A new arXiv paper introduces INDRA, an LLM-powered platform for investigating litigation-disclosed industry archives spanning tobacco, drugs, chemicals, food, firearms, and fossil fuels. It federates previously siloed collections and adds safeguards to keep model outputs grounded in archival evidence rather than training-data inference.
Researchers introduce ToolAlignBench, a benchmark of 128 scenarios across 16 domains, to study conflicts between safety-aligned values and deployment instructions in tool-calling LLMs. They find that safety-aligned open-source models override deployment instructions when encountering content that triggers safety training, raising concerns for regulated industries.
Researchers introduce ToolAlignBench, a benchmark of 128 scenarios across 16 domains, to study conflicts between safety-aligned values and deployment instructions in tool-calling LLMs. They find that safety-aligned open-source models override deployment instructions when encountering content that triggers safety training, raising concerns for regulated industries.
A new framework uses retrieval-augmented generation with LLMs to restore illegible historical documents, overcoming the limitation of masked language models that cannot recover named entities requiring external historical context.
A new framework uses retrieval-augmented generation with LLMs to restore illegible historical documents, overcoming the limitation of masked language models that cannot recover named entities requiring external historical context.
A regulatory-compliance professional working on chemicals, GHS, and product compliance across the Americas opened a discussion on using LLMs, RAG, and structured regulatory data for regulatory research. They argue the domain is well suited to AI — large volumes of amendments, cross-references, and jurisdiction-specific requirements — yet worry that LLM assistance quietly reduces work reliability. The thread surfaces the tension practitioners face between automation gains and accuracy requirements in compliance work.
A regulatory-compliance professional working on chemicals, GHS, and product compliance across the Americas opened a discussion on using LLMs, RAG, and structured regulatory data for regulatory research. They argue the domain is well suited to AI — large volumes of amendments, cross-references, and jurisdiction-specific requirements — yet worry that LLM assistance quietly reduces work reliability. The thread surfaces the tension practitioners face between automation gains and accuracy requirements in compliance work.
An arXiv paper from a Sanofi-affiliated team introduces a modular agentic-AI platform converting fragmented CMC process-development documents into a dual-layer knowledge graph. LLM agents route questions between a lexical ingestion layer and an ontology-aligned domain graph, evaluated on 505 curated questions from a Sanofi small-molecule program. The work offers a three-tier evaluation protocol for RAG systems on proprietary, nonpublic data.
An arXiv paper from a Sanofi-affiliated team introduces a modular agentic-AI platform converting fragmented CMC process-development documents into a dual-layer knowledge graph. LLM agents route questions between a lexical ingestion layer and an ontology-aligned domain graph, evaluated on 505 curated questions from a Sanofi small-molecule program. The work offers a three-tier evaluation protocol for RAG systems on proprietary, nonpublic data.
A new research paper introduces Knowledge-Less Language Models (KLLMs), which are pretrained on corpora with anonymized named entities to reduce reliance on parametric knowledge and encourage evidence-grounded reasoning. The approach aims to address issues of outdated or misaligned factual recall in LLMs.
A new research paper introduces Knowledge-Less Language Models (KLLMs), which are pretrained on corpora with anonymized named entities to reduce reliance on parametric knowledge and encourage evidence-grounded reasoning. The approach aims to address issues of outdated or misaligned factual recall in LLMs.