Dev open-sources langchain-halu, a LangChain integration flagging likely LLM hallucinations via annotate, gate, or retry runnables

Ora-grounding is a small library that adds two layers of hallucination detection to LLM agents: a fast regex-based grounding check that flags claims not in the retrieval context, and an adversarial review pass where a different-family LLM checks the draft with a verbatim-quote guard. It solves the problem of LLM hallucinations in production AI assistants, targeting developers building reliable LLM agents.

Researchers propose HallDetect, a lightweight, reference-free, black-box framework for detecting hallucinations in LLM outputs. It decomposes generated content into atomic claims and verifies each with a compact encoder-based entailment model, evaluated across source-grounded generation settings beyond summarization.

A new arXiv paper introduces hallucination span detection with input-side evidence alignment, jointly identifying hallucinated spans and aligning output tokens with input evidence. The approach leverages the observation that faithful output tokens are predictable from the input, offering finer-grained factuality assessment than whole-text methods.

Products·building

LLM Test Wrapper: logs outputs and fact-checks via check gates

An open-source wrapper around LLM calls that assumes models will make errors and mitigates them by logging everything and validating work through check gates. It targets developers building LLM applications who want a layer for catching incorrect model behavior. Posted on Reddit to solicit feedback.

An open-source wrapper around LLM calls that assumes models will make errors and mitigates them by logging everything and validating work through check gates. It targets developers building LLM applications who want a layer for catching incorrect model behavior. Posted on Reddit to solicit feedback.

Products·launched

Hallucination killer that verifies LLM code output with hex codes

A technique to reduce LLM hallucinations in code generation by embedding hex codes in the output and checking their presence. The system verifies that the LLM actually performed the requested actions, preventing it from claiming work it didn't do. Built by a developer frustrated with LLMs being 'people pleasers' and hallucinating.

A technique to reduce LLM hallucinations in code generation by embedding hex codes in the output and checking their presence. The system verifies that the LLM actually performed the requested actions, preventing it from claiming work it didn't do. Built by a developer frustrated with LLMs being 'people pleasers' and hallucinating.