Agentic RAG agents must generate effective queries across diverse retriever backends, yet current agents produce retriever-oblivious queries, leading to large performance gaps when the backend changes. We observe that different retrievers, despite dissimilar architectures, exhibit structured behavioral patterns in their document-return overlap, forming a continuous behavioral space that can guide retriever-adaptive query generation. We propose RAMP (Retriever-Adaptive Multi-retriever Policy), which probes a black-box retriever with shared queries, encodes the resulting overlap patterns into a compact behavioral embedding, and injects it as soft tokens to condition query generation. A single RAMP model replaces four retriever-specific specialists without per-retriever retraining and generalizes to unseen retrievers by interpolating in behavioral space. On four QA benchmarks, RAMP matches 96–98% of four retriever-specific specialists (+3.8 EM over unconditioned training) and reaches 91.8% of the retrained upper bound on unseen retrievers, outperforming routing, few-shot, and fusion alternatives.

Although LLMs perform well on static medical QA, adapting them to multi-turn diagnostic dialogue remains challenging: agents forget confirmed symptoms, ask redundant questions, and fail to narrow differentials, largely because they lack an explicit representation of their evolving clinical understanding. Our key insight is that a structured belief state—encoding accumulated evidence, ranked differential diagnoses with confidence, and prioritized next questions—can simultaneously serve as a supervision target, a basis for rule-based reward computation, and a grounding context for deciding what to ask next. We propose BridgeGap, a two-stage framework in which supervised fine-tuning on silver-annotated belief states teaches the agent to maintain structured clinical reasoning, while reinforcement learning with a belief-progress reward refines questioning strategy through dense, environment-verified per-turn feedback without LLM judges or learned reward models. On HealthBench, BridgeGap (8B) surpasses Doctor-R1 (8B) by 2.84 points and the 4× larger Baichuan-M2-32B by 5.85 points; on AgentClinic, it achieves 44.6% diagnostic accuracy in 10.5 average turns, compared to 37.6% in 15.5 turns for Doctor-R1. Across evaluations, BridgeGap advances interactive diagnosis while preserving factual knowledge on static medical benchmarks.

Large Language Models continue to scale in size and capability, driving substantial computational and memory demands.Mixture-of-Experts (MoE) architectures alleviate this cost by activating only a sparse subset of experts per token, enabling efficient scaling without proportional increases in inference compute.However, quantization in MoE models remains challenging due to heterogeneous sensitivity across experts and their internal linear layers.Existing mixed-precision frameworks such as Mixed-precision Quantization for MoE (MxMoE) require full quantization-loss evaluation for expert–layer–and-bit configurations, incurring prohibitive profiling cost.To address this, we propose FRI-MxMoE, a profiling-free mixed-precision quantization framework built on Fuzzy Rule Interpolation, designed as a drop-in replacement for the loss estimation component in MxMoE. By constructing a fuzzy rule base in the intra-expert layer feature space (bit-width, activation variance, parameter scale), our method predicts quantization error from only sparse samples, eliminating the need for dense profiling.Extensive experiments demonstrate that FRI-MxMoE accelerates the profiling phase by up to 15.7× (on DeepSeek-V2) while achieving comparable or slightly superior zero-shot accuracy (e.g., +1.04% on DeepSeekV2-Lite) compared to the baseline.This enables continuous sensitivity modeling, preserves accuracy under mixed-precision allocation, and reduces offline computation by orders of magnitude.

The Mixture of Experts (MoE) architecture enables efficient model scaling through conditional computation, where only subset of parameters are activated per input. However, this distributed architecture poses unprecedented challenges for model compression, as conventional quantization methods optimized for dense networks prove inadequate. This paper introduces a specialized quantization framework for MoE architectures, motivated by our discovery that weight matrices across expert networks exhibit distinctive channel-wise outlier distributions, necessitating a more nuanced compression approach. Through theoretical analysis incorporating Fisher Information matrices and condition number characteristics, we establish a fundamental relationship between layer functionality and quantization sensitivity, demonstrating that down-projection layers inherently demand higher precision compared to up-projection layers. Leveraging these insights, we develop an automated channel-wise quantization framework that dynamically determines optimal bit-width allocations while maintaining minimal computational overhead through efficient statistical approximations. When evaluated on the Mixtral-8x7b-v0.1 architecture, our methodology demonstrates a 3.96% improvement over existing state-of-the-art approaches across natural language understanding benchmarks, while achieving superior compression ratios.