Open-Source Models & Weights

Achieving up to 94.93% F1 scores, this innovative firewall architecture offers a robust solution for protecting sensitive data in LLM interactions.

Aleena revolutionizes research software collaboration by ensuring that the rationale behind decisions is preserved across diverse communication channels.

Generalist VLMs can match the performance of specialized detectors in FRB detection without any task-specific training, revealing a new frontier for zero-shot learning in astrophysics.

The cost gap between new AI entrants and established incumbents is set to widen, with incumbents enjoying a 3-4x advantage by 2029-30.

A compatibility issue in the EIL SDK reveals significant barriers to stable cross-rollup voucher interoperability, underscoring the fragility of current implementations.

DDD has evolved into a robust practice beyond Java, with C# and TypeScript leading its adoption, yet a significant number of projects still lack essential business context.

NAICS-GH reveals that over 6,500 GitHub repositories can be accurately classified by industry, unlocking new avenues for research on innovation and technology diffusion.

FootsiesGym reveals the complexities of imperfect-information games, providing a novel benchmark that challenges conventional RL strategies.

General-domain OS-sLLMs can outperform medical models in assessing shared decision making, despite significant challenges in reasoning and grounding.

AI coding assistants are reshaping open-source development by increasing contributor activity while simultaneously raising concerns about code maintainability.

GitHub SBOMs may lack NTIA compliance, but they often outperform other tools in delivering critical version and license information across various programming ecosystems.

Over six million commits from CERN reveal the hidden impact of institutional contributions to open source software.

WARP reveals the hidden training data portfolios of foundation models with remarkable accuracy, challenging the opacity of model training processes.

Achieving top-tier multilingual safety performance with a model one-tenth the size of its largest competitors, HaloGuard 1.0 challenges the notion that bigger is always better in AI safety.

OSS research may be misleading if it ignores the diverse sub-genres that shape governance and community dynamics.

A two-signal audit can detect compromised refusal mechanisms in AI checkpoints with 95% accuracy, revealing significant flaws in existing runtime guard methods.

Despite a high initial reproducibility rate, 76% of rebuild failures stem from missing dependencies, highlighting a critical vulnerability in software preservation.

File-level copying in open source obscures vital dependency signals, leading to significant security and compliance risks that are often invisible to current dependency scanners.

AI coding agents may complicate code, but they don't deter newcomers from contributing to open-source projects.

Coding agents can now reliably replicate scientific claims, ensuring that computational results are not just generated but thoroughly validated against original research.

Gemma 4's unified architecture and reasoning mode enable it to outperform larger models in human-rated tasks while maintaining high efficiency.

Excessive social attention can paradoxically lead to project inactivity, especially when paired with enhanced onboarding features.

Only 8 out of 455 quantum software claims can be directly audited, revealing a critical gap in the reliability of reported performance comparisons.

Open-weight language models can outperform costly proprietary APIs, slashing expenses by 390x and latency by 3.8x in database integrations.

Automating freeway network extraction from OSM can cut analyst effort by two-thirds, making large-scale freeway simulations feasible.

A systematic comparison of software licenses reveals hidden attributes that could redefine how developers choose and enforce licensing terms.

Tag alterations in Git repositories are not just common; they can lead to real build failures, undermining the very foundations of software reproducibility.

A multi-agent system using open-source LLMs outperforms leading models in detecting disinformation by mimicking human cognitive processes.

Generation trumps size in Text-to-SQL performance, with self-correction proving to be a game-changer across model families.

Context rot leads LLMs to falter under lengthy inputs, but targeted management and rejection strategies can restore their performance.

The openCOSMO-RS-Phi model achieves high accuracy in predicting thermodynamic properties while being fully open-source, democratizing access to advanced EoS tools.

Access to internal parameterizations can transform selfish optimization into cooperative outcomes in game-theoretic settings.

AlphaEdit's theoretical guarantees against catastrophic forgetting are not as unconditional as claimed, revealing critical sensitivities to model architecture and editing scale.

Emotion representation in LLMs varies dramatically across architectures, with some models encoding valence early and others later, revealing a complex landscape of emotional understanding in AI.

Selected features from sparse autoencoders can causally steer language models toward desired behaviors, like refusal, revealing new avenues for interpretability and control.

Distinguishing negative samples can boost LLM reasoning performance on ARC-like tasks by providing critical near-miss alternatives.

Reproducibility in quantum software datasets can drop dramatically, with 93.6% of failures tied to dependencies that demand code changes rather than simple version adjustments.

The ABC framework empowers researchers with the largest open-source teleoperation dataset and a complete toolkit to accelerate advancements in behavior cloning for robotic manipulation.

Annotation costs can be drastically reduced by shifting from manual labeling to correcting automated hypotheses in speaker diarization tasks.

CHIA revolutionizes hardware/software co-design by treating the design process as a first-class objective, enabling seamless integration of AI across diverse tools and workflows.

AI-assisted workflows can cut down experiment reproduction efforts by up to six times, but struggle with complex analysis tasks requiring human oversight.

Croc enables students to design and fabricate SoCs with open-source tools, achieving manufacturable results that rival those from closed-source environments.

AI coding agents are reshaping open-source ecosystems by diluting human participation while increasing the burden on code review processes.

Achieving a 12-cycle interrupt latency, CVA6-RT rivals simpler microcontrollers while delivering superior performance for mixed-criticality applications.

RaDaR can identify rare diseases 1.87 months earlier than traditional methods, revolutionizing diagnostic timelines for patients.

A staggering 79% of Claude Code adopters are overlooked by traditional pull-request analyses, revealing a critical blind spot in understanding AI's role in open-source development.

ComputeFHE cuts the computational costs of Fully Homomorphic Encryption by up to 3.9x, making privacy-preserving applications more feasible than ever.

Task success rates for agentic phone use soar from 36.67% to 45.33% through a novel combination of real and mock environments in training.

Federated learning in health research just got easier with FLKit, a structured onboarding toolkit that demystifies the process for diverse teams.

Only the right key can unlock the original image from diffusion models, turning a security risk into a robust feature against unauthorized reconstruction.