Post-training reasoning models is computationally expensive, and the allocation between supervised fine-tuning (SFT) and reinforcement learning (RL) substantially affects reasoning quality. Existing methods operate at the training-procedure level and treat the reasoning corpus as homogeneous, ignoring the structural heterogeneity of individual traces. We propose GROVE (Graph-based Routing Over Variant Experts), an adaptive framework that routes each example to SFT or RL based on the structural topology of its reasoning trace. GROVE parses each trace into a Reasoning Tree, scores its complexity via graph entropy, and partitions the corpus through a data-driven Gaussian Mixture Model with no hand-tuned threshold. Low-complexity traces are trained with SFT and high-complexity traces with GRPO as parallel LoRA adapters from a shared frozen base, merged into a single deployable model with no inference-time overhead. Across two training corpora and five reasoning benchmarks, GROVE consistently outperforms single-stage and naive-hybrid baselines, and a reversed-routing ablation confirms that this mapping is intrinsic rather than arbitrary.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable reasoning capabilities yet continue to suffer from hallucination, where generated text contradicts visual content. In this paper, we introduce Dual-Anchor Introspective Decoding (DaID), a novel contrastive decoding framework that dynamically calibrates each token generation by mining the model’s internal perceptual discrepancies. Specifically, DaID identifies a Spotlight layer to amplify visual factual signals and a Shadow layer to suppress textual inertia. By leveraging visual attention distributions to guide this dual-anchor selection process, our method ensures precise, token-specific adaptation. Experimental results across multiple benchmarks and MLLMs demonstrate that DaID significantly mitigates hallucination while enhancing general reasoning capabilities.

Multimodal Large Language Models (MLLMs) demonstrate remarkable reasoning capabilities, yet they remain susceptible to hallucinations. While Multi-Agent Debate (MAD) presents a promising collaborative solution, unguided debates inevitably devolve into a stochastic martingale process and expanding conversational contexts trigger severe visual attention decay. In this paper, we propose EviaMAD, an Evidence-Anchored Multi-Agent Debate framework. To proactively expose internal hallucinations, EviaMAD bifurcates perception into two distinct roles: a Global Observer capturing the holistic scene and a Local Observer scrutinizing salient details. When conflicts arise, a dedicated Analyzer intervenes, distilling the linguistic disagreement into a deterministic bimodal evidence anchor: a physically cropped visual patch of the disputed region paired with key semantic directives. During the subsequent reflection phase, a surgical reallocation mechanism explicitly amplifies the attention weights of this anchor. By compelling agents to ground their reasoning in strictly verified evidence, EviaMAD effectively elicits collective intelligence. Extensive evaluations demonstrate that EviaMAD significantly mitigates hallucinations while elevating the reasoning capabilities of the multi-agent system.

Federated fine-tuning enables privacy-preserving LLM adaptation but faces a critical bottleneck: the disparity between LLMs’ high memory demands and edge devices’ limited capacity. To break the memory barrier, we propose Chain Federated Fine-tuning (ChainFed), an innovative paradigm that forgoes end-to-end updates in favor of a sequential, layer-by-layer manner. It first trains the initial adapter to convergence, freezes its weights, and then proceeds to the next. This iterative train-and-freeze process forms an optimization chain, gradually enhancing the model’s task-specific proficiency. ChainFed further integrates three core techniques: 1) Dynamic Layer Co-Tuning to bridge semantic gaps between sequentially tuned layers and facilitate information flow; 2) Globally Perceptive Optimization to endow each adapter with foresight beyond its local objective; 3) Function-Oriented Adaptive Tuning to automatically identify the optimal fine-tuning starting point. Extensive experiments on multiple benchmarks demonstrate the superiority of ChainFed over existing methods, boosting average accuracy by up to 46.46%.