Multi-agent systems (MAS) have emerged as a promising paradigm for solving complex tasks. Recent work has explored self-evolving MAS that automatically optimize agent capabilities or communication topologies. However, existing methods either learn a topology that remains fixed at inference time or adapt only the topology or capability during inference. We empirically and theoretically show that effective test-time evolution requires jointly adapting both axes, but on different time scales: capabilities should update rapidly to handle emerging subtasks, while the topology should evolve more slowly to preserve coordination stability. We then introduce TacoMAS, a test-time co-evolution framework for dynamic MAS. TacoMAS formulates MAS inference as a task of online graph adaptation, where nodes represent agents with role-specific capabilities and edges define their communication topology. During inference, a fast capability loop updates agent expertise using trajectory-level feedback, while a slow meta-LLM-driven topology loop performs agents' birth-death operations on MAS, including edge edit, agent addition, and agent removal. We further show that this fast-slow design drives MAS evolution toward a task-conditioned stable equilibrium. Experiments on four benchmarks demonstrate that TacoMAS outperforms nearly 20 multi-agent baselines, achieving an average improvement of 13.3

Contemporary progress in Large Language Models (LLMs) has revealed notable inferential capacities via reinforcement learning (RL) employing verifiable rewards. However, “zero-RL” approaches relying on fixed prompt templates introduce substantial sampling inefficiencies for weak LLMs, as most problems generate invalid outputs during accuracy-driven filtration. To solve this, we propose Cog-Rethinker, a novel hierarchical metacognitive RL framework. Cog-Rethinker enhances the rollout procedure by improving sample utilization through a two-stage framework leveraging human cognition. First, it prompts the policy to decompose zero-accuracy problems into subproblems. Second, it prompts the policy to refine answers by referencing previous wrong solutions. Moreover, to enable cold-starts and maintain train-test consistency, Cog-Rethinker applies supervised fine-tuning using correct samples from these stages. Experimental results demonstrate Cog-Rethinker’s superior performance on mathematical reasoning benchmarks and its improved sample efficiency that accelerates convergence compared to baselines.

Traditional recommender systems rely on passive feedback mechanisms that limit users to simple choices such as like and dislike. However, these coarse-grained signals fail to capture users' nuanced behavior motivations and intentions. In turn, current systems cannot also distinguish which specific item attributes drive user satisfaction or dissatisfaction, resulting in inaccurate preference modeling. These fundamental limitations create a persistent gap between user intentions and system interpretations, ultimately undermining user satisfaction and harming system effectiveness. To address these limitations, we introduce the Interactive Recommendation Feed (IRF), a pioneering paradigm that enables natural language commands within mainstream recommendation feeds. Unlike traditional systems that confine users to passive implicit behavioral influence, IRF empowers active explicit control over recommendation policies through real-time linguistic commands. To support this paradigm, we develop RecBot, a dual-agent architecture where a Parser Agent transforms linguistic expressions into structured preferences and a Planner Agent dynamically orchestrates adaptive tool chains for on-the-fly policy adjustment. To enable practical deployment, we employ simulation-augmented knowledge distillation to achieve efficient performance while maintaining strong reasoning capabilities. Through extensive offline and long-term online experiments, RecBot shows significant improvements in both user satisfaction and business outcomes.

Traditional social science research often requires designing complex experiments across vast methodological spaces and depends on real human participants, making it labor-intensive, costly, and difficult to scale. Here we present S-Researcher, an LLM-agent-based platform that assists researchers in conducting social science research more efficiently and at greater scale by "siliconizing" both the research process and the participant pool. To build S-Researcher, we first develop YuLan-OneSim, a large-scale social simulation system designed around three core requirements: generality via auto-programming from natural language to executable scenarios, scalability via a distributed architecture supporting up to 100,000 concurrent agents, and reliability via feedback-driven LLM fine-tuning. Leveraging this system, S-Researcher supports researchers in designing social experiments, simulating human behavior with LLM agents, analyzing results, and generating reports, forming a complete human-AI collaborative research loop in which researchers retain oversight and intervention at every stage. We operationalize LLM simulation research paradigms into three canonical reasoning modes (induction, deduction, and abduction) and validate S-Researcher through systematic case studies: inductive reproduction of cultural dynamics consistent with Axelrod's theory, deductive testing of competing hypotheses on teacher attention validated against survey data, and abductive identification of a cooperation mechanism in public goods games confirmed by human experiments. S-Researcher establishes a new human–AI collaborative paradigm for social science, in which computational simulation augments human researchers to accelerate discovery across the full spectrum of social inquiry.

In large language model-based agents, memory serves as a critical capability for achieving personalization by storing and utilizing users' information. Although some previous studies have adopted memory to implement user personalization, they typically focus on preference alignment and simple question-answering. However, in the real world, complex tasks often require multi-hop reasoning on a large amount of user information, which poses significant challenges for current memory approaches. To address this limitation, we propose the multi-hop personalized reasoning task to explore how different memory mechanisms perform in multi-hop reasoning over personalized information. We explicitly define this task and construct a dataset along with a unified evaluation framework. Then, we implement various explicit and implicit memory methods and conduct comprehensive experiments. We evaluate their performance on this task from multiple perspectives and analyze their strengths and weaknesses. Besides, we explore hybrid approaches that combine both paradigms and propose the HybridMem method to address their limitations. We demonstrate the effectiveness of our proposed model through extensive experiments. To benefit the research community, we release this project at https://github.com/nuster1128/MPR.

Sequential recommenders are commonly trained with one-hot next-item labels under a causal (prefix-only) objective aligned with inference. While deployment-compatible, this supervision offers little insight into relative preferences among non-target items. Yet logged interaction sequences contain an additional supervisory source: interactions following the target often reveal how user intent evolves, making the target easier to interpret. We treat these future interactions as training-only privileged information, available during learning but not at inference. This raises a natural question: can future interactions provide richer supervision while keeping training aligned with inference-time prediction? We propose Privileged Self-Distillation (PSD), a framework that separates learning-time information from inference-time information. PSD applies two attention masks to the same backbone: a future-aware view yields a privileged teacher distribution conditioned on past and future interactions, while a prefix-only view yields the student distribution used for deployment. Distilling the privileged distribution converts future interactions into training-only supervision rather than inference-time inputs. Since both views share a backbone, the teacher's advantage is purely informational, not architectural, removing the need for a separately pretrained teacher and letting its supervision adapt as the student evolves. PSD further uses an advantage-reachability gate to focus distillation on teacher signals likely supported by the observed prefix, along with a momentum-averaged teacher for stable targets. The framework is optimized end-to-end in a single stage, leaving the deployed model and inference cost unchanged. Experiments across public benchmarks and diverse backbones show consistent improvements.

Sequential Recommendation (SeqRec) aims to predict the next item by capturing sequential patterns from users' historical interactions, playing a crucial role in many real-world recommender systems. However, existing approaches predominantly adopt a direct forward computation paradigm, where the final hidden state of the sequence encoder serves as the user representation. We argue that this inference paradigm, due to its limited computational depth, struggles to model the complex evolving nature of user preferences and lacks a nuanced understanding of long-tail items, leading to suboptimal performance. To address this issue, we propose ReaRec, the first inference-time computing framework for recommender systems, which enhances user representations through implicit multi-step reasoning. Specifically, ReaRec autoregressively feeds the sequence's last hidden state into the sequential recommender while incorporating special reasoning position embeddings to decouple the original item encoding space from the multi-step reasoning space. Moreover, we introduce two lightweight reasoning-based learning methods, Ensemble Reasoning Learning (ERL) and Progressive Reasoning Learning (PRL), to further effectively exploit ReaRec's reasoning potential. Extensive experiments on five public real-world datasets and different SeqRec architectures demonstrate the generality and effectiveness of our proposed ReaRec. Remarkably, post-hoc analyses reveal that ReaRec significantly elevates the performance ceiling of multiple sequential recommendation backbones by approximately 30%-50%. Thus, we believe this work can open a new and promising avenue for future research in inference-time computing for sequential recommendation.

Recent advancements in explainable recommendation have greatly bolstered user experience by elucidating the decision-making rationale. However, the existing methods actually fail to provide effective feedback signals for potentially better or worse generated explanations due to their reliance on traditional supervised learning paradigms in sparse interaction data. To address these issues, we propose a novel human-like feedback-driven optimization framework. This framework employs a dynamic interactive optimization mechanism for achieving human-centered explainable requirements without incurring high labor costs. Specifically, we propose to utilize large language models (LLMs) as human simulators to predict human-like feedback for guiding the learning process. To enable the LLMs to deeply understand the task essence and meet user's diverse personalized requirements, we introduce a human-induced customized reward scoring method, which helps stimulate the language understanding and logical reasoning capabilities of LLMs. Furthermore, considering the potential conflicts between different perspectives of explanation quality, we introduce a principled Pareto optimization that transforms the multi-perspective quality enhancement task into a multi-objective optimization problem for improving explanation performance. At last, to achieve efficient model training, we design an off-policy optimization pipeline. By incorporating a replay buffer and addressing the data distribution biases, we can effectively improve data utilization and enhance model generality. Extensive experiments on four datasets demonstrate the superiority of our approach.

Memory is the key component for transforming a stateless LLM into a persistent, evolving agent through experience accumulation, long-horizon planning, and continual self-improvement. Existing memory systems typically take the LLM as the center and design memory operations tailored to a specific backbone. In practice, however, users frequently switch between LLMs, for example using Claude for coding and GPT for writing across tasks, or routing different steps to different backbones within a single task for cost-effective trade-offs. As a result, memory written by one model often needs to be consumed by another. Making upstream memory effectively adapt to and activate downstream LLMs remains a critical yet underexplored problem. To bridge this gap, we shift the perspective from LLM-centric memory design to memory-centric LLM adaptation. Specifically, we approach the above upstream-downstream memory adaptation problem from both the write and read sides, and design two profile-conditioned operators that are jointly trained to optimize how memory is stored and presented for better task completion. To ensure the learned operators generalize across a broad set of LLMs, we propose a minimum-gain sampling curriculum that prioritizes the least-served LLMs during training. To better measure the operators' actual contribution rather than the LLM's own capability, we design a performance-gap reward that compares against a naive memory baseline. Experiments on HotpotQA, 2WikiMultihopQA, and MuSiQue demonstrate that our model consistently outperforms baselines and remains robust under unseen-model replacement.

Memory is critical for LLM-based agents to preserve past observations for future decision-making, where factual memory serves as its foundational part. However, existing approaches to constructing factual memory face several limitations. Textual methods impose heavy context and indexing burdens, while parametric methods suffer from catastrophic forgetting and high costs. To address these challenges, we introduce NextMem, a latent factual memory framework that utilizes an autoregressive autoencoder to efficiently construct latent memory while ensuring accurate reconstruction. For better optimization, we propose a two-stage training process, including autoregressive reconstruction alignment and progressive latent substitution. We also incorporate quantization to reduce storage overhead. Extensive experiments demonstrate that NextMem achieves superior performance, and excels in retrieval, robustness, and extensibility properties. We release our code and model checkpoints at https://github.com/nuster1128/NextMem.

Reinforcement learning for multi-step LLM agents often relies on scalar rewards that indicate success but cannot explain why a trajectory is good or bad. Rubric-based rewards improve interpretability through natural-language criteria, but existing methods share two limitations: they score at the trajectory level, offering no guidance for individual steps; and their scorer is closed-source and static, so it cannot adapt as the agent evolves during training. We propose ARCO (Adaptive Rubric CO-evolution), which generates a per-step rubric and predicts a rubric-conditioned step-level reward for each action, and continually updates this rubric model on on-policy rollouts so that its criteria and scores co-evolve with the agent's improving behavior. Across HotpotQA, 2WikiMultiHopQA, and MuSiQue with two open-source backbones, ARCO achieves the highest EM in all settings over outcome-, rubric-, and process-reward baselines, and analyses show its rubrics are step-specific, robust to design choices, and useful for diagnosing agent behavior. Code and data are available at https://github.com/zihangtian/ARCO.

Large language model (LLM)-powered assistants have recently integrated memory mechanisms that record user preferences, leading to more personalized and user-aligned responses. However, irrelevant personalized memories are often introduced into the context, interfering with the LLM's intent understanding. To comprehensively investigate the dual effects of personalization, we develop RPEval, a benchmark comprising a personalized intent reasoning dataset and a multi-granularity evaluation protocol. RPEval reveals the widespread phenomenon of irrational personalization in existing LLMs and, through error pattern analysis, illustrates its negative impact on user experience. Finally, we introduce RP-Reasoner, which treats memory utilization as a pragmatic reasoning process, enabling the selective integration of personalized information. Experimental results demonstrate that our method significantly outperforms carefully designed baselines on RPEval, and resolves 80

Capturing complex user preferences from sparse behavioral sequences remains a fundamental challenge in sequential recommendation. Recent latent reasoning methods have shown promise by extending test-time computation through multi-step reasoning, yet they exclusively rely on depth-level scaling along a single trajectory, suffering from diminishing returns as reasoning depth increases. To address this limitation, we propose Parallel Latent Reasoning (PLR), a novel framework that pioneers width-level computational scaling by exploring multiple diverse reasoning trajectories simultaneously. PLR constructs parallel reasoning streams through learnable trigger tokens in continuous latent space, preserves diversity across streams via global reasoning regularization, and adaptively synthesizes multi-stream outputs through mixture-of-reasoning-streams aggregation. Extensive experiments on three real-world datasets demonstrate that PLR substantially outperforms state-of-the-art baselines while maintaining real-time inference efficiency. Theoretical analysis further validates the effectiveness of parallel reasoning in improving generalization capability. Our work opens new avenues for enhancing reasoning capacity in sequential recommendation beyond existing depth scaling.

Curriculum learning plays a crucial role in enhancing the training efficiency of large language models (LLMs) on reasoning tasks. However, existing methods often fail to adequately account for variations in prompt difficulty or rely on simplistic filtering mechanisms to select prompt datasets within a narrow criterion range, resulting in significant computational waste. In this work, we approach the problem from the perspective of reinforcement learning gradient optimization, offering a systematic and theoretical investigation into how to improve the training efficiency of LLMs. We identify two key factors influencing training efficiency: the selection of training prompts and the allocation of rollout quantities across different prompts. Our theoretical analysis reveals that the sampling distribution of prompts dictates the convergence rate of gradient descent, while the allocation of the rollout quantity influences the consistency and stability of overall gradient updates. Based on these insights, we propose CurES, an efficient training method that accelerates convergence and employs Bayesian posterior estimation to minimize computational overhead. Experiments demonstrate that our CurES outperforms Group Relative Policy Optimization (GRPO) by \textbf{+3.3} points and \textbf{+4.82} points with 1.5B and 7B models, respectively, and exceeds the best prior sample efficient methods by \textbf{+2.12} points on average across eight math reasoning benchmarks. Our CurES also improves convergence speed compare to baselines such as GRPO.

Simulating realistic user shopping behavior underpins offline evaluation and reinforcement learning in e-commerce scenarios. While recent LLM- and VLM-based simulators have made encouraging progress, reproducing a real browsing session remains difficult for two reasons. (i) Memory Challenge: a shopping session spans dozens of pages, yet existing agents either discard long-range observation histories, losing the evolving user state, or naively concatenate them, overwhelming the context window and even degrading simulation quality. (ii) Optimization Challenge: current user simulators are typically supervised to match each logged action via imitation or step-level rewards; the resulting sessions often display unrealistic patterns, such as over-exploration or excessive passivity, which per-step supervision can neither detect nor correct. To address the above challenges, we present RecVerse, a GUI-grounded simulation agent that perceives pages through screenshots and produces faithful multi-turn trajectories. For the memory challenge, RecVerse adopts a cognitive-inspired hierarchical memory: Working Memory for short-term focus, Episodic Memory for in-session traces, and Preference Memory for high-level intent, with memory updates treated as actions so that the agent adaptively learns when and what to memorize. For the optimization challenge, RecVerse is optimized with a trajectory-level RL objective that scores entire sessions, aligning both macro-level action-type distributions and micro-level shopping intent with real users. We further release USB (User Simulation Benchmark), an interactive e-commerce GUI trajectory dataset for multi-turn user simulation. Experiments show that RecVerse significantly outperforms existing baselines in both behavioral fidelity and intent consistency.

Large language models (LLMs) are increasingly used as stand-ins for human respondents, from opinion polls and simulated survey participants to agent-based social simulations. These uses rest on one assumption: that conditioning a model on who a person is yields answers resembling those of real people from that group. Here we identify and measure benevolence bias, a small but consistent tendency for aligned LLMs to lean toward the kinder, safer, more socially approved answer on value-laden survey questions. Across 18 widely used models, four social-science datasets (ANES, GSS, WVS, and a cross-cultural prospect-theory replication) and six psychological categories, we find that the bias is a stable model property, not a quirk of any one system: it points the same way across models, grows with model size, and traces to the post-training stage. Prompt language and framing change its size but never its direction, and a "malicious persona" stress test shows a one-sided limit: aligned models struggle to play people who are less kind, less prosocial or more harm-tolerant than average. The issue is thus not only a shifted average, but a narrowed range of people the model can imitate. The bias sits in the middle of the answer distribution rather than its tails, and survives changes in sampling temperature and simple prompted reflection. The encouraging news is that it is easy to diagnose and straightforward to fix: a light-touch contrastive calibration, which needs no retraining and works on black-box APIs, brings all six categories back to the human baseline. Our results give researchers a clear map of where aligned LLMs can already be trusted as human stand-ins, where they need care, and a ready-to-use method for closing the gap.

Conversational Recommender Systems (CRSs) enhance user experience through multi-turn interactions, yet evaluating their performance remains challenging. While Large Language Model (LLM) based user simulators are effective, they suffer from three key limitations: (1) Lack of Domain Adaptability: Reliance on fixed prompts and predefined action spaces hinders transfer to novel domains; (2) Limited User Modeling: Inability to accurately replicate subtle linguistic styles and dynamic preferences; (3) Insufficient Evaluation Validity: Existing simulators fail to adequately assess fundamental capabilities and system robustness. To overcome these, we propose AdaptSim, an Adaptive domain and automatic prompt tuning User Simulator. AdaptSim offers an efficient framework for evaluating CRSs by enabling realistic behavior modeling and diverse style generation. It leverages automatic prompt generation and an open action mechanism to reduce manual effort and improve cross-domain flexibility. For response generation, we employ controlled text generation with a "think-then-respond" strategy for fine-grained control over language style. For CRS evaluation, AdaptSim incorporates a novel Breadth-First Search (BFS)-based, turn-level pairwise comparison framework for comprehensive assessment. Extensive experiments across three domains and four LLMs demonstrate that AdaptSim generates realistic dialogues, enabling a highly effective and reliable evaluation of CRS capabilities and robustness.

Large Language Model (LLM)-based agents increasingly rely on memory to learn from experiences over continual interactions. However, storing experiences as independent, flat units leads to substantial redundancy and retrieval conflicts, as similar episodes repeat overlapping content and subtle scene variations cause retrieved memories to offer contradictory guidance. To address this, we introduce residual experience, positing that newly acquired experience is often an incremental variation of existing knowledge. We propose DeltaMem, a framework that organizes experience memory into two independent residual trees, one storing goal-conditioned task experience as reusable skills and another for scene-level environment knowledge. Each tree uses a root node for generalized base experiences and incremental delta nodes for subsequent variations, allowing related experiences to share a common foundation without duplication. For retrieval, a failure-penalized similarity scan locates the best match, reconstructing the full experience via root-to-match chain composition. An autonomous consolidation mechanism distills high-frequency paths into new root nodes, enabling the trees to self-organize from general heuristics to specialized variants. Experiments across diverse interactive environments show that DeltaMem consistently outperforms existing baselines. To facilitate future research, we release the code at https://github.com/import-myself/DeltaMem.