The Archive

Schema-guided reasoning framework for clinical CRF completion via two-stage decomposition handles sparse field prediction with strict output contracts.

LayerTracer framework analyzes hierarchical representations and robustness bottlenecks across diverse LLM architectures including Transformer, GateDeltaNet, and Mamba.

Theoretical analysis of Bayesian softmax-gated mixture-of-experts models with focus on posterior distribution asymptotic behavior.

Cross-lingual quality classification for multilingual pretraining uses embedding-space consistency to transfer filtering strategies from high-resource to low-resource languages.

Multi-agent iterative search framework for scientific idea generation combines combinatorial innovation theory with LLM-based planning to reduce repetitive outputs.

Conceptual critique of generative AI evaluation paradigms argues benchmarks constitute rather than measure models and calls for descriptive alternatives honoring pluralist contexts.

Evian framework provides fine-grained semantic auditing of vision-language model training data to identify logical fallacies and factual errors beyond coarse-grained quality scores.

Scoping review and survey of stuttered-speech research identifies gaps between current systems and end-user needs; recommends grounding speech tech in stakeholder experiences.

Alibaba releases Qwen3.6-27B, open-source 27B dense model with agentic coding surpassing larger models, Apache 2.0 licensed.

Qwen3.6-27B model release announcement on Hugging Face with no additional details.

Multilingual medical QA evaluation shows cross-lingual evidence retrieval and LLM parametric knowledge improve performance on low-resource languages relative to monolingual approaches.

Opus 4.7 benchmark regression reported on SimpleBench relative to 4.6 and 4.5 versions.

LLM evaluation on feature model analysis using semi-formal blueprints shows reasoning-optimized models (Grok 4, Gemini 2.5 Pro) achieve 88-89% accuracy vs solver oracles.

10x Science has raised a $4.8 million seed round to help pharmaceutical researchers understand complex molecules.

Efficient symbolic computation algorithm for identifying causal effects under latent confounding reduces complexity of Gröbner basis approach in linear structural causal models.

CHASM dataset benchmarks MLLMs on detecting covert ads in Chinese social media with 4,992 curated instances.

Deterministic dropout formulation replaces stochastic masking with explicit regularization terms for Transformer training.

Distinct Leaf Enumeration (DLE) method reduces inference compute by deterministically enumerating decode trees vs. sampling with replacement.

We fine-tuned Qwen3-32B into a tool-calling agent that generates Flow automations from natural language—faster, cheaper, and more accurate than the frontier model it replaced, with a weekly retraining flywheel built on real merchant data.

COBALT formal verification tool targets arithmetic vulnerabilities (CWE-190/191/195) in AI sandbox infrastructure post-April 2026 Claude Mythos escape.

Certified malware detector using randomized smoothing and feature ablation provides provable robustness against adversarial evasion.

Decentralized learning framework achieves centralized performance by sharing Gibbs measures under ERM-RER without dataset pooling.

Knowledge Capsules add structured nonparametric memory units to LLMs for stable multi-hop reasoning and knowledge updates without retraining.

Graph neural network predicts network-level traffic flows by modeling heterogeneous bidirectional graphs of IP, Port, and Connection nodes.

Google releases TPU 8t and TPU 8i chips designed for next-generation AI supercomputing efficiency.

Google releases eighth-generation TPUs with two specialized chips designed for agentic AI workloads.

Google wants AI startups on its cloud and has showcased a long list of them at its annual conference.

Generative AI is being infused into Google's popular feature within Maps.

Mira Murati's Thinking Machines Lab has signed a multi-billion-dollar deal with Google Cloud for AI infrastructure powered by Nvidia's latest GB300 chips, TechCrunch has exclusively learned.

Sequential calibration framework for vision-language-action robotics models quantifies uncertainty across partial trajectory observations.