LightThinker

Compresses chain-of-thought into learned gist tokens to cut reasoning cost (EMNLP 2025), with 2026 LightThinker-Plus checkpoints (Qwen3-30B-A3B, Llama) released on HuggingFace. Sibling LightMem (ICLR 2026, ~1K★) brings lightweight memory-augmented generation to LLM agents. From ZJUNLP (Ningyu Zhang).