AI Paper Atlas

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

ParaTool: Shifting Tool Representations from Context to Parameters

Zekai Yu, Qi Meng, Qizhi Chu, Yu Hao, Chuan Shi, Cheng Yang

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning generalization and delta compression · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
encode each tool into its own LoRA parameters instead of providing docs and examples in contextLlama3.1-8B-Instruct · Stable ToolBench (average)pass rate (%): 65.67 -> 69.57Improved
use per-tool parameters instead of a single global LoRA (ablation)ParaTool · BFCL-V2 averageAST accuracy (%): 49.66 -> 76.23 (Live, Llama3.1-8B)Improved
use soft tool composition instead of hard top-1 selection (ablation)ParaTool · BFCL Live Parallel Multiple (Llama3.1-8B)AST accuracy: drop of nearly 17% with Top-1Improved
use trained gating weights instead of averaging tool parameters (ablation)ParaTool · BFCL (Llama3.1-8B)AST accuracyImproved
add parametric tool fine-tuning stage (ablation)ParaTool · BFCL (Llama3.1-8B)AST accuracy: 27.4% relative improvementImproved
use oracle tool selector instead of learned gate (ablation)ParaTool · BFCL Live Parallel MultipleAST accuracy: gap of 6.0%No clear effect
use learned gating network instead of retrieval-based tool selectors (ablation)ParaTool · BFCL Parallel MultipleAST accuracy: 31.7% relative gain over best selectorImproved
add more usage examples in the contextICL tool calling · Llama3.1-8B and Qwen2.5-7Btool calling accuracyMixed
parameterized tools instead of contextinference cost · Stable ToolBenchFLOPs: average reduction of 92.22%Improved
fine-tune ToolLLaMA-style LoRA on all toolsLlama3.1-8B-Instruct · BFCL Non-live Parallel MultipleAST accuracy (%): 85.50 -> 58.50Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy