Large Language Models (LLMS) revolutionize AI by showing significant capabilities in text generation and problem solving. However, the key limitations still exist in their default values “Think quickly” Method – Output based on a single query without iterative refinement. recent “Think slowly” Methods such as thinking chains drive rest problems into smaller steps, which are still subject to static initial knowledge and cannot dynamically integrate new information during the inference process. This gap becomes apparent in complex tasks that require real-time knowledge updates, such as multi-hop question and answer or adaptive code generation.
The current methods of enhancing LLM reasoning are divided into two categories. Search Authorized Generation (RAG) The system preloads external knowledge, but often introduces irrelevant information, which hinders efficiency and accuracy. Tree-based search algorithm Monte Carlo Tree Search (MCTS) Achieve structured exploration of inference paths, but lacks a mechanism for contextual knowledge integration. For example, although LATS (LLM-driven MCT) introduces the evaluation and reflection phase, it still runs within the model’s initial knowledge boundary. These methods encounter difficulties in balancing the breadth of exploration, contextual relevance, and computational efficiency, often resulting in overly broad or insufficient responses.
In this article, a team of researchers from Digital Security Group, Qihoo 360, proposed Related chains (coats) The framework to address these limitations through two key innovations. First, Associated memory mechanism Enable dynamic knowledge integration during the reasoning process to mimic human cognitive associations. The method with static rags allows information to be retrieved in advance, and the jacket activates knowledge retrieval based on specific reasoning steps, the same as the mathematician, recalls the relevant theorems only if needed in the proof. second, Optimized MCT algorithm This association process is combined through a novel four-stage cycle: selection, knowledge association, quality assessment and value backpropagation. This creates a feedback loop where each reasoning step can trigger the target knowledge update, as shown in Figure 4 of the original implementation.
The core of the jacket is the dual-flow reasoning structure. When processing queries, the system simultaneously explores possible inference paths through the MCTS tree while maintaining the associated memory bank. Search for each node in the tree (represents the inference step) to generate two contents (g(n))related knowledge (am(n)) and
Assign score balance answer quality (fg) Relevance to knowledge (fone)and β The relative importance of controlling them. This ensures that associations are closely related to evolving reasoning processes rather than introducing tangential information.
The performance evaluation of the jacket highlights its advantages over existing reasoning enhancement techniques. The framework is benchmarked based on qualitative and quantitative indicators of various tasks. Qualitative evaluation involves complex query responses, and the jacket shows richer and more comprehensive answers compared to baseline models such as QWEN2.5-32B and CHATGPT. It is worth noting that it introduces other categories of reasoning, such as ethical and regulatory considerations, while other models do not. Quantitative evaluations are conducted in two main areas: knowledge-intensive Q&A and code generation. For the retrieval-based generation (RAG) task, coats were compared with Nativerag, Ircot, Hipporag, Lats, and KAG of the HotPotQA and 2 Wikimultihopqa datasets. Indicators such as exact match (EM) and F1 scores confirm the excellent performance of the jacket, demonstrating its ability to produce exact and context-sensitive answers. In code generation, the coating enhancement model is better than the data sets on data sets such as humans, MBPP and HUMANEVAL-X (QWEN2.5-CODER-7B-7B-INSTRUCTION, QWEN2.5-CODER-14B-14B- INSTRUCT) performance can achieve its adaptability to the domain-specific reasoning tasks.
This work establishes a new paradigm for LLM reasoning by integrating dynamic knowledge associations with structured searches. Unlike previous static enhancement methods, Coat’s real-time memory updates can adapt context-aware reasoning to emerging information needs. Technological innovations in MCT optimization and dual evaluation provide a blueprint for combining external knowledge systems with modern LLM. Although the current implementation relies on predefined external brains, the architecture naturally supports plug-in integration with emerging tools such as LLM proxy and real-time web search. These advances suggest that the next frontier of AI inference may be a system that dynamically interweaves internal computing with targeted external knowledge retrieval, just as human experts consult with references in complex problem-solving processes.
Check Paper. All credit for this study is attributed to the researchers on the project. Also, don’t forget to follow us twitter And join us Telegram Channel and LinkedIn GrOUP. Don’t forget to join us 75K+ ml reddit.
Recommended open source AI platform: “Intellagent is an open source multi-agent framework that evaluates complex dialogue AI systems” (promoted)
The AI framework with Marktechpost first emerged to enhance LLM reasoning: an AI framework.
