Aug 2026· Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2· 0 citations· 15 references
TL;DR
This work proposes Step-wise Training for In-context Reasoning (STIR), a model to dynamically decide when to retrieve a single logically consistent next step, just using the current problem and its intermediate state as the query.
Abstract
Large language models (LLMs) have shown strong performance in mathematical reasoning, supported by approaches such as In-Context Learning (ICL) and Retrieval-Augmented Generation (RAG). However, existing methods often provide problem-level examples, which is too coarse-grained for multi-step reasoning to cause informational redundancy, and structural misalignment. To address this limitation, we propose Step-wise Training for In-context Reasoning (STIR) to provide step-synchronized and logically targeted guidance to enhance the model's mathematical reasoning capabilities. STIR enables a model to dynamically decide when to retrieve a single logically consistent next step, just using the current problem and its intermediate state as the query. First, We decompose expert solutions into Step-Level Reasoning Units inspired by human thinking patterns. Leveraging this data, a Step Retriever is trained for logical continuity to map current reasoning states to relevant subsequent steps. Then a Step Reasoner is trained to decide when to retrieve tailored step examples and incorporates this guidance into reasoning. We further extend STIR with a Process-aware Reinforcement Learning phase using Group Relative Policy Optimization to learn to self-formulate search queries and optimizes the decision-making policy. Experiments on seven benchmarks demonstrate that STIR achieves accuracy improvements ranging from 1.86% to 17.96%, maintaining lower token efficiency than baselines. Analysis via our proposed DSM, TCN and RCR metrics shows that STIR improves reasoning capability, achieving DSM scores ranging from 4.54 to 29.87 across backbones and significant improvements over the baseline in both TCN and RCR.
Large language models (LLMs) excel at text understanding and generation, yet still struggle to reliably understand and apply externally provided procedural rules at scale. To evaluate this capability, we introduce RuleWorld, a large-scale benchmark that reformulates rules as globally reusable abstract units rather than instance-specific facts. In RuleWorld, several scenarios, including single-rule, parallel multi-rule, and multi-hop reasoning, are settled for comprehensive evaluation. We further propose DynaRule, an end-to-end framework that injects the given rules into the KV cache and turns retrieval into an internal, learnable, step-wise process. Specifically, DynaRule employs Stacked Step-Level Attention Training with a specialtoken to enable dynamic rule re-attention and updating during inference. In this way, the model can re-attend to the most relevant rules at each step, dynamically replacing outdated ones to support more stable multi-step reasoning. Experiments on RuleWorld show that existing LLMs face challenges under large rule pools, while DynaRule improves average QA accuracy by up to 19 points and achieves over 85% Recall@1 at 10K rules, outperforming strong baselines by large margins. We make our code and dataset available here: https://github.com/SharkSpicy-NLP/Beyond-Factual-Knowledge.
Bohan Yu, Pengfei Cao, Chen Han et al.· 0 citations
Experimental results show that PD-CoTD significantly enhances the mathematical reasoning abilities of SLMs and provides an effective solution for deploying powerful reasoning in resource-constrained environments.
Zhe Wang, Long Zhang, Fengbo Zheng et al.· International journal of sof...· 0 citations
Large language models (LLMs) often generate inaccurate answers due to their reliance on static internal knowledge. Retrieval-augmented generation (RAG) addresses this limitation by integrating external knowledge and excelling at single-hop queries. However, it struggles with multi-hop questions that require cross-document reasoning. Existing methods, such as graph structured RAG or question decomposition, often lack dynamic decomposition and effective filtering, which leads to lower efficiency and accuracy. To overcome these limitations, we propose Dynamic Decomposition and Filtering for Multi-Hop Reasoning-Augmented Generation (D2F-ReAG), a novel paradigm that adaptively controls reasoning depth by judging the reliability of the root-level reasoning. If the root reasoning is reliable, the model directly generates the answer. Otherwise, the question is logically decomposed into sub-questions, and the verified reasoning derived from these sub-questions is used to refine the root reasoning. Experiments on three multi-hop benchmarks demonstrate the effectiveness of our method in handling complex multi-hop questions.
Jiaoyang Li, Junhao Ruan, Shengwei Tang et al.· 0 citations
Large language model agents increasingly rely on reusable skills to extend their capabilities beyond parametric knowl- edge. However, retrieving the appropriate skill from a large- scale library remains challenging because realistic user re- quests are often concise and underspecified, stating only the task goal while leaving the required capabilities and execu- tion steps implicit. Existing benchmarks provide limited cov- erage of such requests. To address this gap, we introduce SkillReason-Bench, a large-scale cross-domain benchmark containing 3,729 queries and a retrieval corpus of 61,228 skills spanning nine domains. We further propose SkillRea- son, a two-stage framework that uses chain-of-thought rea- soning as training-time supervision for skill retrieval. In Stage I, capability reasoning traces generated by a stronger teacher provide explicit supervision through contrastive learning, re- trieval distribution alignment, and language modeling, en- couraging the retriever to internalize capability reasoning in its query representation. In Stage II, a retrieval-guided GRPO objective encourages the model to explore reasoning trajecto- ries better suited to its own capabilities and more effective for retrieval. At inference, SkillReason directly encodes the orig- inal query without autoregressive CoT generation, preserv- ing efficient query-only retrieval. Extensive experiments on SkillReason-Bench, SkillRet, and SRA-Bench show that Skill- Reason achieves state-of-the-art performance across all three benchmarks, demonstrating that reasoning-enhanced training better bridges the semantic gap between high-level task goals and skill capabilities.
Donghong Jiang, Endian Lin, Luoping Cui et al.· 0 citations
Large Language Models (LLMs) demonstrate significant potential in sequential recommendation, and leveraging their Chain-of-Thought (CoT) reasoning capabilities can further unlock profound user preference understanding. However, deploying explicit CoT reasoning in real-world systems faces prohibitive challenges: (i) the conflict between the large model scale required for high-fidelity reasoning and the resource constraints of online services, and (ii) the excessive latency introduced by auto-regressive rationale generation. To address these issues, we propose I Reasoning via Multi-Teacher Distillation (IRMD), a novel framework that 'compiles' the reasoning abilities of large teacher LLMs into a lightweight student Small Language Model (SLM). IRMD first employs a Multi-Teacher CoT Synthesis with Dual-Constraint Rejection Sampling module to generate a high-quality, diverse set of reasoning paths. Subsequently, our Annealing-Scheduled Reasoning Distillation strategy progressively trains the student to internalize this logic, transitioning from mimicking explicit CoT to performing purely implicit reasoning. Extensive experiments on multiple benchmark datasets demonstrate that IRMD significantly outperforms state-of-the-art baselines in both recommendation accuracy and inference efficiency. Our code is accessible at https://github.com/Cxx-0/IRMD.
Large language models (LLMs) have demonstrated that explicitly performing step-by-step thinking before producing final outputs can substantially improve performance on complex tasks, as exemplified by recent reasoning-oriented models such as OpenAI O1 and DeepSeek R1. Inspired by these advancements, we propose the O1 Embedder, a novel approach aiming to endow retrieval models with similar capabilities to address challenges like multi-task retrieval, zero-shot retrieval, and tasks requiring intensive reasoning of complex relationships. The O1 Embedder generates preliminary thoughts for input queries before document retrieval. To realize this objec-tive, we address two fundamental challenges in integrating thinking mechanisms into dense retrieval. First, retrieval tasks lack explicit supervision for intermediate thinking processes, making it difficult to define thoughts that are truly useful for retrieval. We address this challenge with a data synthesis framework following an Exploration-Refinement process, ensuring alignment with retrieval utility. Second, effectively integrating thought generation with representation learning requires a uni-fied modeling framework that can jointly support generation and embedding within a single model. O1 Embedder addresses this challenge by jointly
Ruiran Yan, Wen Xiong, Ze Liu et al.· Annual Meeting of the Associ...· 2 citations