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

RCP-Merging: Merging Long Chain-of-Thought Models with Domain-Specific Models by Considering Reasoning Capability as Prior

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Chain-of-thought training and analysis · 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.

Proposes RCP-Merging, which merges a long-CoT reasoning model with domain-specific LLMs by treating reasoning weights as a Bayesian prior (diagonal Fisher penalty) combined with domain sensitivity-based masking, evaluated on BioMedicine and Finance domains.

Problem

Merging domain-specific LLMs with long chain-of-thought reasoning models causes reasoning degradation, gibberish output, and output collapse with existing merging methods.

Approach

Treats the reasoning model's weights as a prior, using a diagonal Fisher Information-based reasoning preservation indicator to penalize domain task-vector updates that deviate from reasoning-critical weights, combined with a first-order domain knowledge sensitivity score and a sign-majority mask before adding the reasoning task vector back.

Key results

• On the BioMedicine merge (Qwen2.5-7B), RCP-Merging reaches 49.4 average vs. 39.9 for Sens-Merging (the best baseline), with 55.5 on PubMedQA and 54.1 on MedQA.
• On Llama3.1-8B, RCP-Merging achieves 68.3 average vs. 59.1 for CAT-Merging, with 73.2 on HumanEval and 75.8 on ARC-C.
• On the 1.5B setting, RCP-Merging achieves 47.8 average vs. 37.2 for CAT-Merging; the paper reports a 14.3% average gibberish rate for RCP-Merging versus 82.3% for TIES.

Surprising findings

• The merged model outperforms the original domain-specific model on PubMedQA (+4.5%) and MedQA (+0.7%), despite merging aimed at retaining rather than improving domain performance.
• On LiveCodeBench, RCP-Merging (18.4) is below Sens-Merging (21.7) and CAT-Merging (20.1) in the 7B BioMedicine setting.

Limitations

• Gibberish and output-collapse analysis relies on a GPT-4 evaluator rather than human evaluation.
• Hyperparameter lambda must be tuned per model size (0.3 for 7B, 0.7 for 1.5B), and the reasoning coefficient is set heuristically.
• Evaluation covers only two domains (BioMedicine and Finance) and three base model families/sizes.

Experimental setup

Datasets: GSM8K AIME2024 HumanEval LiveCodeBench PubMedQA MedQA ARC-C GPQA ConvFinQA

Benchmarks: GSM8K AIME2024 HumanEval LiveCodeBench PubMedQA MedQA ARC-C GPQA ConvFinQA

Metrics: accuracy pass@1 gibberish rate Distinct-N perplexity generated token length

Baselines: Linear averaging Task Arithmetic TIES-Merging DARE Linear DARE TIES FuseLLM DARE TIES & AIM FuseLLM & AIM Sens-Merging CAT-Merging

Builds on: Task Arithmetic TIES-Merging DARE Laplace approximation Sens-Merging CAT-Merging

Closest papers by meaning

• Unlocking the Capabilities of Thought: A Reasoning Boundary Framework to Quantify and Optimize Chain-of-Thought (NeurIPS 2024 · similarity 0.88)
• MergeBench: A Benchmark for Merging Domain-Specialized LLMs (NeurIPS 2025 · similarity 0.88)
• Long Chain-of-Thought Reasoning Across Languages (ICLR 2026 · similarity 0.88)
• LongCoT: Benchmarking Long-Horizon Chain-of-Thought Reasoning (ICML 2026 · similarity 0.88)
• The CoT Encyclopedia: Analyzing, Predicting, and Controlling how a Reasoning Model will Think (ICLR 2026 · similarity 0.87)
• CoT Vectors: Transferring and Probing the Reasoning Mechanisms of LLMs (ICLR 2026 · similarity 0.87)