User simulators are essential for the scalable training and evaluation of interactive AI systems. However, existing approaches often rely on shallow user profiling, struggle to maintain persona consistency over long interactions, and are largely limited to English or single-domain settings. We present MUSE, a multi-domain Chinese user simulation framework designed to generate human-like, controllable, and behaviorally consistent responses. First, we propose Iterative Profile Self-Evolution (IPSE), which gradually optimizes user profiles by comparing and reasoning discrepancies between simulated trajectories and real dialogue behaviors. We then apply Role-Reversal Supervised Fine-Tuning to improve local response realism and human-like expression. To enable fine-grained behavioral alignment, we further train a specialized rubric-based reward model and incorporate it into rubric-guided multi-turn reinforcement learning, which optimizes the simulator at the dialogue level and enhances long-horizon behavioral consistency. Experiments show that MUSE consistently outperforms strong baselines in both utterance-level and session-level evaluations, generating responses that are more realistic, coherent, and persona-consistent over extended interactions.

Group Relative Policy Optimization (GRPO) significantly enhances the reasoning performance of Large Language Models (LLMs). However, this success heavily relies on expensive external verifiers or human rules. Such dependency not only leads to significant computational costs and training latency, but also yields sparse rewards that hinder optimization efficiency. To address these challenges, we propose Latent-GRPO, a framework that derives intrinsic rewards directly from latent space geometry. Crucially, our empirical analysis reveals a compelling geometric property: terminal token representations of correct reasoning trajectories form dense clusters with high intra-class similarity, whereas incorrect trajectories remain scattered as outliers. In light of this discovery, we introduce the Iterative Robust Centroid Estimation (IRCE) algorithm, which generates dense, continuous rewards by mitigating magnitude fluctuations via spherical projection and estimating a robust “truth centroid” through iterative aggregation. Experimental results on multiple datasets show that our method maintains model performance while achieving a training speedup of over 2x compared to baselines. Furthermore, extensive results demonstrate strong generalization ability and robustness.

Real-time voice agents face a tension: end-to-end models often lack deep reasoning, while cascaded ASR–LLM–TTS pipelines achieve stronger reasoning but defer LLM inference until after ASR finalizes the transcript, creating an avoidable post–turn-end silent gap. We target this gap under complex spoken requests that admit a relatively determinate answer, and treat it as system-induced rather than communicatively motivated. Existing cascaded streaming strategies attempt to reduce it via mechanical segmentation (e.g., fixed chunks, VAD-based splitting) or speculative generation, but they frequently either break semantic units or waste computation on predictions that must be rolled back. We propose LTS-VoiceAgent, a Listen–Think–Speak framework that explicitly separates when to think from how to reason incrementally. It features a Dynamic Semantic Trigger to detect meaningful prefixes, and a Dual-Role Stream Orchestrator that coordinates a background Thinker (for state maintenance) and a foreground Speaker (for speculative solving). This parallel design enables "thinking while listening" without committing audible content before turn end. We also introduce a Pause-and-Repair benchmark that injects natural disfluencies as a semantic-level streaming stress test. Experiments across VERA, Spoken-MQA, BigBenchAudio, and our benchmark show that LTS-VoiceAgent achieves a stronger accuracy–latency–efficiency trade-off than serial cascaded baselines and existing streaming strategies.

Real-world deployments of large language models (LLMs) increasingly rely on user-provided instructions that are long, precise, and tailored to specific workflows. In such settings, aggregate leaderboard performance is often a weak proxy for a model’s ability to faithfully execute the particular instruction that a practitioner intends to deploy. To address this gap, we introduce FlexBench/FlexEval, an instruction-conditioned evaluation framework that instantiates a specialized test suite from a user-provided target instruction. Given a target instruction, FlexBench (i) derives a set of verifiable evaluation dimensions and (ii) generates an instruction-conditioned conversation corpus. FlexEval then aggregates dimension-level judgments into two complementary metrics: Coverage, which measures workflow progress, and Achievement, which measures conditional correctness. We evaluate the framework on nine multi-turn target instructions spanning diverse scenario categories and workflow topologies, across nine contemporary LLMs over 3,888 generated dialogues. The results show that model rankings and failure modes vary substantially across target instructions, highlighting the need for instruction-conditioned evaluation beyond fixed aggregate benchmarks. Our source code is publicly available at https://github.com/Coryyy/FlexBench.

Effective memory management is essential for large language model agents to navigate long-horizon tasks. Recent research has explored using Reinforcement Learning to develop specialized memory manager agents. However, existing approaches rely on final task performance as the primary reward, which results in severe reward sparsity and ineffective credit assignment, providing insufficient guidance for individual memory operations. To this end, we propose Fine-Mem, a unified framework designed for fine-grained feedback alignment. First, we introduce a Chunk-level Step Reward to provide immediate step-level supervision via auxiliary chunk-specific question answering tasks. Second, we devise Evidence-Anchored Reward Attribution to redistribute global rewards by anchoring credit to key memory operations, based on the specific memory items utilized as evidence in reasoning. Together, these components enable stable policy optimization and align local memory operations with the long-term utility of memory. Experiments on Memalpha and MemoryAgentBench demonstrate that Fine-Mem consistently outperforms strong baselines, achieving superior success rates across various sub-tasks. Further analysis reveals its adaptability and strong generalization capabilities across diverse model configurations and backbones

High-quality chain-of-thought has demonstrated strong potential for unlocking the reasoning capabilities of large language models. However, current paradigms typically treat the reasoning process as an indivisible sequence, lacking an intrinsic mechanism to quantify step-wise information gain. This granularity gap manifests in two limitations: inference inefficiency from redundant exploration without explicit guidance, and optimization difficulty due to sparse outcome supervision or costly external verifiers. In this work, we propose CoT-Flow, a framework that reconceptualizes discrete reasoning steps as a continuous probabilistic flow, quantifying the contribution of each step toward the ground-truth answer. Built on this formulation, CoT-Flow enables two complementary methodologies: flow-guided decoding, which employs a greedy flow-based decoding strategy to extract information-efficient reasoning paths, and flow-based reinforcement learning, which constructs a verifier-free dense reward function. Experiments on challenging benchmarks demonstrate that CoT-Flow achieves a superior balance between inference efficiency and reasoning performance.

Reinforcement Learning with Verifiable Rewards (RLVR) is a key paradigm for improving large-scale reasoning models. Unlike supervised fine-tuning (SFT), RLVR exhibits distinct optimization dynamics and are sensitive to the preservation of pre-trained geometric structures. However, existing parameter-efficient methods face key limitations in this regime. Low-rank adaptation methods, such as PiSSA, are primarily designed for Supervised Fine-Tuning (SFT) and do not account for the distinct optimization dynamics and geometric structures of RLVR. Conversely, directly fine-tuning the unstructured sparse parameter subspace favored by RLVR encounters efficiency bottlenecks on modern hardware. To address these challenges, we propose GeoRA (Geometry-Aware Low-Rank Adaptation), a low-rank adaptation method tailored for RLVR. Specifically, GeoRA exploits the anisotropic and compressible structure of RL update subspace, and extracts its principal directions via Singular Value Decomposition (SVD) to initialize low-rank adapters, while freezing residual components as a structural anchor during training. This design preserves the pre-trained structure and enables efficient dense computation. Experiments on Qwen and Llama models from 1.5B to 32B parameters show that GeoRA consistently outperforms strong low-rank baselines across RLVR settings in mathematics, medicine, and coding, while showing stronger generalization and less forgetting on out-of-domain tasks.