case-based reasoning

Large language models (LLMs) often struggle with tool usage due to difficulties in balancing reasoning depth and structural validity, leading to redundant or structurally flawed outputs. To address this, this work proposes CAST, a novel framework that introduces case-driven learning into LLM tool invocation for the first time. CAST constructs structured cases from historical execution trajectories, extracting task complexity and failure patterns, and integrates fine-grained rewards with reinforcement learning to adaptively calibrate reasoning strategies and proactively avoid structural errors. Experimental results demonstrate that CAST improves execution accuracy by up to 5.85 percentage points on BFCLv2 and ToolBench benchmarks while reducing average reasoning length by 26%, substantially mitigating high-impact structural failures.

This work proposes a precedent-guided reasoning framework to address the inefficiency and performance degradation in large reasoning models caused by verbose and repetitive chain-of-thought processes. Inspired by human problem-solving through precedents, the method adaptively selects semantically relevant and low-perplexity precedents during inference and dynamically internalizes their solution patterns via lightweight adapters. This shifts reasoning from exhaustive self-exploration to efficient guided generation. Notably, it is the first approach to enable dynamic construction and utilization of a precedent set at test time. Experiments across mathematical reasoning, scientific question answering, and code generation demonstrate that the framework significantly shortens reasoning trajectories while maintaining or even improving accuracy, achieving an excellent trade-off between precision and efficiency.

To address repetitive outputs and difficulties in adapting to dynamic code structures when large language models (LLMs) generate functional test scripts, this paper proposes Re4, an enhanced case-based reasoning (CBR) framework built upon a 4R cycle—Retrieve, Reuse, Revise, and Retain—to construct a historical repository of test intents and corresponding scripts. Key contributions include: (1) semantic- and script-structure-aware re-ranking with fine-tuned retrieval; (2) reuse strategy optimization via joint supervised fine-tuning (SFT) and Proximal Policy Optimization (PPO)-based reinforcement learning; and (3) pseudo-labeling to enhance generalization under low-resource conditions. Evaluated on two Huawei Datacom product units, the CBR+Re4 approach achieves significant improvements in script accuracy (+28.6%) and executability (+34.1%), while effectively mitigating redundant generation. The framework demonstrates robust practical efficacy in industrial deployment.

Automated reasoning systems often struggle to adapt solving strategies to unique, challenging problem instances due to reliance on static or benchmark-dependent configuration. Method: We propose the first purely online, single-instance-driven adaptive tuning paradigm. It dynamically decomposes problems into subtasks via a divide-and-conquer structure and employs online reinforcement learning to optimize strategy selection in real time—without any historical benchmarks or offline pretraining. Feedback signals are derived solely from the current instance, and the framework integrates a SAT solver with a neural network verification module to enable closed-loop optimization. Contribution/Results: Our approach enables fully online policy evolution, discovering non-standard solving paths. Experiments demonstrate significant improvements in solving efficiency and success rates on both SAT solving and neural network formal verification tasks, validating the effectiveness and cross-task generalizability of single-instance adaptive tuning.

This work addresses tasks whose outputs decompose into multiple interdependent components. We propose a Monte Carlo Tree Search (MCTS)-based reasoning and search framework that explicitly models “revision actions” as first-class operations in the search space—departing from conventional left-to-right autoregressive generation and enabling dynamic refinement of intermediate outputs. The framework employs a large language model as a learned heuristic evaluator and defines a customizable action space comprising both generation and revision operators. Evaluated on story outline optimization, mini crossword solving, and constrained text generation, our method achieves new state-of-the-art performance: +30% improvement in engagement score, +16% word-level accuracy, and +10% concept coverage. These gains demonstrate substantial advances in both reasoning fidelity and controllability for complex, structured generation tasks.

This work addresses the challenge that large reasoning models often generate excessively lengthy reasoning chains due to “overthinking,” making it difficult to compress output while preserving logical coherence. To tackle this, the authors propose Stepwise Adaptive Thinking (SAT), a novel framework that enables step-level adaptive control of reasoning for the first time. SAT models the reasoning process as a finite state machine and employs a lightweight process reward model to dynamically select among inference modes—slow, normal, fast, or skip—based on local difficulty, thereby achieving difficulty-aware, progressive pruning. Extensive experiments across nine large language models and seven benchmarks demonstrate that SAT reduces reasoning token usage by 40% on average while consistently maintaining or even improving accuracy.

Current test-time scaling approaches for solving challenging mathematical problems suffer from high computational costs and diminishing returns. This work formulates test-time scaling as an instance-level policy routing problem guided by output divergence and introduces a training-free, divergence-aware mechanism that dynamically selects among lightweight reasoning, majority voting, or rewrite-and-reconstruct strategies based on the difficulty of each input instance. Evaluated across seven mathematical benchmarks and three large language models, the proposed method achieves accuracy gains of 3%–7% while substantially reducing sampling overhead, effectively balancing performance and computational efficiency.

Traditional reasoning methods exhibit limited performance in complex tasks requiring backtracking and correction, particularly lacking effective recovery mechanisms in delayed-failure scenarios. This work proposes Pyligent, a framework that models reasoning as a verification-guided search process. By employing a task-specific verifier to label the success or failure of reasoning paths, Pyligent converts the search tree into three types of supervision signals—“continue,” “complete,” and “backtrack”—to train the model. Notably, it introduces explicit supervision on failed branches for the first time, enabling the model to actively recognize errors and initiate backtracking, thereby transcending the conventional paradigm of merely imitating successful trajectories. Experiments demonstrate that Pyligent improves solution rates by 72.7 percentage points on Hidden Graph tasks and by 13–27 percentage points on Sudoku and Blocksworld, significantly enhancing the model’s error-correction and recovery capabilities.

This work addresses the limitations of existing legal judgment prediction methods, which typically rely on static, single-step reasoning and lack both verifiable inference processes and adaptability to judicial evolution. To overcome these challenges, we propose a self-optimizing multi-agent collaborative framework that emulates the role division of a judicial panel. Our approach generates verifiable reasoning chains through a traceable draft–verify–revise workflow and introduces a Hybrid Judicial Memory (HJM) mechanism coupled with a micro-instruction paradigm. This paradigm distills multi-agent interaction trajectories into updatable and transferable micro-instructions, enabling evolutionary learning across cases. Evaluated on the CAIL2018 benchmark, our method achieves state-of-the-art performance and demonstrates strong generalization capabilities on the rigorously time-partitioned CJO2025 dataset.

This work proposes a novel medical reasoning framework that addresses the limitations of current language models in complex clinical decision-making, where failures often stem from an inability to effectively access procedural knowledge and past cases. The approach formulates clinical question answering as an agent-based task and introduces an explicit yet unified retrieval mechanism that jointly accesses a structured skill repository—comprising guidelines and protocols—and an experience bank of verified reasoning trajectories. By integrating step-aware retrieval with a lightweight test-time adaptation module, the framework aligns the model’s intermediate reasoning steps with established clinical logic. Evaluated across multiple medical question-answering benchmarks, the method significantly outperforms strong retrieval-augmented generation (RAG) and prompt-only baselines, yielding improvements in accuracy, reliability, and traceability of clinical reasoning.