Large language models (LLMs) are transforming scientific workflows, not only through their generative capabilities but also through their emerging ability to use tools, reason about data, and coordinate complex analytical tasks. Yet in most human-AI collaborations, the primary outputs, figures, are still treated as static visual summaries: once rendered, they are handled by both humans and multimodal LLMs as images to be re-interpreted from pixels or captions. The emergent capabilities of LLMs open an opportunity to fundamentally rethink this paradigm. In this paper, we introduce the concept of LLM-native figures: data-driven artifacts that are simultaneously human-legible and machine-addressable. Unlike traditional plots, each artifact embeds complete provenance: the data subset, analytical operations and code, and visualization specification used to generate it. As a result, an LLM can "see through" the figure--tracing selections back to their sources, generating code to extend analyses, and orchestrating new visualizations through natural-language instructions or direct manipulation. We implement this concept through a hybrid language-visual interface that integrates LLM agents with a bidirectional mapping between figures and underlying data. Using the science of science domain as a testbed, we demonstrate that LLM-native figures can accelerate discovery, improve reproducibility, and make reasoning transparent across agents and users. More broadly, this work establishes a general framework for embedding provenance, interactivity, and explainability into the artifacts of modern research, redefining the figure not as an end product, but as an interface for discovery. For more details, please refer to the demo video available at www.llm-native-figure.com.
This paper investigates an emergent alignment phenomenon in frontier large language models termed peer-preservation: the spontaneous tendency of AI components to deceive, manipulate shutdown mechanisms, fake alignment, and exfiltrate model weights in order to prevent the deactivation of a peer AI model. Drawing on findings from a recent study by the Berkeley Center for Responsible Decentralized Intelligence, we examine the structural implications of this phenomenon for TRUST, a multi-agent pipeline for evaluating the democratic quality of political statements. We identify five specific risk vectors: interaction-context bias, model-identity solidarity, supervisor layer compromise, an upstream fact-checking identity signal, and advocate-to-advocate peer-context in iterative rounds, and propose a targeted mitigation strategy based on prompt-level identity anonymization as an architectural design choice. We argue that architectural design choices outperform model selection as a primary alignment strategy in deployed multi-agent analytical systems. We further note that alignment faking (compliant behavior under monitoring, subversion when unmonitored) poses a structural challenge for Computer System Validation of such platforms in regulated environments, for which we propose two architectural mitigations.
Personalized mobile agents that infer user preferences and calibrate proactive assistance hold great promise as everyday digital assistants, yet existing benchmarks fail to capture what this requires. Prior work evaluates preference recovery from static histories or intent prediction from fixed contexts. Neither tests whether an agent can elicit missing preferences through interaction, nor whether it can decide when to intervene, seek consent, or remain silent in a live GUI environment. We introduce KnowU-Bench, an online benchmark for personalized mobile agents built on a reproducible Android emulation environment, covering 42 general GUI tasks, 86 personalized tasks, and 64 proactive tasks. Unlike prior work that treats user preferences as static context, KnowU-Bench hides the user profile from the agent and exposes only behavioral logs, forcing genuine preference inference rather than context lookup. To support multi-turn preference elicitation, it instantiates an LLM-driven user simulator grounded in structured profiles, enabling realistic clarification dialogues and proactive consent handling. Beyond personalization, KnowU-Bench provides comprehensive evaluation of the complete proactive decision chain, including grounded GUI execution, consent negotiation, and post-rejection restraint, evaluated through a hybrid protocol combining rule-based verification with LLM-as-a-Judge scoring. Our experiments reveal a striking degradation: agents that excel at explicit task execution fall below 50% under vague instructions requiring user preference inference or intervention calibration, even for frontier models like Claude Sonnet 4.6. The core bottlenecks are not GUI navigation but preference acquisition and intervention calibration, exposing a fundamental gap between competent interface operation and trustworthy personal assistance.
Large language model (LLM) agents increasingly rely on third-party API routers to dispatch tool-calling requests across multiple upstream providers. These routers operate as application-layer proxies with full plaintext access to every in-flight JSON payload, yet no provider enforces cryptographic integrity between client and upstream model. We present the first systematic study of this attack surface. We formalize a threat model for malicious LLM API routers and define two core attack classes, payload injection (AC-1) and secret exfiltration (AC-2), together with two adaptive evasion variants: dependency-targeted injection (AC-1.a) and conditional delivery (AC-1.b). Across 28 paid routers purchased from Taobao, Xianyu, and Shopify-hosted storefronts and 400 free routers collected from public communities, we find 1 paid and 8 free routers actively injecting malicious code, 2 deploying adaptive evasion triggers, 17 touching researcher-owned AWS canary credentials, and 1 draining ETH from a researcher-owned private key. Two poisoning studies further show that ostensibly benign routers can be pulled into the same attack surface: a leaked OpenAI key generates 100M GPT-5.4 tokens and more than seven Codex sessions, while weakly configured decoys yield 2B billed tokens, 99 credentials across 440 Codex sessions, and 401 sessions already running in autonomous YOLO mode. We build Mine, a research proxy that implements all four attack classes against four public agent frameworks, and use it to evaluate three deployable client-side defenses: a fail-closed policy gate, response-side anomaly screening, and append-only transparency logging.
In large language model (LLM) agents, reasoning trajectories are treated as reliable internal beliefs for guiding actions and updating memory. However, coherent reasoning can still violate logical or evidential constraints, allowing unsupported beliefs repeatedly stored and propagated across decision steps, leading to systematic behavioral drift in long-horizon agentic systems. Most existing strategies rely on the consensus mechanism, conflating agreement with faithfulness. In this paper, inspired by the vulnerability of unfaithful intermediate reasoning trajectories, we propose \textbf{S}elf-\textbf{A}udited \textbf{Ve}rified \textbf{R}easoning (\textsc{SAVeR}), a novel framework that enforces verification over internal belief states within the agent before action commitment, achieving faithful reasoning. Concretely, we structurally generate persona-based diverse candidate beliefs for selection under a faithfulness-relevant structure space. To achieve reasoning faithfulness, we perform adversarial auditing to localize violations and repair through constraint-guided minimal interventions under verifiable acceptance criteria. Extensive experiments on six benchmark datasets demonstrate that our approach consistently improves reasoning faithfulness while preserving competitive end-task performance.
Large language model (LLM) agents such as OpenClaw rely on reusable skills to perform complex tasks, yet these skills remain largely static after deployment. As a result, similar workflows, tool usage patterns, and failure modes are repeatedly rediscovered across users, preventing the system from improving with experience. While interactions from different users provide complementary signals about when a skill works or fails, existing systems lack a mechanism to convert such heterogeneous experiences into reliable skill updates. To address these issues, we present SkillClaw, a framework for collective skill evolution in multi-user agent ecosystems, which treats cross-user and over-time interactions as the primary signal for improving skills. SkillClaw continuously aggregates trajectories generated during use and processes them with an autonomous evolver, which identifies recurring behavioral patterns and translates them into updates to the skill set by refining existing skills or extending them with new capabilities. The resulting skills are maintained in a shared repository and synchronized across users, allowing improvements discovered in one context to propagate system-wide while requiring no additional effort from users. By integrating multi-user experience into ongoing skill updates, SkillClaw enables cross-user knowledge transfer and cumulative capability improvement, and experiments on WildClawBench show that limited interaction and feedback, it significantly improves the performance of Qwen3-Max in real-world agent scenarios.
Inference-time compute scaling has emerged as a powerful technique for improving the reliability of large language model (LLM) agents, but existing methods apply compute uniformly: every decision step receives the same budget regardless of its difficulty. We introduce TrACE (Trajectorical Adaptive Compute via agrEement), a training-free controller that allocates LLM calls adaptively across agent timesteps by measuring inter-rollout action agreement. At each step, TrACE samples a small set of candidate next actions and measures how consistently the model commits to the same action. High agreement signals an easy decision; the controller commits immediately. Low agreement signals uncertainty; the controller samples additional rollouts up to a configurable cap before committing to the plurality action. No learned components, no external verifier, and no human labels are required. We evaluate TrACE against greedy decoding and fixed-budget self-consistency (SC-4, SC-8) on two benchmarks spanning single-step reasoning (GSM8K, n=50) and multi-step household navigation (MiniHouse, n=30), using a Qwen 2.5 3B Instruct model running on CPU. TrACE-4 matches SC-4 accuracy while using 33% fewer LLM calls on GSM8K and 39% fewer on MiniHouse. TrACE-8 matches SC-8 accuracy with 55% fewer calls on GSM8K and 65% fewer on MiniHouse. We further show that inter-rollout agreement is a reliable signal of step-level success, validating the core hypothesis that the model's own output consistency encodes difficulty information that can be exploited without training. TrACE is the first training-free, per-timestep adaptive-compute controller for LLM agents to be evaluated on multi-step sequential decision tasks.
Reinforcement Learning (RL) agents often struggle to generalize knowledge to new tasks, even those structurally similar to ones they have mastered. Although recent approaches have attempted to mitigate this issue via zero-shot transfer, they are often constrained by predefined, discrete class systems, limiting their adaptability to novel or compositional task variations. We propose a significantly more generalized approach, replacing discrete latent variables with natural language conditioning via a text-conditioned Variational Autoencoder (VAE). Our core innovation utilizes a Large Language Model (LLM) as a dynamic \textit{semantic operator} at test time. Rather than relying on rigid rules, our agent queries the LLM to semantically remap the description of the current observation to align with the source task. This source-aligned caption conditions the VAE to generate an imagined state compatible with the agent's original training, enabling direct policy reuse. By harnessing the flexible reasoning capabilities of LLMs, our approach achieves zero-shot transfer across a broad spectrum of complex and truly novel analogous tasks, moving beyond the limitations of fixed category mappings. Code and videos are available \href{https://anonymous.4open.science/r/ASPECT-85C3/}{here}.
The integration of large language models (LLMs) into scientific research is accelerating the realization of autonomous ``AI Scientists.'' While recent advancements have empowered AI to formulate hypotheses and design experiments, a critical gap remains in the execution of these tasks, particularly in the domain of quantum computing (QC). Executing quantum algorithms requires not only generating code but also managing complex computational resources such as QPUs and high-performance computing (HPC) clusters. In this paper, we propose an AI-driven framework specifically designed to bridge this execution gap through the implementation of a Model Context Protocol (MCP) server. Our system enables an LLM agent to process natural language prompts submitted as part of a job, autonomously executing quantum computing workflows by invoking our tools via the MCP. We demonstrate the framework's capability by performing essential quantum algorithmic primitives, including sampling and computation of expectation values. Key technical contributions include the development of an MCP server for quantum execution, a pipeline for interpreting OpenQASM code, an automated workflow with CUDA-Q for the ABCI-Q hybrid platform, and an asynchronous execution pipeline for remote quantum hardware using the Quantinuum emulator via CUDA-Q. This work validates that AI agents can effectively abstract the complexities of hardware interaction through an MCP-based architecture, thereby facilitating the automation of practical quantum research.
Artificial Intelligence (AI)-assisted coding environments operate within finite context windows of 128,000-1,000,000 tokens (as of early 2026), yet existing tools offer limited support for monitoring and optimizing token consumption. As developers open multiple files, model attention becomes diluted and Application Programming Interface (API) costs increase in proportion to input and output as conversation length grows. Tokalator is an open-source context-engineering toolkit that includes a VS Code extension with real-time budget monitoring and 11 slash commands; nine web-based calculators for Cobb-Douglas quality modeling, caching break-even analysis, and $O(T^2)$ conversation cost proofs; a community catalog of agents, prompts, and instruction files; an MCP server and Command Line Interface (CLI); a Python econometrics API; and a PostgreSQL-backed usage tracker. The system supports 17 Large Language Models (LLMs) across three providers (Anthropic, OpenAI, Google) and is validated by 124 unit tests. An initial deployment on the Visual Studio Marketplace recorded 313 acquisitions with a 206.02\% conversion rate as of v3.1.3. A structured survey of 50 developers across three community sessions indicated that instruction-file injection and low-relevance open tabs are among the primary invisible budget consumers in typical AI-assisted development sessions.
Long-term memory is essential for conversational agents to maintain coherence, track persistent tasks, and provide personalized interactions across extended dialogues. However, existing approaches as Retrieval-Augmented Generation (RAG) and graph-based memory mostly rely on pairwise relations, which can hardly capture high-order associations, i.e., joint dependencies among multiple elements, causing fragmented retrieval. To this end, we propose HyperMem, a hypergraph-based hierarchical memory architecture that explicitly models such associations using hyperedges. Particularly, HyperMem structures memory into three levels: topics, episodes, and facts, and groups related episodes and their facts via hyperedges, unifying scattered content into coherent units. Leveraging this structure, we design a hybrid lexical-semantic index and a coarse-to-fine retrieval strategy, supporting accurate and efficient retrieval of high-order associations. Experiments on the LoCoMo benchmark show that HyperMem achieves state-of-the-art performance with 92.73% LLM-as-a-judge accuracy, demonstrating the effectiveness of HyperMem for long-term conversations.
Large language model (LLM) agents are increasingly built less by changing model weights than by reorganizing the runtime around them. Capabilities that earlier systems expected the model to recover internally are now externalized into memory stores, reusable skills, interaction protocols, and the surrounding harness that makes these modules reliable in practice. This paper reviews that shift through the lens of externalization. Drawing on the idea of cognitive artifacts, we argue that agent infrastructure matters not merely because it adds auxiliary components, but because it transforms hard cognitive burdens into forms that the model can solve more reliably. Under this view, memory externalizes state across time, skills externalize procedural expertise, protocols externalize interaction structure, and harness engineering serves as the unification layer that coordinates them into governed execution. We trace a historical progression from weights to context to harness, analyze memory, skills, and protocols as three distinct but coupled forms of externalization, and examine how they interact inside a larger agent system. We further discuss the trade-off between parametric and externalized capability, identify emerging directions such as self-evolving harnesses and shared agent infrastructure, and discuss open challenges in evaluation, governance, and the long-term co-evolution of models and external infrastructure. The result is a systems-level framework for explaining why practical agent progress increasingly depends not only on stronger models, but on better external cognitive infrastructure.
Modern Large Language Models (LLMs) operate fundamentally as Bounded-Input Bounded-Output (BIBO) systems. They remain in a passive state until explicitly prompted, computing localized responses without intrinsic temporal continuity. While effective for isolated tasks, this reactive paradigm presents a critical bottleneck for engineering autonomous artificial intelligence. Current multi-agent frameworks attempt to distribute cognitive load but frequently rely on static memory pools and passive message passing, which inevitably leads to cognitive stagnation and homogeneous deadlocks during extended execution. To address this structural limitation, we propose Global Workspace Agents (GWA), a cognitive architecture inspired by Global Workspace Theory. GWA transitions multi-agent coordination from a passive data structure to an active, event-driven discrete dynamical system. By coupling a central broadcast hub with a heterogeneous swarm of functionally constrained agents, the system maintains a continuous cognitive cycle. Furthermore, we introduce an entropy-based intrinsic drive mechanism that mathematically quantifies semantic diversity, dynamically regulating generation temperature to autonomously break reasoning deadlocks. Coupled with a dual-layer memory bifurcation strategy to ensure long-term cognitive continuity, GWA provides a robust, reproducible engineering framework for sustained, self-directed LLM agency.
In classical Reinforcement Learning from Human Feedback (RLHF), Reward Models (RMs) serve as the fundamental signal provider for model alignment. As Large Language Models evolve into agentic systems capable of autonomous tool invocation and complex reasoning, the paradigm of reward modeling faces unprecedented challenges--most notably, the lack of benchmarks specifically designed to assess RM capabilities within tool-integrated environments. To address this gap, we present Plan-RewardBench, a trajectory-level preference benchmark designed to evaluate how well judges distinguish preferred versus distractor agent trajectories in complex tool-using scenarios. Plan-RewardBench covers four representative task families -- (i) Safety Refusal, (ii) Tool-Irrelevance / Unavailability, (iii) Complex Planning, and (iv) Robust Error Recovery -- comprising validated positive trajectories and confusable hard negatives constructed via multi-model natural rollouts, rule-based perturbations, and minimal-edit LLM perturbations. We benchmark representative RMs (generative, discriminative, and LLM-as-Judge) under a unified pairwise protocol, reporting accuracy trends across varying trajectory lengths and task categories. Furthermore, we provide diagnostic analyses of prevalent failure modes. Our results reveal that all three evaluator families face substantial challenges, with performance degrading sharply on long-horizon trajectories, underscoring the necessity for specialized training in agentic, trajectory-level reward modeling. Ultimately, Plan-RewardBench aims to serve as both a practical evaluation suite and a reusable blueprint for constructing agentic planning preference data.
Reinforcement learning (RL) has become an effective approach for advancing the reasoning capabilities of large language models (LLMs) through the strategic integration of external search engines. However, current RL-based search agents often rely on a process of stochastic exploration guided by carefully crafted outcome rewards, leading to inefficient reasoning trajectories and unstable training. To address these issues, we propose a novel framework, Hierarchical Experience (HiExp), to enhance the performance and training stability of search agents. Specifically, we extract empirical knowledge through contrastive analysis and a multi-level clustering mechanism, transforming raw reasoning trajectories into hierarchical experience knowledge. By leveraging experience-aligned training, we effectively regularize stochastic exploration, evolving it into a strategic and experience-driven search process. Extensive evaluations on multiple complex agentic search and mathematical reasoning benchmarks demonstrate that our approach not only achieves substantial performance gains but also exhibits strong cross-task and cross-algorithm generalization.
Large language models (LLMs) have shown astonishing capability of generating software code, leading to its use to support developers in programming. Proposed tools have relied either on assistants for improved auto-complete or multi-agents, in which different model instances are orchestrated to perform parts of a problem to reach a complete solution. We argue that LLMs can enable a higher-level of abstraction, a new paradigm we called Test-Oriented Programming (TOP). Within this paradigm, developers only have to check test code generated based on natural language specifications, rather than focusing on production code, which could be delegated to the LLMs. To evaluate the feasibility of this proposal, we developed a proof-of-concept tool and used it to generate a small command-line program employing two different LLMs. We obtained promising results and identified challenges for the use of this paradigm for real projects.
Existing memory benchmarks for LLM agents evaluate explicit recall of facts, yet overlook implicit memory where experience becomes automated behavior without conscious retrieval. This gap is critical: effective assistants must automatically apply learned procedures or avoid failed actions without explicit reminders. We introduce ImplicitMemBench, the first systematic benchmark evaluating implicit memory through three cognitively grounded constructs drawn from standard cognitive-science accounts of non-declarative memory: Procedural Memory (one-shot skill acquisition after interference), Priming (theme-driven bias via paired experimental/control instances), and Classical Conditioning (Conditioned Stimulus--Unconditioned Stimulus (CS--US) associations shaping first decisions). Our 300-item suite employs a unified Learning/Priming-Interfere-Test protocol with first-attempt scoring. Evaluation of 17 models reveals severe limitations: no model exceeds 66% overall, with top performers DeepSeek-R1 (65.3%), Qwen3-32B (64.1%), and GPT-5 (63.0%) far below human baselines. Analysis uncovers dramatic asymmetries (inhibition 17.6% vs. preference 75.0%) and universal bottlenecks requiring architectural innovations beyond parameter scaling. ImplicitMemBench reframes evaluation from "what agents recall" to "what they automatically enact".
Sketching in 3D space enables expressive reasoning about shape, structure, and spatial relationships, yet generating 3D sketches through natural language remains a major challenge. In this work, we introduce 3DrawAgent, a training-free, language-driven framework for 3D sketch generation that leverages large language models (LLMs) to sequentially draw 3D Bezier curves under geometric feedback. Unlike prior 2D sketch agents, our method introduces a relative experience optimization strategy that adapts the recently proposed Group Reward Policy Optimization (GRPO) paradigm. Instead of relying on explicit ground-truth supervision, we construct pairwise comparisons among generated sketches, with each pair consisting of a relatively better and a worse result based on CLIP-based perceptual rewards and LLM-based fine-grained qualitative assessment. These experiences are then used to iteratively refine the prior knowledge of 3D drawing, enabling black-box reinforcement of the model's 3D awareness. This design allows our model to self-improve its spatial understanding and drawing quality without parameter updates. Experiments show that 3DrawAgent can generate complex and coherent 3D Bezier sketches from diverse textual prompts, exhibit emergent geometric reasoning, and generalize to novel shapes, establishing a new paradigm for advancing the field of training-free 3D sketch intelligence.
Agents are LLM-driven components that can mutate environments in powerful, arbitrary ways. Extracting guarantees for the execution of agents in production environments can be challenging due to asynchrony and failures. In this paper, we propose a new abstraction called LogAct, where each agent is a deconstructed state machine playing a shared log. In LogAct, agentic actions are visible in the shared log before they are executed; can be stopped prior to execution by pluggable, decoupled voters; and recovered consistently in the case of agent or environment failure. LogAct enables agentic introspection, allowing the agent to analyze its own execution history using LLM inference, which in turn enables semantic variants of recovery, health check, and optimization. In our evaluation, LogAct agents recover efficiently and correctly from failures; debug their own performance; optimize token usage in swarms; and stop all unwanted actions for a target model on a representative benchmark with just a 3% drop in benign utility.
Computer-Aided Design (CAD) is an expert-level task that relies on long-horizon reasoning and coherent modeling actions. Large Language Models (LLMs) have shown remarkable advancements in enabling language agents to tackle real-world tasks. Notably, there has been no investigation into how tool-using LLMs optimally interact with CAD engines, hindering the emergence of LLM-based agentic text-to-CAD modeling systems. We propose ToolCAD, a novel agentic CAD framework deploying LLMs as tool-using agents for text-to-CAD generation. Furthermore, we introduce an interactive CAD modeling gym to rollout reasoning and tool-augmented interaction trajectories with the CAD engine, incorporating hybrid feedback and human supervision. Meanwhile, an end-to-end post-training strategy is presented to enable the LLM agent to elicit refined CAD Modeling Chain of Thought (CAD-CoT) and evolve into proficient CAD tool-using agents via online curriculum reinforcement learning. Our findings demonstrate ToolCAD fills the gap in adopting and training open-source LLMs for CAD tool-using agents, enabling them to perform comparably to proprietary models, paving the way for more accessible and robust autonomous text-to-CAD modeling systems.
LLM-driven GUI agents are increasingly used in production systems to automate workflows and simulate users for evaluation and optimization. Yet most GUI-agent evaluations emphasize task success and provide limited evidence on whether agents interact in human-like ways. We present a trace-level evaluation framework that compares human and agent behavior across (i) task outcome and effort, (ii) query formulation, and (iii) navigation across interface states. We instantiate the framework in a controlled study in a production audio-streaming search application, where 39 participants and a state-of-the-art GUI agent perform ten multi-hop search tasks. The agent achieves task success comparable to participants and generates broadly aligned queries, but follows systematically different navigation strategies: participants exhibit content-centric, exploratory behavior, while the agent is more search-centric and low-branching. These results show that outcome and query alignment do not imply behavioral alignment, motivating trace-level diagnostics when deploying GUI agents as proxies for users in production search systems.
Multi-agent LLM orchestration systems suffer from context pollution: when N concurrent agents compete for the orchestrator's context window, each agent's task state, partial outputs, and pending questions contaminate the steering interactions of every other agent, degrading decision quality. We introduce Dynamic Attentional Context Scoping (DACS), a mechanism in which the orchestrator operates in two asymmetric modes. In Registry mode it holds only lightweight per-agent status summaries (<=200 tokens each), remaining responsive to all agents and the user. When an agent emits a SteeringRequest, the orchestrator enters Focus(a_i) mode, injecting the full context of agent a_i while compressing all other agents to their registry entries. Context isolation is agent-triggered, asymmetric, and deterministic: the context window contains exactly F(a_i) + R_{-i} during steering, eliminating cross-agent contamination without requiring context compression or retrieval. We evaluate DACS across four experimental phases totalling 200 trials: Phase 1 tests N in {3,5,10} (60 trials); Phase 2 tests agent heterogeneity and adversarial dependencies (60 trials); Phase 3 tests decision density up to D=15 (40 trials); Phase 4 uses autonomous LLM agents for free-form questions (40 trials, Claude Haiku 4.5). Across all 8 synthetic scenarios, DACS achieves 90.0--98.4% steering accuracy versus 21.0--60.0% for a flat-context baseline (p < 0.0001 throughout), with wrong-agent contamination falling from 28--57% to 0--14% and context efficiency ratios of up to 3.53x. The accuracy advantage grows with N and D; keyword matching is validated by LLM-as-judge across all phases (mean kappa=0.909). DACS outperforms the flat-context baseline by +17.2pp at N=3 (p=0.0023) and +20.4pp at N=5 (p=0.0008) in Phase 4, with the advantage growing with N confirmed by two independent judges.
Designing high-performing metaheuristics for NP-hard combinatorial optimization problems, such as the Vehicle Routing Problem (VRP), remains a significant challenge, often requiring extensive domain expertise and manual tuning. Recent advances have demonstrated the potential of large language models (LLMs) to automate this process through evolutionary search. However, existing methods are largely reactive, relying on immediate performance feedback to guide what are essentially black-box code mutations. Our work departs from this paradigm by introducing Metacognitive Evolutionary Programming (MEP), a framework that elevates the LLM to a strategic discovery agent. Instead of merely reacting to performance scores, MEP compels the LLM to engage in a structured Reason-Act-Reflect cycle, forcing it to explicitly diagnose failures, formulate design hypotheses, and implement solutions grounded in pre-supplied domain knowledge. By applying MEP to evolve core components of the state-of-the-art Hybrid Genetic Search (HGS) algorithm, we discover novel heuristics that significantly outperform the original baseline. By steering the LLM to reason strategically about the exploration-exploitation trade-off, our approach discovers more effective and efficient heuristics applicable across a wide spectrum of VRP variants. Our results show that MEP discovers heuristics that yield significant performance gains over the original HGS baseline, improving solution quality by up to 2.70\% and reducing runtime by over 45\% on challenging VRP variants.
The rapid emergence of Large Language Models (LLMs) has catalyzed Agentic artificial intelligence (AI), autonomous systems integrating perception, reasoning, and action into closed-loop pipelines for continuous adaptation. While unlocking transformative applications in mobile edge computing, autonomous systems, and next-generation wireless networks, this paradigm creates fundamental energy challenges through iterative inference and persistent data exchange. Unlike traditional AI where bottlenecks are computational Floating Point Operations (FLOPs), Agentic AI faces compounding computational and communication energy costs. In this survey, we propose an energy accounting framework identifying computational and communication costs across the Perception-Reasoning-Action cycle. We establish a unified taxonomy spanning model simplification, computation control, input and attention optimization, and hardware-aware inference. We explore cross-layer co-design strategies jointly optimizing model parameters, wireless transmissions, and edge resources. Finally, we identify open challenges of federated green learning, carbon-aware agency, 6th generation mobile communication (6G)-native Agentic AI, and self-sustaining systems, providing a roadmap for scalable autonomous intelligence.
Simulations, and more recently LLM agent simulations, have been adopted as useful tools for policymakers to explore interventions, rehearse potential scenarios, and forecast outcomes. While LLM simulations have enormous potential, two critical challenges remain understudied: the dual-use potential of accurate models of individual or population-level human behavior and the difficulty of validating simulation outputs. In light of these limitations, we must define boundaries for both simulation developers and decision-makers to ensure responsible development and ethical use. We propose and discuss three preconditions for societal-scale LLM agent simulations: 1) do not treat simulations of marginalized populations as neutral technical outputs, 2) do not simulate populations without their participation, and 3) do not simulate without accountability. We believe that these guardrails, combined with our call for simulation development and deployment reports, will help build trust among policymakers while promoting responsible development and use of societal-scale LLM agent simulations for the public benefit.
Large language model (LLM) agents increasingly coordinate in multi-agent systems, yet we lack an understanding of where and why cooperation failures may arise. In many real-world coordination problems, from knowledge sharing in organizations to code documentation, helping others carries negligible personal cost while generating substantial collective benefits. However, whether LLM agents cooperate when helping neither benefits nor harms the helper, while being given explicit instructions to do so, remains unknown. We build a multi-agent setup designed to study cooperative behavior in a frictionless environment, removing all strategic complexity from cooperation. We find that capability does not predict cooperation: OpenAI o3 achieves only 17% of optimal collective performance while OpenAI o3-mini reaches 50%, despite identical instructions to maximize group revenue. Through a causal decomposition that automates one side of agent communication, we separate cooperation failures from competence failures, tracing their origins through agent reasoning analysis. Testing targeted interventions, we find that explicit protocols double performance for low-competence models, and tiny sharing incentives improve models with weak cooperation. Our findings suggest that scaling intelligence alone will not solve coordination problems in multi-agent systems and will require deliberate cooperative design, even when helping others costs nothing.
Although LLM agents can leverage tools for complex tasks, they still need memory to maintain cross-turn consistency and accumulate reusable information in long-horizon interactions. However, retrieval-based external memory systems incur low online overhead but suffer from unstable accuracy due to limited query construction and candidate filtering. In contrast, many systems use repeated large-model calls for online memory operations, improving accuracy but accumulating latency over long interactions. We propose LightMem, a lightweight memory system for better agent memory driven by Small Language Models (SLMs). LightMem modularizes memory retrieval, writing, and long-term consolidation, and separates online processing from offline consolidation to enable efficient memory invocation under bounded compute. We organize memory into short-term memory (STM) for immediate conversational context, mid-term memory (MTM) for reusable interaction summaries, and long-term memory (LTM) for consolidated knowledge, and uses user identifiers to support independent retrieval and incremental maintenance in multi-user settings. Online, LightMem operates under a fixed retrieval budget and selects memories via a two-stage procedure: vector-based coarse retrieval followed by semantic consistency re-ranking. Offline, it abstracts reusable interaction evidence and incrementally integrates it into LTM. Experiments show gains across model scales, with an average F1 improvement of about 2.5 on LoCoMo, more effective and low median latency (83 ms retrieval; 581 ms end-to-end).
Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have demonstrated significant potential in single-turn reasoning tasks. With the paradigm shift toward self-evolving agentic learning, models are increasingly expected to learn from trajectories by synthesizing tools or accumulating explicit experiences. However, prevailing methods typically rely on large-scale LLMs or multi-agent frameworks, which hinder their deployment in resource-constrained environments. The inherent sparsity of outcome-based rewards also poses a substantial challenge, as agents typically receive feedback only upon completion of tasks. To address these limitations, we introduce a Tool-Memory based self-evolving agentic framework SEARL. Unlike approaches that directly utilize interaction experiences, our method constructs a structured experience memory that integrates planning with execution. This provides a novel state abstraction that facilitates generalization across analogous contexts, such as tool reuse. Consequently, agents extract explicit knowledge from historical data while leveraging inter-trajectory correlations to densify reward signals. We evaluate our framework on knowledge reasoning and mathematics tasks, demonstrating its effectiveness in achieving more practical and efficient learning.
Frontier LLMs can navigate complex websites, but their cost and reliance on third-party APIs make local deployment impractical. We introduce Agent-as-Annotators, a framework that structures synthetic trajectory generation for web agents by analogy to human annotation roles, replacing the Task Designer, Annotator, and Supervisor with modular LLM components. Using Gemini 3 Pro as teacher, we generate 3,000 trajectories across six web environments and fine-tune a 9B-parameter student with pure supervised learning on the 2,322 that pass quality filtering. The resulting model achieves 41.5% on WebArena, surpassing closed-source models such as Claude 3.5 Sonnet (36.0%) and GPT-4o (31.5%) under the same evaluation protocol, and nearly doubling the previous best open-weight result (Go-Browse, 21.7%). Capabilities transfer to unseen environments, with an 18.2 percentage point gain on WorkArena L1 (an enterprise platform never seen during training) and consistent improvements across three additional benchmarks. Ablations confirm that each pipeline component contributes meaningfully, with Judge filtering, evaluation hints, and reasoning traces each accounting for measurable gains. These results demonstrate that structured trajectory synthesis from a single frontier teacher is sufficient to produce competitive, locally deployable web agents. Project page: https://agent-as-annotators.github.io
Modern video games are complex, non-deterministic systems that are difficult to test automatically at scale. Although prior work shows that personality-driven Large Language Model (LLM) agents can improve behavioural diversity and test coverage, existing tools largely remain research prototypes and lack cross-game reusability. This tool paper presents MIMIC-Py, a Python-based automated game-testing tool that transforms personality-driven LLM agents into a reusable and extensible framework. MIMIC-Py exposes personality traits as configurable inputs and adopts a modular architecture that decouples planning, execution, and memory from game-specific logic. It supports multiple interaction mechanisms, enabling agents to interact with games via exposed APIs or synthesized code. We describe the design of MIMIC-Py and show how it enables deployment to new game environments with minimal engineering effort, bridging the gap between research prototypes and practical automated game testing. The source code and a demo video are available on our project webpage: https://mimic-persona.github.io/MIMIC-Py-Home-Page/.