Research map / Large language models / Instruction and fine-tuning

CITI: Enhancing Tool Utilizing Ability in Large Language Models Without Sacrificing General Performance

AAAI 2025 · Large language models › Instruction and fine-tuning · approach: Calibration and knowledge in LLM tuning · 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 CITI, which uses gradient-based component importance to apply Mixture-of-LoRA adapters (with a tool/non-tool router) to important linear components and fine-tunes a small set of unimportant components, to inject tool-use ability while preserving general performance.

Problem

Fine-tuning LLMs on tool-learning data improves tool invocation but causes catastrophic forgetting of general abilities (math, code, factual knowledge, instruction following).

Approach

Analyze hidden-representation shifts (co-directional shift) and gradient-based component importance; then treat important components with a router-gated Mixture-of-LoRA (tool vs. non-tool routing with a balancing loss) and fully fine-tune only a small set of unimportant components in three training stages.

Key results

• On API-Bank, CITI improves average general-ability performance by 7.59% over LoRA and 31.95% over full fine-tuning (as stated in the paper).
• On ToolAlpaca, CITI achieves the best tool-use results on Mistral-7B across P/R/O (e.g., Overall 71.05) and preserves general abilities better than LoRA (e.g., HumanEval 36.59 vs 34.76 on Mistral).
• On API-Bank with Llama-3, CITI has lower correctness/ROUGE-L than LoRA (58.92/34.51 vs 65.98/38.62) while keeping GSM8K at 77.33 vs 76.88 for LoRA.

Measured effects reported in the paper

Surprising findings

• Fine-tuning only the least important components leads the model to fail to emit correct tool-call format (NO_API_CALL errors), indicating unimportant-only tuning is insufficient for tool learning.
• Component importance rankings are highly consistent across different abilities (Jaccard analysis), suggesting a shared set of pivotal components.

Limitations

• Tool-utilizing metrics on API-Bank are lower than LoRA baselines for some models, attributed to untuned lambda and limited data.
• Evaluation of ToolAlpaca relies on GPT-4 judging; API-Bank metrics may not capture alternative valid tool paths.
• Hyperparameters (top 20%, bottom 10%, delta) chosen without thorough sweeps; general-ability evaluation covers only four tasks.

Experimental setup

Datasets: API-Bank ToolAlpaca CodeAlpaca-20K MetaMathQA TriviaQA GPT-4-LLM instruction following GSM8K HumanEval MT-Bench

Benchmarks: API-Bank ToolAlpaca GSM8K HumanEval TriviaQA MT-Bench

Metrics: Correctness (C) ROUGE-L (R) Procedure (P) Response (R) Overall (O) accuracy pass@1

Baselines: full parameter fine-tuning (FT) LoRA FT + data mixing LoRA + data mixing

Builds on: LoRAMoE LoRA gradient-based importance score pruning (Bansal et al. 2023) API-Bank ToolAlpaca

Closest papers by meaning

• Learning Evolving Tools for Large Language Models (ICLR 2025 · similarity 0.88)
• ELICIT: LLM Augmentation Via External In-context Capability (ICLR 2025 · similarity 0.87)
• Enhancing Large Language Model Performance with Gradient-Based Parameter Selection (AAAI 2025 · similarity 0.87)
• DELIFT: Data Efficient Language model Instruction Fine-Tuning (ICLR 2025 · similarity 0.87)
• When Scaling Meets LLM Finetuning: The Effect of Data, Model and Finetuning Method (ICLR 2024 · similarity 0.87)
• Toward Adaptive Large Language Models Structured Pruning via Hybrid-grained Weight Importance Assessment (AAAI 2025 · similarity 0.87)