over 50% through innovative skill-guided learning.
CURE redefines context management for tabular foundation models, achieving significant performance gains in stream learning by intelligently managing uncertainty and redundancy.
Generalist agents need to remember distinct information to navigate conflicting optimal actions across environments, challenging the notion that current state observations are sufficient.
Code-Augur reveals hidden vulnerabilities in software by transforming agentic assumptions into explicit security specifications, leading to unprecedented detection rates.
SAGE achieves a statistically robust increase in user retention for a mental-health chatbot by effectively navigating the complexities of prompt optimization through agent-guided exploration.
Misfired alignment in LLMs can lead to a 18.9% failure rate in reasoning about stereotypes, revealing a critical flaw in current safety-oriented training methods.
Sub-agents can now communicate failure states and rationales, boosting response reliability by over 10% in complex multi-agent systems.
SMART redefines brain atlas construction by achieving state-of-the-art forecasting accuracy while maintaining interpretability and scalability in high-dimensional medical imaging.
TGCM outperforms existing methods by effectively disentangling overlapping APT campaigns, achieving robust separation even in complex interleaved scenarios.
STARE not only prevents policy entropy collapse but also enhances accuracy by up to 8% across diverse tasks, showcasing a new frontier in stable RL training.
Leadership in multi-agent LLM teams only matters when initial consensus fails, challenging the assumption that stronger control always leads to better outcomes.
DeXposure-Claw transforms DeFi risk supervision by integrating structured evidence with LLM decision-making, drastically reducing false alarms.
Clarification F1 scores improved by up to 73% with a novel uncertainty decomposition method, unlocking new capabilities for LLM agents in ambiguous task environments.
ClaMPAPP achieves superior diagnostic performance in pediatric appendicitis by combining LLMs for feature extraction with a robust machine-learning classifier, challenging the efficacy of end-to-end LLMs in clinical settings.
Social intelligence in AI emerges not from isolated skills but through a dynamic coevolution with humans over time.
When complexity in symbolic systems exceeds a threshold, LLMs risk losing semantic understanding, leading to outputs that are misleadingly coherent but fundamentally flawed.
Dynamic analysis can detect all attack classes in ML models while achieving near-zero false positives, outperforming traditional static scanning methods.
JustDiag reveals that accountability in root cause analysis hinges on explicit justification rather than just fluent conclusions.
Mesh inference achieves centralized optimality without sharing internal states, challenging traditional models of collective decision-making.
A strong separation between sign rank and $\mathbb{Z}_2$-index reveals new insights into the complexity of binary concept classes.
Almost no computational overhead while ensuring high integrity verification for cloud-based encrypted control computations.
Forgetting in continual learning is not random; it clusters in a few critical output-space directions, revealing a surprising structure in how models adapt to new tasks.
C2FL restores robust collective adaptation in mobile, privacy-sensitive environments by leveraging spatial clustering and temporal drift mitigation strategies.
BLITZ achieves superior null calibration and faster performance, making it a game-changer for nonparametric conditional independence testing in causal discovery.
Traditional metrics mask critical error patterns in legal AI, but LegalHalluLens reveals a staggering 40-point gap in performance across claim types, offering a clearer path to trustworthy deployment.
