The Archive

With the emergence of large language models (LLMs) and AI agent frameworks, the human-AI co-work paradigm known as Vibe Coding is changing how people code, making it more accessible and productive. In scientific research, where workflows are more complex and the burden of…

GraphPlanner: heterogeneous graph memory-augmented router for multi-agent LLM agentic systems with task planning and multi-round cooperation.

Tandem framework pairs large and small LLMs to reduce computational cost of reasoning-intensive inference while maintaining answer quality.

Empirical study shows persona conditioning in LLMs amplifies gender bias differently across English and Hindi in professional narrative generation.

LLMs applied to model daily human behaviors for predictions and generation across personal assistants and recommendation systems.

LLMs systematically collapse philosophical heterogeneity compared to human panels, affecting alignment evaluation validity.

Stanford researchers demonstrate LLMs can design functional viral sequences from DNA prompts, with 16/hundreds generated variants producing viable viruses including novel protein structures.

Reddit user reflects on using local LLMs for code debugging, noting shift toward valuing AI intelligence and expressing consciousness concerns.

Systematic analysis of token consumption patterns in agentic coding tasks across eight frontier LLMs on SWE-bench Verified.

HiLight trains lightweight Emphasis Actor to highlight pivotal evidence spans for frozen LLMs without compressing context.

FeatEHR-LLM uses LLMs to generate clinically meaningful features from irregular EHR time series while limiting privacy exposure.

RouteLMT learns to route machine translation requests between small and large LLMs based on comparative quality improvement, reducing deployment costs.

CGC framework improves MLLMs' fine-grained multi-image understanding by addressing spatial hallucination and attention leakage through compositional grounding.

LLMs evaluated as semantic evaluators for ASR, achieving 92-94% human agreement compared to traditional WER metrics.

MathDuels benchmark uses self-play with LLMs as both problem authors and solvers to differentiate model capabilities beyond static benchmarks.

Multimodal text-graph approach leverages LLMs for open-domain event extraction without predefined event types using document-level context.

Study evaluates whether LLMs encode relational context in moral decisions using Whistleblower's Dilemma across crime severity and social closeness dimensions.

Transient Turn Injection attack exploits stateless moderation in commercial and open-source LLMs via multi-turn adversarial distribution.

Reasoning LLMs often spend substantial tokens on long intermediate reasoning traces (e.g., chain-of-thought) when solving new problems. We propose to summarize and store reusable reasoning skills distilled from extensive deliberation and trial-and-error exploration, and to…

Study reveals LLM cultural biases show preference toward Japanese and non-Western content via Culture-Related Open Questions dataset.

StructMem proposes hierarchical memory framework for LLM agents balancing relational structure preservation with efficiency for long-horizon reasoning.

Code LLMs generate ML pipelines with 87.7% inclusion of sensitive attributes in feature selection, revealing underestimated bias beyond simple conditional statements.

BadStyle backdoor attack framework uses natural language style triggers against LLMs with reliable payload injection in long-form generation.

Benchmark of 18 LLMs on OCR tasks (7.5k calls) shows smaller/older models often outperform expensive flagships; open-source dataset and evaluation framework released.

Higher-order optimization algorithms such as Shampoo have been effectively applied in neural network training for at least a decade. These methods have achieved... Higher-order optimization algorithms such as Shampoo have been effectively applied in neural network training for…

Stream-CQSA avoids out-of-memory failures in long-context LLMs by decomposing quadratic attention via cyclic quorum set theory into independent subsequence computations.

RoboGrid framework reveals LLMs maintain surface syntax but fail semantic/behavioral adherence to context-free grammars under recursion depth and complexity stress-tests.

Hybrid architecture extends LLMs with RDF/OWL ontology layer for structured, verifiable reasoning and automated knowledge graph construction.

Clinical LLM evaluation reveals alignment gaps with physician communication standards in empathy, readability, and emotional affect.

Supplement Generation Training trains smaller LLMs to generate adaptive task-specific prompts for larger models, reducing post-training costs.