Efficient reasoning is critical for large reasoning models (LRMs) deployed in agent systems with strict context constraints. Existing methods predominantly rely on global length penalties that overlook step-level structure, inadvertently compressing critical derivations while retaining low-value detours. Hence, we argue for a principled, step-aware alternative that preserves critical steps and compresses low-value detours. However, it raises three fundamental challenges: the difficulty of obtaining step-level supervision, the need for reward signals that induce differentiated length pressure, and the requirement to balance training stability with exploration on complex problems. We propose SmartThinker, a step-level token allocation framework that learns to selectively preserve or compress individual reasoning steps. At its core is Step-Level Length Control Policy Optimization (SCPO), which provides label-free feedback by online estimation of each step’s contribution from the change in correct-answer likelihood after removing it. We further introduce reasoning density, defined as contribution per token, to relax length pressure on high-value steps while intensifying it on low-density ones. Finally, SCPO stabilizes training through length-bias-corrected advantages and difficulty-adaptive clipping, achieving superior accuracy-efficiency trade-offs across nine in-domain and out-of-domain benchmarks.

Reasoning efficiency in Text-to-SQL is an important yet underexplored problem. Chain-of-Thought (CoT) reasoning models simplify traditional Text-to-SQL workflows and significantly improve the ability to handle complex queries. However, existing methods typically generate long reasoning processes for all queries. This leads to high token consumption and inference latency, increasing system cost and limiting practical deployment. To address this issue, we propose Ada-SQL, an adaptive reasoning model for Text-to-SQL that jointly optimizes reasoning mode selection and token budget control via a two-stage reinforcement learning framework coupled with a token budget model. This enables the model to adaptively decide whether to reason and how much reasoning budget to allocate based on query difficulty. Experimental results show that Ada-SQL reduces output token consumption by more than 65% while maintaining competitive execution accuracy, achieving a favorable performance-efficiency trade-off. Code and models will be released at https://github.com/sxhText2SQL/Ada-SQL.

Retrieval-Augmented Generation (RAG) systems have become a standard approach for grounding large language models in external knowledge. However, they are constrained by a decoupled architecture: retrieval and reasoning operate as separate stages, with retrieved text merely prepended as passive context. This prevents deep integration of knowledge into the model’s parametric reasoning, leading to fragmented responses for complex queries requiring multi-document synthesis or conflict resolution. To bridge this gap, we propose NeuRAG, an end-to-end Neuralized RAG framework that unifies knowledge retrieval and fusion through Hyper-Neurons—parameterized modules encoding entire documents directly into the model’s parameter space. In NeuRAG, each document is encoded as a lightweight LoRA module, conceptualized as a knowledge neuron. These neurons collectively form a document-adaptive Hyper-Layer, which dynamically activates and fuses knowledge neurons via an attention mechanism conditioned on the input hidden-state query. This enables the model to jointly retrieve and reason within a single forward pass, seamlessly integrating external knowledge into its inference pathway. Extensive experiments across multiple datasets and LLMs demonstrate NeuRAG’s strong and consistent performance as a promising novel RAG paradigm.

Recent studies have increasingly explored the combination of existing LoRA modules for effective adaptation to unseen tasks in data-scarce scenarios. However, current LoRA selection methods typically rely on a few task samples, making it difficult to capture the full scope of task-relevant information. Furthermore, even after selection, a knowledge gap remains between the selected LoRA modules and the target task, which existing coarse-grained LoRA aggregation strategies struggle to bridge. To address these challenges, we propose Selection and Convolution for LoRA aggregation (SC-LoRA), a two-stage framework that first selects appropriate LoRA modules based on parameter clustering and then aggregates them using a convolutional LoRA aggregator. Our LoRA selection strategy ensures comprehensive coverage of task-relevant LoRA modules by leveraging their distance in the parameter space. Building on this, the convolutional LoRA aggregator extracts useful knowledge in a fine-grained manner, seamlessly bridging the gap to the target task. Our experiments demonstrate that SC-LoRA excels in aggregating multiple LoRA modules for effective adaptation to unseen tasks.