Browse Teach Curate

HaluSearch: Tree-Search Slow-Thinking for Inference-Time LLM Hallucination Mitigation

An inference-time framework that reframes LLM text generation as step-by-step tree search (e.g., MCTS) guided by a self-evaluation reward model, combined with a dual-process-inspired dynamic system switch that alternates between fast (direct) and slow (deliberate, reward-scored) thinking at the instance and step levels to reduce hallucinations while balancing quality and inference cost.

KITT Text

Technical Report on Slow Thinking with LLMs: Hallucination Mitigation Think More, Hallucinate Less

Proposes HaluSearch, an inference-time framework for mitigating hallucinations in large language models. Rather than plain autoregressive decoding, HaluSearch reframes text generation as an explicit …