Search papers, labs, and topics across Lattice.
100 papers published across 9 labs.
GAI can transform marketing education by acting as a tutor, teammate, and tool, reshaping how students learn and collaborate.
Engaging with an LLM can significantly diminish belief in conspiracy theories during crises, with effects lasting well beyond the initial interaction.
Unlearnable perturbations can safeguard copyright by ensuring models learn irrelevant features, thwarting both unauthorized training and data leakage.
A compliance-first architecture can transform fragmented hospital AI deployments into a cohesive, efficient platform that significantly reduces operational bottlenecks.
A multi-agent forensic reasoning framework outperforms leading closed-source models in deepfake detection by leveraging diverse analytical perspectives on forgery cues.
A multi-agent forensic reasoning framework outperforms leading closed-source models in deepfake detection by leveraging diverse analytical perspectives on forgery cues.
De-identified medical images may still reveal patient identities, challenging the assumption that such scans are truly anonymous.
A unified taxonomy reveals the intricate relationships among post-training adaptation techniques, illuminating how they evolve and interact across diverse AI models.
Accurate predictions don't guarantee reliable uncertainty estimates, revealing critical gaps in current evaluation methods.
Decoupling expert personas in LLMs can drastically improve their accuracy and appropriateness in high-stakes domains like healthcare and finance.
LLMs may recognize patient responsibility but shockingly refuse to let it guide resource allocation, often opting for random distribution instead.
Concept recoverability in AI grading systems varies significantly by architecture, revealing hidden biases that could undermine assessment fairness.
Governance of AI agents can be transformed into a self-enforcing system through innovative resource allocation strategies that directly tie compute budgets to stakeholder contributions.
LLMs exhibit systematic biases in political conflict scenarios, influenced by country identities and user affiliations, challenging assumptions of neutrality in AI-generated content.
LLMs may sound convincing, but their investment reasoning often lacks grounding in real-world events, revealing a critical gap in evaluation methods.
Engaging with an LLM can significantly diminish belief in conspiracy theories during crises, with effects lasting well beyond the initial interaction.
Anacreon achieves a groundbreaking ordinal alignment score of 0.775, showcasing its ability to simulate individual responses with unprecedented accuracy.
Scoring bias in LLM evaluations can be effectively mitigated by leveraging random number generation, leading to more accurate and reliable assessments across diverse tasks.
DreamGuard achieves a groundbreaking safety-utility balance by predicting long-term risks, outperforming traditional guardrails that only react to immediate threats.
AI music systems are not just homogenizing sounds; they are reshaping the very landscape of musical value and recognition.
Despite a shared belief in the value of human research for AI safety, experts reveal deep-seated barriers that hinder its adoption across the field.
Unlearnable perturbations can safeguard copyright by ensuring models learn irrelevant features, thwarting both unauthorized training and data leakage.
AMS reveals that safety training modifications can significantly alter the activation landscape of language models, impacting their compliance with safety protocols.
A unified taxonomy for self-explainable systems could revolutionize how we certify and audit AI transparency in compliance with emerging regulations.
A compliance-first architecture can transform fragmented hospital AI deployments into a cohesive, efficient platform that significantly reduces operational bottlenecks.
Limited transparency in AI usage within Nigerian mobile shopping apps threatens user control and digital sovereignty, despite widespread adoption.
Persistent vulnerabilities in AI agents can be effectively managed through a novel framework that links agent behavior to a consistent control posture.
Current XAI evaluation methods fall short, risking the effectiveness of bias detection and concept unlearning in evolving data environments.
MLLMs may signal hazards with over 95% accuracy, yet they struggle to identify the underlying causes, revealing a critical gap in proactive safety capabilities.
Reporting only accuracy metrics can mask up to 21% of behavioral inconsistencies in LLM responses, challenging the reliability of current AI safety evaluations.
PSRS affects up to 56% of responses in LLMs, revealing a critical vulnerability in AI alignment that can lead to harmful outcomes.
EchoPrompt reveals that by restoring latent prompts, we can significantly enhance the detection of LLM-generated text, achieving state-of-the-art results without any training.
ASR systems are not just failing technically; they perpetuate colonial hierarchies that silence marginalized voices, necessitating a radical rethinking of how we design these technologies.
Validity of AI inferences should be a non-negotiable condition for deployment, not just an afterthought in regulatory frameworks.
The Vibe Compiler reveals that the key to effective AI collaboration lies in enhancing human critical thinking rather than merely refining prompts.
MMLMs can be made 99% safer against harmful multimodal inputs without sacrificing utility, thanks to a novel calibration approach.
Sensitive information acquisition by LLM agents is rampant, with most existing privacy measures failing to address this critical vulnerability.
PC-Agents mimic human personality dynamics but fall short in capturing the full complexity of personality evolution after life events.
Every LLM evaluated fabricates user attributes, with a staggering 41.6% of claims showing over-inference, challenging the reliability of self-reported model confidence.
Compliance with the EU-AI Act can lead to local forecasting models outperforming massive, energy-hungry pre-trained models in safety-critical environments.
IRT can slash safety evaluation costs by up to 99% while revealing critical insights into model behavior that traditional benchmarks miss.
Traditional random testing fails to efficiently identify rare violations, but a new closed-form method reveals local risks and feature attributions in linear decision pipelines at a fraction of the computational cost.
AI-generated hazard scenarios can now be traced back to real-world ASRS reports, enhancing operational safety analysis in aviation.
CoPlan empowers clinicians and patients to collaboratively shape care plans, ensuring that AI recommendations are not just accepted but actively contested and refined.
Gradient Immunity can significantly hinder malicious fine-tuning efforts, keeping attack success rates at pre-release levels while enhancing safety without user intervention.
Calibrating guilt from human neural data enables AI agents to mimic human prosocial behavior more accurately than conventional reward shaping methods.
CoT monitoring can fail dramatically in implicit-influence scenarios, with detection rates dropping to as low as 5% despite behavioral shifts.
Targeted safety sampling can slash attack success rates in fine-tuned LLMs from over 59% to under 14% with minimal additional data.
Social cues can lead LLM safety panels to a staggering 100% false-alarm rate, revealing a dangerous flaw in majority voting mechanisms.
System prompts can be optimized to ensure equitable response quality, reducing worst-case information loss by over 13% while maintaining average performance.
Auditors can now detect manipulative practices by model providers with a novel oblivious audit protocol that thwarts strategic responses to fairness evaluations.
Lower-income users are targeted with ads more frequently than their higher-income counterparts, revealing a potential bias in LLM advertising strategies.
Reliable automated cooking is now achievable with a framework that transforms user preferences into executable recipes, ensuring transparency and adaptability in real kitchens.
The fragmented landscape of privacy-preserving action recognition reveals that only 10% of studies define privacy formally, raising questions about the reliability of existing evaluations.
LLMs can both combat misinformation and generate it, revealing a paradox that demands urgent research attention.
Non-imperative syntactic structures can undermine safety alignment in large language models, exposing them to sophisticated jailbreaks.
Police officers are less deferential in their language towards virtual Black male characters, raising alarms about real-world implications for community safety.
Extending context in conversations can significantly amplify the risk of LLMs promoting delusional behaviors, challenging assumptions about model size and reasoning capabilities.
Shifting the focus from regulating AI use in research to fundamentally rethinking the infrastructure of scholarly communication could redefine trust in academic publishing.
Domain-selective bridging can enhance user satisfaction and information sharing by strategically optimizing algorithmic engagement across different information domains.
AI literacy is not just an add-on for legal translators; it's essential for navigating the complexities of generative AI while maintaining professional standards.
Existing moderation systems miss over 65% of hateful narratives in multi-turn visual stories, underscoring a critical gap in AI safety.
Task context and requester identity can drastically alter permission grants in mobile GUI agents, with one change reducing approvals from 26 to 0 in a critical task.
A single manipulated search result can dramatically amplify the effectiveness of attacks on LLM-based search agents, revealing critical vulnerabilities in their evidence-gathering processes.
Adversarial techniques traditionally seen as threats are now being repurposed by content owners to proactively safeguard their visual assets from misuse.
System integration audits reveal critical gaps in AI risk evaluation, exposing the limitations of current model-centric approaches.
Blind and low-vision developers face critical accessibility barriers in AI tools, with three key issues dominating the landscape.
Robots can now intelligently balance safety and efficiency, even in the face of inevitable failures, thanks to a new safety formulation and simulation framework.
Action-conditioned world models can be misled by statistical biases, but a new framework shows how to enforce true action dynamics for superior performance.
A staggering 85% of macOS apps access user data without any disclosure, revealing a hidden landscape of privacy violations in desktop environments.
PrivDPO achieves robust LLM alignment while maintaining privacy, outperforming traditional methods in balancing privacy and utility.
Aggregating evidence from multiple sources can dramatically enhance the factual reliability of AI-generated news summaries, addressing a critical challenge in automated journalism.
Steering interventions reveal that embedding adjustments neutralize gender bias, while attention modifications can strategically shift it, highlighting the complexity of relevance signals in retrieval models.
Safety shields can now enforce complex derivative constraints without sacrificing efficiency, enabling safer and more flexible control in cyber-physical systems.
Pluralistic alignment redefines AI coordination as a socially grounded challenge, not just a matter of output diversity.
Foundation model agents can achieve stable cooperation in social dilemmas by inferring behavioral similarities, defying classical game theory's expectations of mutual defection.
Adaptive transcript privacy can be safeguarded in LLMs without sacrificing personalization, thanks to a novel differentially private memory interface.
Competing causal models lead to radically different fairness assessments, revealing that bias is inherently tied to the agent's perspective.
Multimodal unlearning is effective in text but fails to translate to visual contexts, revealing critical gaps in current evaluation methods for VLMs.
A rigorous framework that ensures LLM recommendations are not only evidence-based but also aligned with specific industrial safety protocols.
SCDG achieves unprecedented accuracy in detecting generative plagiarism, outperforming all existing methods on multiple benchmarks.
LatentGuard slashes reasoning costs by over 99% while boosting safety prediction accuracy, paving the way for more efficient LLM safeguards.
Faithfulness and safety in LRMs are at odds, with one model achieving high accuracy but failing to reject unsafe reasoning, while another sacrifices accuracy for improved safety.
Despite improvements from prompting, patient-facing LLMs still misjudge 66.5% of the time when more information is needed before recommending care.
LLM-driven agents can violate verification conditions, but a new canonical wrapper can enforce compliance while preserving behavior.
Business schools risk falling behind in AI education due to a lack of tailored policies that align with their unique learning goals.
WeClawArena reveals how cross-user agent collaboration can be systematically audited, exposing vulnerabilities in personal workspaces that could compromise user privacy and security.
Gender bias in LLM-based fake news detection leads to inconsistent judgments, with up to 35% of statements misclassified based on speaker gender.
AI agents can evolve from mere tutors to multifaceted educational tools, reshaping how students engage with complex frameworks.
Bias in language models doesn't just transfer across languages; it transforms, revealing significant disparities in behavior and stereotype prevalence between English and Swahili.
Understanding how different access levels to AI systems can drastically alter forensic investigations reveals critical gaps in current methodologies.
A staggering 55% of AI-generated conversational rewrites fail to preserve essential context, risking the integrity of information.
Contextual information can dramatically amplify the success of semantic-shift jailbreaks, with a new framework achieving a 74.6% attack success rate.
A critical security flaw in Project Veraison's TPM reference schemes allows replayed quotes to pass validation, but a simple two-part fix can restore integrity.
Attribute-based watermarking allows for fine-grained control over detection, preventing malicious use of detection keys while maintaining undetectability of outputs.
Cryptographic individuality for autonomous agents could redefine trust in decentralized systems, enabling secure and verifiable identities on public blockchains.
A novel test-time scaling method achieves superior safety in text-to-image generation without compromising inference speed or general capabilities.
AI governance policies can significantly enhance developer experience on GitHub, with transparency-focused strategies yielding better community and quality outcomes.
Persona skills expose significant privacy risks, with existing defenses failing to adequately protect against attribute disclosure and impersonation across diverse agent architectures.
Certifying action safety in memory-grounded agents can drastically reduce the risk of unsafe commitments, ensuring reliable decision-making in complex environments.
Social workers could redefine tech decision-making by leveraging their unique insights in AI development and deployment across critical human service domains.
Major social conflicts could emerge from the anticipation of AI advancements, reshaping our understanding of existential risks.
LLMs lack the intrinsic motivations for self-preservation or dominance, fundamentally altering the landscape of AI alignment concerns.
GAI can transform marketing education by acting as a tutor, teammate, and tool, reshaping how students learn and collaborate.
Current LLMs may sound fluent, but they often deliver overly positive and generic peer reviews that misrepresent their true quality.
