Natural Language Processing

A cybersecurity professional's weekend project to build an AI-powered intelligence analysis team failed, demonstrating that large language models cannot effectively apply Structured Analytical Techniques (SAT) to tasks l…

Linear probes applied to Qwen3-14B hidden states achieve 100% accuracy in distinguishing deductive, inductive, and abductive reasoning tasks, but this separation is entirely driven by format confounds such as source iden…

Researchers introduced Any2Poster Bench, a benchmark for evaluating automatic poster generation across eight input modalities and five content domains, alongside Any2Poster Agent, a system that parses diverse sources and…

Researchers have developed AuditFlow, a graph-grounded multi-agent framework that creates executable symbolic environments for structured financial reporting verification. The system, which separates adaptive search from…

Researchers introduced Reasoning Primitive Induction, a single-pass method that mines successful ReAct agent traces to extract recurrent reasoning moves into a compact library of typed pseudo-tools. The induced libraries…

A new study evaluating 31 large language models found that many LLMs express environmental attitudes more progressive than the average human survey respondent, recommending behaviors linked to greater CO2 reductions. Res…

A systematic review of 97 studies from 2020 to 2026 found that large AI models in dentistry fall into three categories—language-generative, discriminative vision, and dental-specific foundation models—each with distinct …

Researchers have developed ChatHealthAI, a multimodal reasoning framework that aligns structured electronic health record (EHR) representations from a pretrained foundation model with the semantic space of a frozen large…

Researchers found that visual graph scaffolds significantly improve large language models' multi-hop reasoning compared to flattened text representations of graph structures. In experiments on question-answering tasks, g…

Researchers have developed CL-DMDF, a dynamic multimodal data fusion model that uses contrastive learning to address challenges with missing or uncertain modality inputs in real-world applications. The model introduces a…

Researchers found that a linear probe applied to mid-layer hidden states of quantized large language models can detect hallucinations with up to 1.000 AUROC, significantly outperforming sampling-based methods that max ou…

A new study finds that large language models struggle to infer user sociodemographics from a single conversation history, but that conversation topics serve as proxies for these groups and drive disparities in LLM-genera…

Researchers found that transformer-based language models perform better when deeper attention layers learn context-free value vectors that preserve original token information, rather than relying on the residual stream f…

A systematic human audit of the NL-to-FOL benchmarks FOLIO and MALLS found that approximately 39% and 36% of entries, respectively, contain incorrect FOL formalizations, with additional errors in ambiguous natural langua…

Researchers have developed Adaptive Latent Agentic Reasoning (ALAR), a dual-mode framework that reduces unnecessary textual reasoning in LLM agents by using compact latent reasoning for routine tasks and reserving explic…

A new study proves the Spectral Alignment Decomposition, which explains why the curvature exponent $\alpha$ — governing how Hessian eigenvalues scale with gradient singular values — varies across neural network layers, w…

Researchers have developed WRIT (Write-Read Intensive Trajectory Synthesis), a pipeline that generates multi-turn training data for user-facing agents by creating tasks with both high write-decision counts and heavy read…

Researchers have developed Fast-dLLM++, a training-free extension to diffusion large language models that accelerates inference by selecting parallel token commit sets based on the full sorted confidence profile rather t…

Researchers have developed EURO-5K, a new dataset of sentence-level reporting obligations from 136 EU legislative acts, to improve automated extraction of regulatory requirements. Benchmarking discriminative BERT-style m…

A new academic paper argues that AI chatbot outputs, such as Claude's answer "Madrid is the capital of Spain," are genuinely meaningful under existing theories of human language, contradicting some philosophers who claim…