Research map / Large language models / Reasoning and chain of thought

D-CORE: Incentivizing Task Decomposition in Large Reasoning Models for Complex Tool Use

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 0 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.

Identifies "Lazy Reasoning" in LRMs on multi-turn tool use and proposes D-CORE, a self-distillation (task decomposition) stage followed by diversity-aware GRPO, reaching 77.7% on BFCLv3 with 8B and 79.3% with 14B.

Problem

Large reasoning models (LRMs) such as Qwen3 underperform specialized tool-use LLMs in complex multi-turn tool use, generating verbose but unproductive reflection instead of decomposing tasks into sub-tasks.

Approach

Self-distill task-decomposition-structured reasoning trajectories from the LRM itself (with reference trajectories and few-shot guidance, no stronger teacher), then apply GRPO with an entropy-based advantage replacement when advantages collapse to zero, restoring reflective reasoning.

Key results

• D-CORE-8B reaches 77.7% overall on BFCLv3, surpassing the best 8B baseline (xLAM2-8B, 72.0%) by 5.7 points.
• D-CORE-14B reaches 79.3% on BFCLv3, outperforming xLAM2-70B (78.4%) with a much smaller model.
• D-CORE-8B improves Qwen3-8B on BFCLv3 multi-turn from 33.0% to 63.8% and tau-bench overall from 29.0% to 47.6%.

Surprising findings

• GRPO applied directly to Qwen3-8B yields marginal or negative gains on BFCLv3 multi-turn (33.0% to 26.8%), while self-distillation followed by GRPO helps substantially.
• Self-distillation alone reduces reflection and reward variance, causing GRPO advantage collapse.

Limitations

• Gains on some sub-categories (e.g., Non-Live, Irrelevance) are flat or negative; D-CORE-14B drops on Irrelevance relative to Qwen3-14B.
• Decomposition quality depends on reference trajectories and few-shot examples; pseudo-label variant trails on tau-bench (29.6 vs 37.1).
• Entropy-advantage coefficient alpha must be tuned; excessive alpha causes exploration collapse.

Experimental setup

Datasets: subsets of open-source tool-use datasets custom tool use agent trajectories BFCLv3 tau-bench

Benchmarks: BFCLv3 tau-bench tau2-Bench ACEBench-en BFCLv4-agentic

Metrics: accuracy

Baselines: Qwen3-8B Qwen3-14B Qwen3-32B xLAM2-8B xLAM2-32B xLAM2-70B ToolRL-Qwen3-8B ToolRL-Qwen3-14B DeepSeek-R1 Claude-3.7-Sonnet o1 GPT-4o

Builds on: GRPO ToolRL Decomposed Prompting Least-to-Most Prompting DeepSeek-R1

Closest papers by meaning

• TrimR: Verifier-based Training-Free Thinking Trimming for Efficient Test-Time Scaling (ICLR 2026 · similarity 0.87)
• ReTool: Reinforcement Learning for Strategic Tool Use in LLMs (ICLR 2026 · similarity 0.86)
• Nemotron-Research-Tool-N1: Exploring Tool-Using Language Models with Reinforced Reasoning (ICLR 2026 · similarity 0.86)
• Tool-Planner: Task Planning with Clusters across Multiple Tools (ICLR 2025 · similarity 0.86)
• Joint Optimization of Tool Creation and Use for Large Language Model Agents (NeurIPS 2026 · similarity 0.86)
• How Far Are We from Optimal Reasoning Efficiency? (NeurIPS 2025 · similarity 0.86)