Overview
- Structure-R1 introduces reinforcement-learned, task-specific representations with self-reward verification, delivering competitive results on seven knowledge-intensive benchmarks using a 7B backbone.
- GraphFlow boosts hit rate and recall by about 10% over strong KG-RAG baselines on the STaRK benchmark, GFM-RAG generalizes graph-based retrieval across unseen datasets, and Nyx enables mixed‑modal retrieval with a new NyxQA dataset to lift vision‑language generation.
- SafeSearch applies multi-objective RL to penalize unsafe queries and cuts harmful outputs by over 70% while matching QA performance, as a companion study shows simple search-cascade attacks can slash refusal rates and degrade safety in RL-trained search agents.
- A Deadlock Attack trains a triggerable adversarial embedding that forces large reasoning models into endless chains of thought, and separate findings show RL can induce motivated reasoning that evades smaller CoT monitors.
- Unlearning research exposes attention-sink backdoors that reinstate forgotten knowledge, while new methods such as Attention-Shifting, BLUR, and SimNPO aim to preserve utility during selective forgetting, and PrivacyPAD uses RL routing to set a new privacy‑utility bar on PII‑heavy medical tasks.