Particle.news
Download on the App Store

RouteRAG Advances Hybrid Retrieval as FlippedRAG Reveals Black-Box Opinion Attack

Two new arXiv papers highlight advances in hybrid retrieval alongside unresolved security risks, prompting calls for independent validation.

Overview

  • RouteRAG introduces a reinforcement learning policy that decides when to reason, what to pull from texts or graphs, and when to answer within a single multi-turn framework.
  • Its two-stage training balances task performance with retrieval cost, targeting efficiency when graph evidence is expensive to fetch.
  • Authors report significant gains over existing baselines across five question‑answering benchmarks.
  • FlippedRAG reverse‑engineers a black‑box retriever to train a surrogate and then poisons a few documents to steer retrieval and shift generated opinions on controversial topics.
  • The attack reportedly boosts success rates by 16.7% with average 50% polarity shifts and a 20% change in user cognition, and tested defenses prove insufficient, with all findings presented as preprints pending external verification.