Search papers, labs, and topics across Lattice.
59 papers published across 3 labs.
Turn your Jupyter notebooks into one-click installable desktop apps with LabConstrictor, democratizing access to computational methods for researchers without DevOps expertise.
Despite their general prowess, open-source LLMs still lag behind proprietary models in the nuanced task of dating texts, even after fine-tuning.
Can a dedicated research program keep a smaller, local LLM competitive against global giants in the rapidly evolving AI landscape?
An AI-integrated agile education platform accelerates practice-relevant AI research by closing the theory-practice gap in software development.
Single-domain watermarks are fundamentally insufficient against modern adversarial toolsets, as spatial and latent watermarks exhibit orthogonal vulnerabilities to generative and geometric attacks, respectively.
Turn your Jupyter notebooks into one-click installable desktop apps with LabConstrictor, democratizing access to computational methods for researchers without DevOps expertise.
Despite their general prowess, open-source LLMs still lag behind proprietary models in the nuanced task of dating texts, even after fine-tuning.
Can a dedicated research program keep a smaller, local LLM competitive against global giants in the rapidly evolving AI landscape?
An AI-integrated agile education platform accelerates practice-relevant AI research by closing the theory-practice gap in software development.
Single-domain watermarks are fundamentally insufficient against modern adversarial toolsets, as spatial and latent watermarks exhibit orthogonal vulnerabilities to generative and geometric attacks, respectively.
A fully open-source speech understanding model, OSUM-Pangu, proves that competitive performance is achievable on non-CUDA hardware, challenging the dominance of GPU-centric ecosystems.
Speech-aware LLMs are surprisingly bad at speaker verification, but a simple embedding injection trick closes the gap with dedicated systems while preserving the LLM's language abilities.
A 4B-parameter model, InternVL-U, outperforms 14B-parameter models in multimodal generation and editing, proving that size isn't everything.
GNNs don't just detect time series anomalies better, they also offer a crucial interpretability boost for real-world diagnosis.
Automating the messy process of turning open-source code into LLM tools unlocks a new level of agent capabilities, outperforming even commercial LLMs.
Pretrained ALiBi transformers suffer from a widespread attention collapse that can be surgically repaired to yield a 25% perplexity improvement, suggesting that standard pretraining leaves performance on the table.
Open-source LLMs can now rival proprietary systems in extracting crucial cancer progression data from radiology reports, unlocking scalable analysis while preserving patient privacy.
Forget parameter conflicts: representational incompatibility is the real culprit behind LLM merging failures, setting fundamental limits on which tasks can be successfully combined.
Forget ensembling or retraining: model merging lets you Frankenstein LLMs for specialized skills at minimal cost.
Don't build a domain-specific model just because you can: fine-tuning a general-purpose model can achieve comparable performance on common tasks, saving significant resources.
Mamba-2's efficiency doesn't require custom CUDA kernels: XLA's compiler optimizations are enough to unlock near-optimal performance across diverse hardware.
Open-sourcing a fully reproducible, optimized Band-Split RNN for music separation, this paper reveals the surprisingly large gap between published results and what can be achieved with a faithful reimplementation, even with significant effort.
Open-source TTS gets a serious upgrade with Fish Audio S2, offering instruction-following control via natural language and production-ready streaming performance.
AI-powered cyber reasoning can now find real-world bugs in open-source software thanks to a new framework that liberates DARPA's AI Cyber Challenge systems from their inaccessible cloud origins.
Democratized LLM pre-training is now a reality: Covenant-72B proves you can train a competitive 72B model with untrusted peers over the internet, opening the door to broader participation and reduced costs.
Ditch the da Vinci: this open-source surgical robotics platform brings precision and flexibility to autonomous laparoscopic procedures using standard industrial robots.
Turns out, buying stars and downloads for open-source software doesn't actually trick developers into using it.
IronEngine achieves 100% task completion on file operation benchmarks by decoupling planning quality from execution capability via a novel three-phase pipeline and intelligent tool routing.
Bridging the gap between narrative descriptions and workflow implementations, CoPaLink automatically links bioinformatics tools mentioned in papers to their usage in code, boosting reproducibility.
A 4B-parameter model, Meissa, rivals the performance of much larger proprietary models in medical agent tasks, offering a cost-effective and privacy-preserving alternative for clinical applications.
Tabular foundation models, despite excelling in point estimate benchmarks, need proper scoring rules like CRPS to reliably evaluate their probabilistic regression capabilities, revealing a crucial blind spot in current evaluation practices.
LLMs can automate and improve thematic analysis of qualitative data, achieving expert-level alignment in clinical domains through iterative codebook refinement.
A new 30B open-weight LLM trained on 34 European languages achieves state-of-the-art performance on low-resource languages with significantly less compute, proving that clever training beats brute force.
Bridge the trust gap in cloud-based LLM services with AFTUNE, a practical framework that lets you audit proprietary fine-tuning and inference without prohibitive overhead.
Zero-shot multilingual TTS models stumble when synthesizing Kashmiri, but a script-aware, flow-based adaptation strategy unlocks intelligible speech.
Steer LLMs like never before with AI Steerability 360, an open-source toolkit that unifies input, structural, state, and output steering methods under a common pipeline.
Chain-of-Thought prompting doesn't always improve LLMs' ability to solve discrete optimization problems, and surprisingly, "disordered" datasets can sometimes boost performance on simpler tasks.
AI models can detect injected thoughts, but they often have no idea *what* those thoughts are, relying on content-agnostic anomaly detection and then guessing common concepts.
LLMs often know the answer long before their "reasoning" suggests, wasting tokens on performative chain-of-thought.
Achieve state-of-the-art autopilot performance with a codebase that's significantly leaner and more modular, unlocking faster iteration for robotics researchers.
Forget slow, end-to-end models: building real-time voice agents hinges on a cascaded streaming pipeline, as demonstrated by a new tutorial achieving sub-second latency.
VLMs can now dynamically adapt to changing deployment environments with user-controlled authorization, thanks to a new framework that protects intellectual property while maintaining performance.
MQED-QD offers a unified, open-source workflow for simulating exciton dynamics in complex nanophotonic environments, enabling the rational design of nanoscale architectures.
A terminal-native coding agent, OPENDEV, achieves robust autonomous software engineering by enforcing explicit reasoning phases and prioritizing context efficiency, offering a blueprint for secure and extensible AI assistance.
Data augmentation can boost a TF-IDF model to near state-of-the-art hate speech detection accuracy on certain datasets, rivaling much larger transformer models.
Censored LLMs offer a surprisingly natural and effective environment for stress-testing methods that aim to elicit truthfulness and detect deception.
Uncovered: six distinct archetypes of Public Sector Open Source Program Offices (OSPOs) that reveal how different organizational structures drive OSS adoption and collaboration.
LLM privacy on shared accelerators doesn't have to break the bank: GELO achieves strong obfuscation with only 20-30% latency overhead, defeating common attacks.
A new lattice-based transaction scheme offers financial institutions a post-quantum secure and auditable distributed ledger solution that existing Ring-CT models can't provide.
Forget specialized models: a single segmentation framework, trained on diverse historical maps, now achieves state-of-the-art performance across collections, scales, and regions.
A new model, MUTEX, achieves 60% token-level F1 score on Urdu toxic span detection, providing the first supervised baseline for a challenging low-resource language.
By merging models on the Fisher-Rao manifold, this work achieves stable and accurate LLM merging even with many heterogeneous models, overcoming the representation collapse issues plaguing simpler weight averaging techniques.
Ditch matplotlib for blazing-fast, GPU-powered dimensionality reduction and visualization on your Mac with mlx-vis.
Forget text prompts: vector prompt interfaces are the key to unlocking scalable and stable LLM customization.
A compromised 5G network can hijack or disable UAVs, revealing a major security gap in current UAV communication protocols.
Unleash the power of AI-assisted audio annotation with LabelBuddy, the open-source tool that lets you plug in your own models and build richer, more nuanced music representations.
Two-bit quantization can nearly match the performance of larger models on Polish language tasks, but beware: some methods that look good on paper fail catastrophically when generating text.
LLMs' performance on False Belief Tests isn't just about size – it's profoundly skewed by how you phrase the question, revealing that models learn stereotypical responses to mental-state vocabulary during pre-training.
Think your LLM's code is anonymous? This paper shows you can fingerprint it with high accuracy, even across different programming languages.
Finally, a lightweight, dependency-free Python library streamlines Vietnamese text normalization, handling everything from currency to acronyms without needing GPUs or external APIs.
Nominal techniques, a principled approach to variable binding, are now accessible as a practical Agda library.
Data quality, not just model size, reigns supreme: Phi-4-reasoning-vision-15B proves that smaller, open-weight multimodal models can achieve competitive performance through rigorous data curation and architecture choices.
LLMs aren't just chatbots; they're surprisingly good motivational interviewers, even outperforming human therapists on key metrics and fooling psychiatrists in distinguishability tests.
Ditch gradient sharing: PTOPOFL uses persistent homology to communicate only 48-dimensional topological summaries in federated learning, slashing reconstruction risk by 4.5x while boosting AUC.
