DARE‑Agent: Domain‑Aware, Resource‑Efficient, Evidence‑grounded Agentic RAG

ICLR 2026 Conference Submission17258 Authors

Keywords: Agentic RAG, Agents, DPO

Abstract: LLM agents for deep research have advanced open‑domain reasoning, yet deployments in specialized domains still fail along three critical axes: unverifiable answers, uncontrolled cost, and domain‑agnostic retrieval that undermines authority/recency. Prevailing evaluations focus narrowly on answer accuracy, overlooking process‑level metrics such as citation correctness, minimal sufficient evidence (MSE), and the accuracy–cost trade‑off, while many training setups rely on complex, hard‑to‑reproduce online RL. We reframe research‑agent quality as a multi‑objective problem spanning accuracy, verifiability, and resource efficiency, and introduce DARE‑Agent, a domain‑aware, resource‑efficient, evidence‑grounded agentic RAG framework. DARE‑Agent integrates a learnable Domain‑Aware Gating mechanism into a short, auditable trajectory: the agent proposes domain‑conditioned controls over retrieval and evidence, and an executor clips them to safe ranges. Training combines SFT with Direct Preference Optimization over multiple sampled trajectories, using a composite preference score that balances accuracy, verifiability, cost, and redundancy; retrieved tokens are loss‑masked for stability. In a reproducible fixed‑corpus setting plus small live‑web subsets, DARE‑Agent delivers competitive accuracy while consistently improving citation precision, reducing MSE, and yielding stronger accuracy–cost Pareto fronts under matched budgets; it also raises authority/recency hit rates.

Primary Area: generative models

Submission Number: 17258