Skip to content

Interact-RAG: Reason and Interact with the Corpus, Beyond Black-Box Retrieval

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Agentic RAG treats retrieval as white-box; agents reason and interact with corpus

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Interact-RAG replaces the black-box retrieval paradigm in agentic RAG by exposing a Corpus Interaction Engine that gives LLM agents fine-grained, programmatic control over retrieval. An end-to-end agent is trained via SFT on synthesized trajectories followed by GRPO-based RL, achieving a 22.5% relative improvement over the previous best across six QA benchmarks.

Problem

Existing agentic RAG systems treat retrieval as an opaque black box: the agent can only issue natural-language queries to an embedding-based retriever and passively receive text chunks. This forces agents into inefficient query-reformulation loops when initial retrieval fails—they cannot inspect retrieval state, adjust weighting between retrieval strategies, anchor on entities, or prune noisy context. The limitation is especially acute on complex multi-hop questions where evidence is phrased differently from the query or where semantically similar but irrelevant documents cause distraction.

Method

Interact-RAG introduces three interlocking components:

1. Corpus Interaction Engine — exposes a structured action space \(\mathcal{A}_{CI}\) of seven primitives across three categories: - Multi-Faceted Retrieval: semantic_search, exact_search (Full-Text Search via SQLite), and weighted_fusion(w_s, w_e) to blend them. - Anchored Matching: entity_match(entity) to restrict retrieval to a named entity. - Context Shaping: include_docs, exclude_docs, adjust_scale(n) to sculpt the live retrieval context.

2. Reasoning-Enhanced Workflow — a training-free, three-module pipeline: a global planner decomposes the query; an adaptive reasoner issues Proceed or Reflect-&-Refine directives at each step; an executor translates directives into concrete parameterized tool calls. The workflow also serves as a data-synthesis engine, producing 4.8K clean SFT trajectories (only successful traces retained).

3. Two-Stage Training — SFT on synthesized trajectories (token loss masked over retrieved content), followed by GRPO with a sparse outcome reward: $\(R(\tau) = -1 + \mathbf{1}_{\{\tau_\text{valid}\}} + \mathbf{1}_{\{\tau_\text{valid}\}} \cdot \mathbf{1}_{\{y_\text{ans}\}}\)$ The \(-1\) baseline is absorbed by GRPO's group normalization. Total training data: 12K QA pairs on Qwen3-8B (vs. 170K for Search-R1).

Key Contributions

  • Corpus Interaction Engine that dismantles the black-box retrieval abstraction with seven composable action primitives.
  • Hierarchical reasoning workflow (planner → reasoner → executor) enabling zero-shot use and high-quality trajectory synthesis.
  • Two-phase training (SFT + GRPO-based RL) producing a unified end-to-end agent without multi-module inference overhead.
  • Demonstrated data efficiency: competitive performance trained on 12K pairs vs. baselines using up to 170K pairs.

Results

  • Average EM: Interact-RAG (Qwen3-8B) scores 52.8 EM / 62.2 F1, vs. 43.1 / 52.7 for Search-R1 (second best) — +22.5% relative gain on EM averaged across six benchmarks.
  • Multi-hop gains: on MuSiQue, 34.8 EM vs. 25.5 (Search-R1) — +36.4% relative; on 2WikiMultiHopQA, 69.6 vs. 50.9 — +36.7% relative.
  • Single-hop: NQ 51.6 EM (+14.2%), PopQA 56.0 EM (+23.6% relative over Search-R1).
  • Qwen2.5-7B variant: Interact-RAG-7B scores 47.5 avg EM, +10.1% relative over Search-R1-7B.
  • Ablation (EM drops vs. full model on 2Wiki / MuSiQue / PopQA):
  • w/o Interaction: −8.9% / −10.9% / −10.4%
  • w/o SFT: −15.2% / −21.9% / −6.8%
  • w/o RL: −6.3% / −16.9% / −18.6%
  • Training-free workflow outperforms MA-RAG on 2Wiki (60.1 vs. 45.5 EM) without any fine-tuning.

Limitations

  • Retrieval corpus is fixed to a 2018 Wikipedia dump; generalization to heterogeneous or live corpora is not demonstrated in the main experiments.
  • Exact search and entity match primitives build an additional full-text index (SQLite FTS); storage and indexing overhead are not quantified at scale.
  • SFT trajectories are synthesized by Qwen-Plus under the workflow; trajectory quality depends on the teacher model's capability, potentially introducing systematic biases.
  • Baselines using Qwen2.5-7B official checkpoints are compared against an 8B Qwen3 model in the main table; backbone differences partially confound the comparison.
  • The action space is corpus-specific (document IDs, scale parameters); portability to retrieval systems without FTS or document-level metadata is unclear.

Relevance to Agentic AI / LLM Agents

Interact-RAG directly advances the design of tool-using agents by replacing a coarse, single-action retrieval interface with a structured, multi-primitive API—showing that exposing internal retrieval state to the agent substantially improves complex reasoning. The SFT-then-RL training pipeline (trajectory synthesis via a workflow → GRPO policy refinement) is a generalizable recipe for teaching agents fine-grained tool use without large human-labeled datasets. The work demonstrates that data efficiency can be achieved by coupling high-quality synthetic trajectory generation with RL, reducing required data by an order of magnitude relative to comparable end-to-end trained agents. This pattern—structured tool APIs plus hybrid imitation/RL training—is broadly applicable to other agentic domains beyond RAG.