Large language models (LLMs) are increasingly used to support question answering and decision-making in high-stakes, domain-specific settings such as natural hazard response and infrastructure planning, where effective answers must convey fine-grained, decision-critical details. However, existing evaluation frameworks for retrieval-augmented generation (RAG) and open-ended question answering primarily rely on surface-level similarity, factual consistency, or semantic relevance, and often fail to assess whether responses provide the specific information required for domain-sensitive decisions. To address this gap, we propose a multi-dimensional, reference-free evaluation framework that assesses LLM outputs along four complementary dimensions: specificity, robustness to paraphrasing and semantic perturbations, answer relevance, and context utilization. We introduce a curated dataset of 1,412 domain-specific question-answer pairs spanning 40 professional roles and seven natural hazard types to support systematic evaluation. We further conduct human evaluation to assess inter-annotator agreement and alignment between model outputs and human judgments, which highlights the inherent subjectivity of open-ended, domain-specific evaluation. Our results show that no single metric sufficiently captures answer quality in isolation and demonstrate the need for structured, multi-metric evaluation frameworks when deploying LLMs in high-stakes applications.
Large Language Models (LLMs) are reshaping recommender systems by leveraging extensive world knowledge and semantic reasoning to interpret user intent. However, effectively integrating these capabilities with collaborative signals while avoiding prohibitive inference latency remains a critical bottleneck. To address this, we propose a trajectory-driven internalization framework to develop a Single-agent Trajectory-Aligned Recommender (STAR). Specifically, to internalize complex reasoning capabilities into a single efficient model, we first design a multi-agent teacher system capable of multi-turn tool usage and reflection. This teacher utilizes a Collaborative Signal Translation mechanism to explicitly convert latent behavioral patterns into descriptive natural language evidence to enhance reasoning accuracy. Subsequently, a trajectory-driven distillation pipeline transfers this agentic logic, including planning, tool usage, and self-reflection, into the compact STAR model. Extensive experiments demonstrate that STAR surpasses its teacher by 8.7% to 39.5% while eliminating iterative latency, paving the way for real-time, reasoning-enhanced recommendation.
This paper introduces AnalyticsGPT, an intuitive and efficient large language model (LLM)-powered workflow for scientometric question answering. This underrepresented downstream task addresses the subcategory of meta-scientific questions concerning the "science of science." When compared to traditional scientific question answering based on papers, the task poses unique challenges in the planning phase. Namely, the need for named-entity recognition of academic entities within questions and multi-faceted data retrieval involving scientometric indices, e.g. impact factors. Beyond their exceptional capacity for treating traditional natural language processing tasks, LLMs have shown great potential in more complex applications, such as task decomposition and planning and reasoning. In this paper, we explore the application of LLMs to scientometric question answering, and describe an end-to-end system implementing a sequential workflow with retrieval-augmented generation and agentic concepts. We also address the secondary task of effectively synthesizing the data into presentable and well-structured high-level analyses. As a database for retrieval-augmented generation, we leverage a proprietary research performance assessment platform. For evaluation, we consult experienced subject matter experts and leverage LLMs-as-judges. In doing so, we provide valuable insights on the efficacy of LLMs towards a niche downstream task. Our (skeleton) code and prompts are available at: https://github.com/lyvykhang/llm-agents-scientometric-qa/tree/acl.
Tool-integrated reasoning (TIR) enables LLM agents to solve tasks through planning, tool use, and iterative revision, but outcome-only reinforcement learning in this setting suffers from sparse, delayed rewards and weak step-level credit assignment. In long-horizon TIR trajectories, an early irrecoverable mistake can determine success or failure, making it crucial to localize the first irrecoverable step and leverage it for fine-grained credit assignment. We propose Error-Localized Policy Optimization (ELPO), which localizes the first irrecoverable step via binary-search rollout trees under a fixed rollout budget, converts the resulting tree into stable learning signals through hierarchical advantage attribution, and applies error-localized adaptive clipping to strengthen corrective updates on the critical step and its suffix. Across TIR benchmarks in math, science QA, and code execution, ELPO consistently outperforms strong Agentic RL baselines under comparable sampling budgets, with additional gains in Pass@K and Major@K scaling, rollout ranking quality, and tool-call efficiency. Our code will be publicly released soon.
Long-horizon planning is widely recognized as a core capability of autonomous LLM-based agents; however, current evaluation frameworks suffer from being largely episodic, domain-specific, or insufficiently grounded in persistent economic dynamics. We introduce EcoGym, a generalizable benchmark for continuous plan-and-execute decision making in interactive economies. EcoGym comprises three diverse environments: Vending, Freelance, and Operation, implemented in a unified decision-making process with standardized interfaces, and budgeted actions over an effectively unbounded horizon (1000+ steps if 365 day-loops for evaluation). The evaluation of EcoGym is based on business-relevant outcomes (e.g., net worth, income, and DAU), targeting long-term strategic coherence and robustness under partial observability and stochasticity. Experiments across eleven leading LLMs expose a systematic tension: no single model dominates across all three scenarios. Critically, we find that models exhibit significant suboptimality in either high-level strategies or efficient actions executions. EcoGym is released as an open, extensible testbed for transparent long-horizon agent evaluation and for studying controllability-utility trade-offs in realistic economic settings.
While Large Language Models (LLM) enable non-experts to specify open-world multi-robot tasks, the generated plans often lack kinematic feasibility and are not efficient, especially in long-horizon scenarios. Formal methods like Linear Temporal Logic (LTL) offer correctness and optimal guarantees, but are typically confined to static, offline settings and struggle with computational scalability. To bridge this gap, we propose a neuro-symbolic framework that grounds LLM reasoning into hierarchical LTL specifications and solves the corresponding Simultaneous Task Allocation and Planning (STAP) problem. Unlike static approaches, our system resolves stochastic environmental changes, such as moving users or updated instructions via a receding horizon planning (RHP) loop with real-time perception, which dynamically refines plans through a hierarchical state space. Extensive real-world experiments demonstrate that our approach significantly outperforms baseline methods in success rate and interaction fluency while minimizing planning latency.
Understanding an agent's goals helps explain and predict its behaviour, yet there is no established methodology for reliably attributing goals to agentic systems. We propose a framework for evaluating goal-directedness that integrates behavioural evaluation with interpretability-based analyses of models' internal representations. As a case study, we examine an LLM agent navigating a 2D grid world toward a goal state. Behaviourally, we evaluate the agent against an optimal policy across varying grid sizes, obstacle densities, and goal structures, finding that performance scales with task difficulty while remaining robust to difficulty-preserving transformations and complex goal structures. We then use probing methods to decode the agent's internal representations of the environment state and its multi-step action plans. We find that the LLM agent non-linearly encodes a coarse spatial map of the environment, preserving approximate task-relevant cues about its position and the goal location; that its actions are broadly consistent with these internal representations; and that reasoning reorganises them, shifting from broader environment structural cues toward information supporting immediate action selection. Our findings support the view that introspective examination is required beyond behavioural evaluations to characterise how agents represent and pursue their objectives.
Generating step-by-step "how-to" procedures is a key LLM capability: how-to advice is commonly requested in chatbots, and step-by-step planning is critical for reasoning over complex tasks. Yet, measuring and improving procedural validity at scale on real-world tasks remains challenging and understudied. To address this, we introduce How2Everything, a scalable framework to evaluate and improve goal-conditioned procedure generation. Our framework includes How2Mine, which mines 351K procedures from 980K web pages across 14 topics and readily scales to larger corpora. From this pool we build How2Bench, a 7K-example evaluation set balanced across topics. To reliably score model outputs, we develop How2Score, an evaluation protocol that uses an LLM judge to detect whether a generation contains any critical failure that would prevent achieving the goal. For low-cost, reproducible evaluation, we distill a frontier model into an open 8B model, achieving 80.5% agreement with human annotators. How2Bench reveals clear scaling trends across model sizes and training stages, providing signal early in pretraining. Finally, RL using How2Score as a reward improves performance on How2Bench by >10 points across three models without systematic regressions on standard benchmarks, with gains robust to superficial source-document memorization or format compliance. Taken together, How2Everything shows how pretraining web data can support a closed loop of capability evaluation and improvement at scale.
The advancement of large language models (LLMs) has significantly accelerated the development of search agents capable of autonomously gathering information through multi-turn web interactions. Various benchmarks have been proposed to evaluate such agents. However, existing benchmarks often construct queries backward from answers, producing unnatural tasks misaligned with real-world needs. Moreover, these benchmarks tend to focus on either locating specific information or aggregating information from multiple sources, while relying on static answer sets prone to data contamination. To bridge these gaps, we introduce GISA, a benchmark for General Information-Seeking Assistants comprising 373 human-crafted queries that reflect authentic information-seeking scenarios. GISA features four structured answer formats (item, set, list, and table), enabling deterministic evaluation. It integrates both deep reasoning and broad information aggregation within unified tasks, and includes a live subset with periodically updated answers to resist memorization. Notably, GISA provides complete human search trajectories for every query, offering gold-standard references for process-level supervision and imitation learning. Experiments on mainstream LLMs and commercial search products reveal that even the best-performing model achieves only 19.30\% exact match score, with performance notably degrading on tasks requiring complex planning and comprehensive information gathering. These findings highlight substantial room for future improvement.
Integration of VLM reasoning with symbolic planning has proven to be a promising approach to real-world robot task planning. Existing work like UniDomain effectively learns symbolic manipulation domains from real-world demonstrations, described in Planning Domain Definition Language (PDDL), and has successfully applied them to real-world tasks. These domains, however, are restricted to tabletop manipulation. We propose UniPlan, a vision-language task planning system for long-horizon mobile-manipulation in large-scale indoor environments, that unifies scene topology, visuals, and robot capabilities into a holistic PDDL representation. UniPlan programmatically extends learned tabletop domains from UniDomain to support navigation, door traversal, and bimanual coordination. It operates on a visual-topological map, comprising navigation landmarks anchored with scene images. Given a language instruction, UniPlan retrieves task-relevant nodes from the map and uses a VLM to ground the anchored image into task-relevant objects and their PDDL states; next, it reconnects these nodes to a compressed, densely-connected topological map, also represented in PDDL, with connectivity and costs derived from the original map; Finally, a mobile-manipulation plan is generated using off-the-shelf PDDL solvers. Evaluated on human-raised tasks in a large-scale map with real-world imagery, UniPlan significantly outperforms VLM and LLM+PDDL planning in success rate, plan quality, and computational efficiency.
Large language models (LLMs) have opened new opportunities for transforming natural language user intents into executable actions. This capability enables embodied AI agents to perform complex tasks, without involvement of an expert, making human-robot interaction (HRI) more convenient. However these developments raise significant security and privacy challenges such as self-preferencing, where a single LLM service provider dominates the market and uses this power to promote their own preferences. LLM oracles have been recently proposed as a mechanism to decentralize LLMs by executing multiple LLMs from different vendors and aggregating their outputs to obtain a more reliable and trustworthy final result. However, the accuracy of these approaches highly depends on the aggregation method. The current aggregation methods mostly use semantic similarity between various LLM outputs, not suitable for robotic task planning, where the temporal order of tasks is important. To fill the gap, we propose an LLM oracle with a new aggregation method for robotic task planning. In addition, we propose a decentralized multi-robot infrastructure based on Hyperledger Fabric that can host the proposed oracle. The proposed infrastructure enables users to express their natural language intent to the system, which then can be decomposed into subtasks. These subtasks require coordinating different robots from different vendors, while enforcing fine-grained access control management on the data. To evaluate our methodology, we created the SkillChain-RTD benchmark made it publicly available. Our experimental results demonstrate the feasibility of the proposed architecture, and the proposed aggregation method outperforms other aggregation methods currently in use.
Large Language Models (LLMs) show promise as planners for embodied AI, but their stochastic nature lacks formal reasoning, preventing strict safety guarantees for physical deployment. Current approaches often rely on unreliable LLMs for safety checks or simply reject unsafe plans without offering repairs. We introduce the Verifiable Iterative Refinement Framework (VIRF), a neuro-symbolic architecture that shifts the paradigm from passive safety gatekeeping to active collaboration. Our core contribution is a tutor-apprentice dialogue where a deterministic Logic Tutor, grounded in a formal safety ontology, provides causal and pedagogical feedback to an LLM planner. This enables intelligent plan repairs rather than mere avoidance. We also introduce a scalable knowledge acquisition pipeline that synthesizes safety knowledge bases from real-world documents, correcting blind spots in existing benchmarks. In challenging home safety tasks, VIRF achieves a perfect 0 percent Hazardous Action Rate (HAR) and a 77.3 percent Goal-Condition Rate (GCR), which is the highest among all baselines. It is highly efficient, requiring only 1.1 correction iterations on average. VIRF demonstrates a principled pathway toward building fundamentally trustworthy and verifiably safe embodied agents.
Databases are the most critical assets for enterprises, and yet they remain largely inaccessible to people who make the most important decisions. In this paper, we describe the Tursio search platform that builds an abstraction layer, aka semantic knowledge graph, over the underlying databases to make them searchable in natural language. Tursio infuses large language models (LLMs) into every part of the query processing stack, including data modeling, query compilation, query planning, and result reasoning. This allows Tursio to process natural language queries systematically using techniques from traditional query planning and rewriting, rather than black-box memorization. We describe the architecture of Tursio in detail and present a comprehensive evaluation on production workloads, and synthetic and realistic benchmarks. Our results show that Tursio achieves high accuracy while being efficient and scalable, making databases truly searchable for non-expert users.
Personal data centralization among dominant platform providers including search engines, social networking services, and e-commerce has created siloed ecosystems that restrict user sovereignty, thereby impeding data use across services. Meanwhile, the rapid proliferation of Large Language Model (LLM)-based agents has intensified demand for highly personalized services that require the dynamic provision of diverse personal data. This presents a significant challenge: balancing the utilization of such data with privacy protection. To address this challenge, we propose Puda (Private User Dataset Agent), a user-sovereign architecture that aggregates data across services and enables client-side management. Puda allows users to control data sharing at three privacy levels: (i) Detailed Browsing History, (ii) Extracted Keywords, and (iii) Predefined Category Subsets. We implemented Puda as a browser-based system that serves as a common platform across diverse services and evaluated it through a personalized travel planning task. Our results show that providing Predefined Category Subsets achieves 97.2% of the personalization performance (evaluated via an LLM-as-a-Judge framework across three criteria) obtained when sharing Detailed Browsing History. These findings demonstrate that Puda enables effective multi-granularity management, offering practical choices to mitigate the privacy-personalization trade-off. Overall, Puda provides an AI-native foundation for user sovereignty, empowering users to safely leverage the full potential of personalized AI.
To address the instability of unguided reasoning paths in standard Chain-of-Thought prompting, recent methods guide large language models (LLMs) by first eliciting a single reasoning strategy. However, relying on just one strategy for each question can still limit performance across diverse tasks. We propose Diverge-to-Induce Prompting (DIP), a framework that first prompts an LLM to generate multiple diverse high-level rationales for each question. Each rationale is then elaborated into a detailed, step-by-step draft plan. Finally, these draft plans are induced into a final plan. DIP enhances zero-shot reasoning accuracy without reliance on resource-intensive sampling. Experiments show that DIP outperforms single-strategy prompting, demonstrating the effectiveness of multi-plan induction for prompt-based reasoning.
Real-world offensive security operations are inherently open-ended: attackers explore unknown attack surfaces, revise hypotheses under uncertainty, and operate without guaranteed success. Existing LLM-based offensive agent evaluations rely on closed-world settings with predefined goals and binary success criteria. To address this gap, we introduce CyberExplorer, an evaluation suite with two core components: (1) an open-environment benchmark built on a virtual machine hosting 40 vulnerable web services derived from real-world CTF challenges, where agents autonomously perform reconnaissance, target selection, and exploitation without prior knowledge of vulnerability locations; and (2) a reactive multi-agent framework supporting dynamic exploration without predefined plans. CyberExplorer enables fine-grained evaluation beyond flag recovery, capturing interaction dynamics, coordination behavior, failure modes, and vulnerability discovery signals-bridging the gap between benchmarks and realistic multi-target attack scenarios.
Large language models (LLMs) are increasingly capable of carrying out long-running, real-world tasks. However, as the amount of context grows, their reliability often deteriorates, a phenomenon known as "context rot". Existing long-context benchmarks primarily focus on single-step settings that evaluate a model's ability to retrieve information from a long snippet. In realistic scenarios, however, LLMs often need to act as agents that explore environments, follow instructions and plans, extract useful information, and predict correct actions under a dynamically growing context. To assess language agents in such settings, we introduce LOCA-bench (a benchmark for LOng-Context Agents). Given a task prompt, LOCA-bench leverages automated and scalable control of environment states to regulate the agent's context length. This design enables LOCA-bench to extend the context length potentially to infinity in a controlled way while keeping the underlying task semantics fixed. LOCA-bench evaluates language agents as a combination of models and scaffolds, including various context management strategies. While agent performance generally degrades as the environment states grow more complex, advanced context management techniques can substantially improve the overall success rate. We open-source LOCA-bench to provide a platform for evaluating models and scaffolds in long-context, agentic scenarios: https://github.com/hkust-nlp/LOCA-bench
Demand forecasting is a cornerstone of e-commerce operations, directly impacting inventory planning and fulfillment scheduling. However, existing forecasting systems often fail during high-impact periods such as flash sales, holiday campaigns, and sudden policy interventions, where demand patterns shift abruptly and unpredictably. In this paper, we introduce EventCast, a modular forecasting framework that integrates future event knowledge into time-series prediction. Unlike prior approaches that ignore future interventions or directly use large language models (LLMs) for numerical forecasting, EventCast leverages LLMs solely for event-driven reasoning. Unstructured business data, which covers campaigns, holiday schedules, and seller incentives, from existing operational databases, is processed by an LLM that converts it into interpretable textual summaries leveraging world knowledge for cultural nuances and novel event combinations. These summaries are fused with historical demand features within a dual-tower architecture, enabling accurate, explainable, and scalable forecasts. Deployed on real-world e-commerce scenarios spanning 4 countries of 160 regions over 10 months, EventCast achieves up to 86.9% and 97.7% improvement on MAE and MSE compared to the variant without event knowledge, and reduces MAE by up to 57.0% and MSE by 83.3% versus the best industrial baseline during event-driven periods. EventCast has deployed into real-world industrial pipelines since March 2025, offering a practical solution for improving operational decision-making in dynamic e-commerce environments.
We study online configuration selection with admission control problem, which arises in LLM serving, GPU scheduling, and revenue management. In a planning horizon with $T$ periods, we consider a two-layer framework for the decisions made within each time period. In the first layer, the decision maker selects one of the $K$ configurations (ex. quantization, parallelism, fare class) which induces distribution over the reward-resource pair of the incoming request. In the second layer, the decision maker observes the request and then decides whether to accept it or not. Benchmarking this framework requires care. We introduce a \textbf{switching-aware fluid oracle} that accounts for the value of mixing configurations over time, provably upper-bounding any online policy. We derive a max-min formulation for evaluating the benchmark, and we characterize saddle points of the max-min problem via primal-dual optimality conditions linking equilibrium, feasibility, and complementarity. This guides the design of \textbf{SP-UCB--OLP} algorithm, which solves an optimistic saddle point problem and achieves $\tilde{O}(\sqrt{KT})$ regret.
Material synthesis planning (MSP) remains a fundamental and underexplored bottleneck in AI-driven materials discovery, as it requires not only identifying suitable precursor materials but also designing coherent sequences of synthesis operations to realize a target material. Although several AI-based approaches have been proposed to address isolated subtasks of MSP, a unified methodology for solving the entire MSP task has yet to be established. We propose MSP-LLM, a unified LLM-based framework that formulates MSP as a structured process composed of two constituent subproblems: precursor prediction (PP) and synthesis operation prediction (SOP). Our approach introduces a discrete material class as an intermediate decision variable that organizes both tasks into a chemically consistent decision chain. For OP, we further incorporate hierarchical precursor types as synthesis-relevant inductive biases and employ an explicit conditioning strategy that preserves precursor-related information in the autoregressive decoding state. Extensive experiments show that MSP-LLM consistently outperforms existing methods on both PP and SOP, as well as on the complete MSP task, demonstrating an effective and scalable framework for MSP that can accelerate real-world materials discovery.
Large language models demonstrate limited capability in proficiency-controlled sentence simplification, particularly when simplifying across large readability levels. We propose a framework that decomposes complex simplifications into manageable steps through dynamic path planning, semantic-aware exemplar selection, and chain-of-thought generation with conversation history for coherent reasoning. Evaluation on five languages across two benchmarks shows our approach improves simplification effectiveness while reducing computational steps by 22-42%. Human evaluation confirms the fundamental trade-off between simplification effectiveness and meaning preservation. Notably, even human annotators struggle to agree on semantic preservation judgments, highlighting the inherent complexity of this task. Our work shows that while step-by-step simplification improves control, preserving semantic fidelity during extensive simplification remains an open challenge.
Large language models (LLMs) are increasingly embedded into recommender systems, where they operate across multiple functional roles such as data augmentation, profiling, and decision making. While prior work emphasizes recommendation performance, the systemic risks of LLMs, such as bias and hallucination, and their propagation through feedback loops remain largely unexplored. In this paper, we propose a role-aware, phase-wise diagnostic framework that traces how these risks emerge, manifest in ranking outcomes, and accumulate over repeated recommendation cycles. We formalize a controlled feedback-loop pipeline that simulates long-term interaction dynamics and enables empirical measurement of risks at the LLM-generated content, ranking, and ecosystem levels. Experiments on widely used benchmarks demonstrate that LLM-based components can amplify popularity bias, introduce spurious signals through hallucination, and lead to polarized and self-reinforcing exposure patterns over time. We plan to release our framework as an open-source toolkit to facilitate systematic risk analysis across diverse LLM-powered recommender systems.
As Low-Rank Adaptation (LoRA) becomes the standard approach for efficiently fine-tuning large language models (LLMs), shared clusters increasingly execute many concurrent LoRA training jobs over the same frozen backbone. While recent advances enable batching (co-locating) multiple adapters during serving, efficient training-time co-location of heterogeneous LoRA adapters presents unique challenges. Jobs often differ in adapter rank, batch size, and resource allocation, and naïve batching can introduce synchronization stalls, communication overheads, and per-job slowdowns that are worse than executing independently. We introduce tLoRA, a framework that enables efficient batch training of multiple LoRA jobs. tLoRA fuses adapters that share the same base model into an elastic shared super-model, exploiting existing distributed training frameworks to derive parallelism plans that share resources effectively. At the kernel level, tLoRA employs a fused LoRA kernel that adaptively reconstructs low-rank computation tiles and schedules rank-aware nano-batches to maximize overlap between computation and communication across adapters. At the scheduling layer, tLoRA incorporates an online, residual-capacity-aware scheduler that adaptively groups jobs to maximize collective throughput. Evaluations using real-world cluster traces demonstrate that tLoRA improves training throughput by 1.2--1.8x, job training completion time by 2.3--5.4x, and GPU utilization by 37%.
Test-time compute allocation in large reasoning models (LRMs) is widely used and has applications in mathematical problem solving, code synthesis, and planning. Recent work has addressed this problem by scaling self-consistency and parallel thinking, adding generic ``thinking tokens'' and prompting models to re-read the question before answering. Unfortunately, these approaches either inject task-agnostic tokens or mandate heuristics that do not explain -- and often ignore -- the \emph{spontaneous} repetition that many LRMs exhibit at the head of their internal chains. In contrast, we analyze and harness the model's tendency to restate the question, which we term the \emph{Echo of Prompt (EOP)}, as a front-loaded, compute-shaping mechanism. We formalize its probabilistic cost by casting echo removal as rejection-based conditioning and defining the \emph{Echo Likelihood Gap} $Δ\mathcal{L}$ as a computable proxy. This provides the missing theoretical link that links early repetition to likelihood gains and downstream accuracy. However, it does not by itself specify how to exploit EOP. Consequently, we develop \emph{Echo-Distilled SFT (ED-SFT)} to instill an ``echo-then-reason'' pattern through supervised finetuning, and \emph{Echoic Prompting (EP)} to re-ground the model mid-trace without training. While promising, quantifying benefits beyond verbosity is non-trivial. Therefore, we conduct length and suffix-controlled likelihood analyses together with layer-wise attention studies, showing that EOP increases answer to answer-prefix attention in middle layers, consistent with an \emph{attention refocusing} mechanism. We evaluate on GSM8K, MathQA, Hendrycks-MATH, AIME24, and MATH-500 under identical decoding settings and budgets, and find consistent gains over baselines. Code is available at https://github.com/hhh2210/echoes-as-anchors.
Large language model (LLM)-based multi-agent systems (MAS) show strong promise for complex reasoning, planning, and tool-augmented tasks, but designing effective MAS architectures remains labor-intensive, brittle, and hard to generalize. Existing automatic MAS generation methods either rely on code generation, which often leads to executability and robustness failures, or impose rigid architectural templates that limit expressiveness and adaptability. We propose Evolutionary Generation of Multi-Agent Systems (EvoMAS), which formulates MAS generation as structured configuration generation. EvoMAS performs evolutionary generation in configuration space. Specifically, EvoMAS selects initial configurations from a pool, applies feedback-conditioned mutation and crossover guided by execution traces, and iteratively refines both the candidate pool and an experience memory. We evaluate EvoMAS on diverse benchmarks, including BBEH, SWE-Bench, and WorkBench, covering reasoning, software engineering, and tool-use tasks. EvoMAS consistently improves task performance over both human-designed MAS and prior automatic MAS generation methods, while producing generated systems with higher executability and runtime robustness. EvoMAS outperforms the agent evolution method EvoAgent by +10.5 points on BBEH reasoning and +7.1 points on WorkBench. With Claude-4.5-Sonnet, EvoMAS also reaches 79.1% on SWE-Bench-Verified, matching the top of the leaderboard.
Internal modelling of the world -- predicting transitions between previous states $X$ and next states $Y$ under actions $Z$ -- is essential to reasoning and planning for LLMs and VLMs. Learning such models typically requires costly action-labelled trajectories. We propose SWIRL, a self-improvement framework that learns from state-only sequences by treating actions as a latent variable and alternating between Forward World Modelling (FWM) $P_θ(Y|X,Z)$ and an Inverse Dynamics Modelling (IDM) $Q_φ(Z|X,Y)$. SWIRL iterates two phases: (1) Variational Information Maximisation, which updates the FWM to generate next states that maximise conditional mutual information with latent actions given prior states, encouraging identifiable consistency; and (2) ELBO Maximisation, which updates the IDM to explain observed transitions, effectively performing coordinate ascent. Both models are trained with reinforcement learning (specifically, GRPO) with the opposite frozen model's log-probability as a reward signal. We provide theoretical learnability guarantees for both updates, and evaluate SWIRL on LLMs and VLMs across multiple environments: single-turn and multi-turn open-world visual dynamics and synthetic textual environments for physics, web, and tool calling. SWIRL achieves gains of 16% on AURORABench, 28% on ByteMorph, 16% on WorldPredictionBench, and 14% on StableToolBench.
The rapid advancement of Large Language Models (LLMs) has catalyzed the development of autonomous agents capable of navigating complex environments. However, existing evaluations primarily adopt a deductive paradigm, where agents execute tasks based on explicitly provided rules and static goals, often within limited planning horizons. Crucially, this neglects the inductive necessity for agents to discover latent transition laws from experience autonomously, which is the cornerstone for enabling agentic foresight and sustaining strategic coherence. To bridge this gap, we introduce OdysseyArena, which re-centers agent evaluation on long-horizon, active, and inductive interactions. We formalize and instantiate four primitives, translating abstract transition dynamics into concrete interactive environments. Building upon this, we establish OdysseyArena-Lite for standardized benchmarking, providing a set of 120 tasks to measure an agent's inductive efficiency and long-horizon discovery. Pushing further, we introduce OdysseyArena-Challenge to stress-test agent stability across extreme interaction horizons (e.g., > 200 steps). Extensive experiments on 15+ leading LLMs reveal that even frontier models exhibit a deficiency in inductive scenarios, identifying a critical bottleneck in the pursuit of autonomous discovery in complex environments. Our code and data are available at https://github.com/xufangzhi/Odyssey-Arena
Temporal knowledge graph question answering (TKGQA) aims to answer time-sensitive questions by leveraging temporal knowledge bases. While Large Language Models (LLMs) demonstrate significant potential in TKGQA, current prompting strategies constrain their efficacy in two primary ways. First, they are prone to reasoning hallucinations under complex temporal constraints. Second, static prompting limits model autonomy and generalization, as it lack optimization through dynamic interaction with temporal knowledge graphs (TKGs) environments. To address these limitations, we propose \textbf{TKG-Thinker}, a novel agent equipped with autonomous planning and adaptive retrieval capabilities for reasoning over TKGs. Specifically, TKG-Thinker performs in-depth temporal reasoning through dynamic multi-turn interactions with TKGs via a dual-training strategy. We first apply Supervised Fine-Tuning (SFT) with chain of thought data to instill core planning capabilities, followed by a Reinforcement Learning (RL) stage that leverages multi-dimensional rewards to refine reasoning policies under intricate temporal constraints. Experimental results on benchmark datasets with three open-source LLMs show that TKG-Thinker achieves state-of-the-art performance and exhibits strong generalization across complex TKGQA settings.
Automated vulnerability reproduction from CVE descriptions requires generating executable Proof-of-Concept (PoC) exploits and validating them in target environments. This process is critical in software security research and practice, yet remains time-consuming and demands specialized expertise when performed manually. While LLM agents show promise for automating this task, existing approaches often conflate exploring attack directions with fixing implementation details, which leads to unproductive debugging loops when reproduction fails. To address this, we propose CVE2PoC, an LLM-based dual-loop agent framework following a plan-execute-evaluate paradigm. The Strategic Planner analyzes vulnerability semantics and target code to produce structured attack plans. The Tactical Executor generates PoC code and validates it through progressive verification. The Adaptive Refiner evaluates execution results and routes failures to different loops: the Tactical Loop for code-level refinement, while the Strategic Loop for attack strategy replanning. This dual-loop design enables the framework to escape ineffective debugging by matching remediation to failure type. Evaluation on two benchmarks covering 617 real-world vulnerabilities demonstrates that CVE2PoC achieves 82.9% and 54.3% reproduction success rates on SecBench.js and PatchEval, respectively, outperforming the best baseline by 11.3% and 20.4%. Human evaluation confirms that generated PoCs achieve comparable code quality to human-written exploits in readability and reusability.
Large language models (LLMs) achieve strong capabilities by scaling model capacity and training data, yet many real-world deployments rely on smaller models trained or adapted from low-resource data. This gap motivates the need for mechanisms to transfer knowledge from large, high-resource models to smaller, low-resource targets. While model merging provides an effective transfer mechanism, most existing approaches assume architecture-compatible models and therefore cannot directly transfer knowledge from large high-resource LLMs to heterogeneous low-resource targets. In this work, we propose a cross-architecture merging framework based on optimal transport (OT) that aligns activations to infer cross-neuron correspondences between heterogeneous models. The resulting transport plans are then used to guide direct weight-space fusion, enabling effective high-resource to low-resource transfer using only a small set of inputs. Extensive experiments across low-resource languages and specialized domains demonstrate consistent improvements over target models.