SEAL-RAG: Loop-Adaptive RAG with On-the-Fly Entity Extraction and Fixed-k Gap Repair

ICLR 2026 Conference Submission17581 Authors

Keywords: Retrieval-Augmented Generation (RAG), Multi-hop Question Answering, Entity Extraction, Gap Specification, Fixed-k Retrieval, Micro-Queries, Evidence Precision@k, Sufficiency Checking, Training-Free Inference, Open-Domain QA, Optimization

TL;DR: Training-free, fixed-k RAG controller that repairs gaps via on-the-fly entity extraction and micro-queries, replacing distractors instead of expanding context to boost answer correctness and evidence precision@k.

Abstract:

We propose SEAL-RAG, a training-free, inference-time controller (no fine-tuning
of retriever, reranker, or generator) for retrieval-augmented generation that targets
multi-hop precision. SEAL executes a fixed retrieval depth k (k = number of
passages retrieved per search/micro-query) in a Search → Extract → Assess →
Loop cycle. A scope-aware sufficiency check aggregates coverage, typed bridging,
corroboration/contradiction, and answerability signals to decide stop vs. targeted
repair. At each loop, SEAL performs on-the-fly, entity-anchored (head, relation,
tail) extraction, maintains a live entity ledger, and builds a gap specification (miss-
ing entities/relations) that triggers one micro-query per repair under the same top-k;
new candidates are merged via entity-first ranking (prefers passages anchoring
those entities) before a single final generation step. On a 1,000-example HotpotQA
validation subset in a shared setup, SEAL improves LLM-judged answer correct-
ness by +10–22 pp (k=1) and +3–13 pp (k=3) vs. SELF-RAG across backbones,
and increases evidence precision@k (gold-title precision) by +12–18 pp at k=3.
These gains are statistically significant (chi-square for correctness; paired two-sided
t-tests for precision/recall/F1; p<0.05). By keeping k fixed and bounding repairs
by T (maximum repair iterations), SEAL yields a predictable, bounded cost profile
while replacing distractors rather than broadening context.

Primary Area: optimization

Submission Number: 17581