Research map / Large language models / Retrieval and knowledge

Trigger3:Refining Query Correction via Adaptive Model Selector

AAAI 2025 · Large language models › Retrieval and knowledge · approach: Knowledge editing in LLMs · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Trigger3, a large-small model collaboration framework with three BERT-based triggers (correction, LLM, fallback) that adaptively route queries between a small Seq2Edit/Seq2Seq corrector and a fine-tuned LLM for query correction.

Problem

Query correction in search engines: small correction models generalize poorly and LLMs over-correct and are costly, so it is unclear when to use each model and which to choose.

Approach

A model-agnostic adaptive selector with three triggers: a Correction Trigger filters already-correct queries, an LLM Trigger calls the LLM only when the small model fails but the LLM can succeed, and a Fallback Trigger returns the original query when neither model corrects it.

Key results

• Trigger3 with BART and Qwen1.5-7B-Chat reaches Char-F0.5 75.63 on Commercial vs 73.21 for BART alone and 71.30 for Cascading (LLM)/GECToR-based cascade.
• On QQ with BART, Trigger3 achieves Char-F0.5 65.27 vs 61.54 for HybridLLM (best baseline).
• With GECToR, Trigger3 uses LLM on 32.09% of Commercial queries (vs 45.63% for Margin Sampling) while reaching F0.5 74.60 vs 67.48.

Measured effects reported in the paper

Surprising findings

• Direct LLM correction without small-model preliminary rewrites over-corrects and performs worse than the small model alone on these tasks.
• Prompt engineering (zero-shot, few-shot, few-shot CoT) gives very low F0.5 (around 13-19), so LLMs had to be fine-tuned to be competitive.

Limitations

• LLMs fine-tuned on only 1,000 queries and the QQ dataset is a modified public dataset rather than a native query correction benchmark.
• Evaluated only on Chinese query correction with two 7B LLMs.
• Commercial dataset is proprietary and its data release status is not stated.

Experimental setup

Datasets: Commercial (short-video search logs, Kuaishou) QQ (modified public search dataset)

Benchmarks: Commercial QQ

Metrics: character-level P/R/F0.5 (ChERRANT) word-level P/R/F0.5 LLM coverage

Baselines: GECToR BART mT5 Single (LLM) Cascading (LLM) Random Routing Meta Routing HybridLLM Random Cascading Margin Sampling

Builds on: GECToR HybridLLM Margin Sampling BERT LoRA

Closest papers by meaning

• MaFeRw: Query Rewriting with Multi-Aspect Feedbacks for Retrieval-Augmented Large Language Models (AAAI 2025 · similarity 0.84)
• Efficient Evaluation of LLM Performance with Statistical Guarantees (NeurIPS 2026 · similarity 0.84)
• Data Efficient Adaptation in Large Language Models via Continuous Low-Rank Fine-Tuning (NeurIPS 2025 · similarity 0.84)
• HonestLLM: Toward an Honest and Helpful Large Language Model (NeurIPS 2024 · similarity 0.83)
• Who Knows the Answer? Finding the Best Model and Prompt for Each Query Using Confidence-Based Search (AAAI 2024 · similarity 0.83)
• Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs (ICLR 2026 · similarity 0.83)