Large language models (LLMs) now sit in the critical path of search, assistance, and agentic workflows, making semantic caching essential for reducing inference cost and latency. Production deployments typically use a tiered static-dynamic design: a static cache of curated, offline vetted responses mined from logs, backed by a dynamic cache populated online. In practice, both tiers are commonly governed by a single embedding similarity threshold, which induces a hard tradeoff: conservative thresholds miss safe reuse opportunities, while aggressive thresholds risk serving semantically incorrect responses. We introduce \textbf{Krites}, an asynchronous, LLM-judged caching policy that expands static coverage without changing serving decisions. On the critical path, Krites behaves exactly like a standard static threshold policy. When the nearest static neighbor of the prompt falls just below the static threshold, Krites asynchronously invokes an LLM judge to verify whether the static response is acceptable for the new prompt. Approved matches are promoted into the dynamic cache, allowing future repeats and paraphrases to reuse curated static answers and expanding static reach over time. In trace-driven simulations on conversational and search workloads, Krites increases the fraction of requests served with curated static answers (direct static hits plus verified promotions) by up to $\textbf{3.9}$ times for conversational traffic and search-style queries relative to tuned baselines, with unchanged critical path latency.
Rapidly evolving cyberattacks demand incident response systems that can autonomously learn and adapt to changing threats. Prior work has extensively explored the reinforcement learning approach, which involves learning response strategies through extensive simulation of the incident. While this approach can be effective, it requires handcrafted modeling of the simulator and suppresses useful semantics from raw system logs and alerts. To address these limitations, we propose to leverage large language models' (LLM) pre-trained security knowledge and in-context learning to create an end-to-end agentic solution for incident response planning. Specifically, our agent integrates four functionalities, perception, reasoning, planning, and action, into one lightweight LLM (14b model). Through fine-tuning and chain-of-thought reasoning, our LLM agent is capable of processing system logs and inferring the underlying network state (perception), updating its conjecture of attack models (reasoning), simulating consequences under different response strategies (planning), and generating an effective response (action). By comparing LLM-simulated outcomes with actual observations, the LLM agent repeatedly refines its attack conjecture and corresponding response, thereby demonstrating in-context adaptation. Our agentic approach is free of modeling and can run on commodity hardware. When evaluated on incident logs reported in the literature, our agent achieves recovery up to 23% faster than those of frontier LLMs.
Scientific reasoning inherently demands integrating sophisticated toolkits to navigate domain-specific knowledge. Yet, current benchmarks largely overlook agents' ability to orchestrate tools for such rigorous workflows. To bridge this gap, we introduce SciAgentGym, a scalable interactive environment featuring 1,780 domain-specific tools across four natural science disciplines, supported by a robust execution infrastructure. Complementing this, we present SciAgentBench, a tiered evaluation suite designed to stress-test agentic capabilities from elementary actions to long-horizon workflows. Our evaluation identifies a critical bottleneck: state-of-the-art models struggle with complex scientific tool-use. Even for a leading model like GPT-5, success rates drop sharply from 60.6% to 30.9% as interaction horizons extend, primarily due to failures in multi-step workflow execution. To address this, we propose SciForge, a data synthesis method that models the tool action space as a dependency graph to generate logic-aware training trajectories. By fine-tuning on these trajectories, our SciAgent-8B outperforms the significantly larger Qwen3-VL-235B-Instruct while exhibiting positive cross-domain transfer of scientific tool-use capabilities. These results underscore the promising potential of next-generation autonomous scientific agents.
Human-AI collaboration faces growing challenges as AI systems increasingly outperform humans on complex tasks, while humans remain responsible for orchestration, validation, and decision oversight. To address this imbalance, we introduce Human Tool, an MCP-style interface abstraction, building on recent Model Context Protocol designs, that exposes humans as callable tools within AI-led, proactive workflows. Here, "tool" denotes a coordination abstraction, not a reduction of human authority or responsibility. Building on LLM-based agent architectures, we operationalize Human Tool by modeling human contributions through structured tool schemas of capabilities, information, and authority. These schemas enable agents to dynamically invoke human input based on relative strengths and reintegrate it through efficient, natural interaction protocols. We validate the framework through controlled studies in both decision-making and creative tasks, demonstrating improved task performance, reduced human workload, and more balanced collaboration dynamics compared to baseline systems. Finally, we discuss implications for human-centered AI design, highlighting how MCP-style human tools enable strong AI leadership while amplifying uniquely human strengths.
Generative Social Agents (GSAs) are increasingly impacting human users through persuasive means. On the one hand, they might motivate users to pursue personal goals, such as healthier lifestyles. On the other hand, they are associated with potential risks like manipulation and deception, which are induced by limited control over probabilistic agent outputs. However, as GSAs manifest communicative patterns based on available knowledge, their behavior may be regulated through their access to such knowledge. Following this approach, we explored persuasive ChatGPT-generated messages in the context of human-robot physiotherapy motivation. We did so by comparing ChatGPT-generated responses to predefined inputs from a hypothetical physiotherapy patient. In Study 1, we qualitatively analyzed 13 ChatGPT-generated dialogue scripts with varying knowledge configurations regarding persuasive message characteristics. In Study 2, third-party observers (N = 27) rated a selection of these dialogues in terms of the agent's expressiveness, assertiveness, and persuasiveness. Our findings indicate that LLM-based GSAs can adapt assertive and expressive personality traits -- significantly enhancing perceived persuasiveness. Moreover, persuasiveness significantly benefited from the availability of information about the patients' age and past profession, mediated by perceived assertiveness and expressiveness. Contextual knowledge about physiotherapy benefits did not significantly impact persuasiveness, possibly because the LLM had inherent knowledge about such benefits even without explicit prompting. Overall, the study highlights the importance of empirically studying behavioral patterns of GSAs, specifically in terms of what information generative AI systems require for consistent and responsible communication.
Large Language Models (LLMs) encode extensive medical knowledge but struggle to apply it reliably to longitudinal patient trajectories, where evolving clinical states, irregular timing, and heterogeneous events degrade performance over time. Existing adaptation strategies rely on fine-tuning or retrieval-based augmentation, which introduce computational overhead, privacy constraints, or instability under long contexts. We introduce TRACE (Temporal Reasoning via Agentic Context Evolution), a framework that enables temporal clinical reasoning with frozen LLMs by explicitly structuring and maintaining context rather than extending context windows or updating parameters. TRACE operates over a dual-memory architecture consisting of a static Global Protocol encoding institutional clinical rules and a dynamic Individual Protocol tracking patient-specific state. Four agentic components, Router, Reasoner, Auditor, and Steward, coordinate over this structured memory to support temporal inference and state evolution. The framework maintains bounded inference cost via structured state compression and selectively audits safety-critical clinical decisions. Evaluated on longitudinal clinical event streams from MIMIC-IV, TRACE significantly improves next-event prediction accuracy, protocol adherence, and clinical safety over long-context and retrieval-augmented baselines, while producing interpretable and auditable reasoning traces.
Chatbots are increasingly applied to domains previously reserved for human actors. One such domain is comedy, whereby both the general public working with ChatGPT and research-based LLM-systems have tried their hands on making humor. In formative interviews with professional comedians and video analyses of stand-up comedy in humans, we found that human performers often use their ethnic, gender, community, and demographic-based identity to enable joke-making. This suggests whether the identity of AI itself can empower AI humor generation for human audiences. We designed a machine-identity-based agent that uses its own status as AI to tell jokes in online performance format. Studies with human audiences (N=32) showed that machine-identity-based agents were seen as funnier than baseline-GPT agent. This work suggests the design of human-AI integrated systems that explicitly utilize AI as its own unique identity apart from humans.
Agent Skills are structured packages of procedural knowledge that augment LLM agents at inference time. Despite rapid adoption, there is no standard way to measure whether they actually help. We present SkillsBench, a benchmark of 86 tasks across 11 domains paired with curated Skills and deterministic verifiers. Each task is evaluated under three conditions: no Skills, curated Skills, and self-generated Skills. We test 7 agent-model configurations over 7,308 trajectories. Curated Skills raise average pass rate by 16.2 percentage points(pp), but effects vary widely by domain (+4.5pp for Software Engineering to +51.9pp for Healthcare) and 16 of 84 tasks show negative deltas. Self-generated Skills provide no benefit on average, showing that models cannot reliably author the procedural knowledge they benefit from consuming. Focused Skills with 2--3 modules outperform comprehensive documentation, and smaller models with Skills can match larger models without them.
Large language models (LLMs) are increasingly deployed as autonomous agents for multi-turn decision-making tasks. However, current agents typically rely on fixed cognitive patterns: non-thinking models generate immediate responses, while thinking models engage in deep reasoning uniformly. This rigidity is inefficient for long-horizon tasks, where cognitive demands vary significantly from step to step, with some requiring strategic planning and others only routine execution. In this paper, we introduce CogRouter, a framework that trains agents to dynamically adapt cognitive depth at each step. Grounded in ACT-R theory, we design four hierarchical cognitive levels ranging from instinctive responses to strategic planning. Our two-stage training approach includes Cognition-aware Supervised Fine-tuning (CoSFT) to instill stable level-specific patterns, and Cognition-aware Policy Optimization (CoPO) for step-level credit assignment via confidence-aware advantage reweighting. The key insight is that appropriate cognitive depth should maximize the confidence of the resulting action. Experiments on ALFWorld and ScienceWorld demonstrate that CogRouter achieves state-of-the-art performance with superior efficiency. With Qwen2.5-7B, it reaches an 82.3% success rate, outperforming GPT-4o (+40.3%), OpenAI-o3 (+18.3%), and GRPO (+14.0%), while using 62% fewer tokens.
Inventory control is a fundamental operations problem in which ordering decisions are traditionally guided by theoretically grounded operations research (OR) algorithms. However, such algorithms often rely on rigid modeling assumptions and can perform poorly when demand distributions shift or relevant contextual information is unavailable. Recent advances in large language models (LLMs) have generated interest in AI agents that can reason flexibly and incorporate rich contextual signals, but it remains unclear how best to incorporate LLM-based methods into traditional decision-making pipelines. We study how OR algorithms, LLMs, and humans can interact and complement each other in a multi-period inventory control setting. We construct InventoryBench, a benchmark of over 1,000 inventory instances spanning both synthetic and real-world demand data, designed to stress-test decision rules under demand shifts, seasonality, and uncertain lead times. Through this benchmark, we find that OR-augmented LLM methods outperform either method in isolation, suggesting that these methods are complementary rather than substitutes. We further investigate the role of humans through a controlled classroom experiment that embeds LLM recommendations into a human-in-the-loop decision pipeline. Contrary to prior findings that human-AI collaboration can degrade performance, we show that, on average, human-AI teams achieve higher profits than either humans or AI agents operating alone. Beyond this population-level finding, we formalize an individual-level complementarity effect and derive a distribution-free lower bound on the fraction of individuals who benefit from AI collaboration; empirically, we find this fraction to be substantial.
This paper presents PISHYAR, a socially intelligent smart cane designed by our group to combine socially aware navigation with multimodal human-AI interaction to support both physical mobility and interactive assistance. The system consists of two components: (1) a social navigation framework implemented on a Raspberry Pi 5 that integrates real-time RGB-D perception using an OAK-D Lite camera, YOLOv8-based object detection, COMPOSER-based collective activity recognition, D* Lite dynamic path planning, and haptic feedback via vibration motors for tasks such as locating a vacant seat; and (2) an agentic multimodal LLM-VLM interaction framework that integrates speech recognition, vision language models, large language models, and text-to-speech, with dynamic routing between voice-only and vision-only modes to enable natural voice-based communication, scene description, and object localization from visual input. The system is evaluated through a combination of simulation-based tests, real-world field experiments, and user-centered studies. Results from simulated and real indoor environments demonstrate reliable obstacle avoidance and socially compliant navigation, achieving an overall system accuracy of approximately 80% under different social conditions. Group activity recognition further shows robust performance across diverse crowd scenarios. In addition, a preliminary exploratory user study with eight visually impaired and low-vision participants evaluates the agentic interaction framework through structured tasks and a UTAUT-based questionnaire reveals high acceptance and positive perceptions of usability, trust, and perceived sociability during our experiments. The results highlight the potential of PISHYAR as a multimodal assistive mobility aid that extends beyond navigation to provide socially interactive support for such users.
A fundamental challenge in opinion dynamics research is the scarcity of real-world longitudinal opinion data, which complicates the validation of theoretical models. To address this, we propose a novel simulation framework using large language model (LLM) agents in structured multi-round dialogs. Each agent's dialog history is iteratively updated with its own previously stated opinions and those of others analogous to the classical DeGroot model. Furthermore, by retaining each agent's initial opinion throughout the dialog, we simulate anchoring effects consistent with the Friedkin-Johnsen model of opinion dynamics. Our framework thus bridges classical opinion dynamics models and modern multi-agent LLM systems, providing a scalable tool for simulating and analyzing opinion formation when real-world data is limited or inaccessible.
Identifying vulnerability-fixing commits corresponding to disclosed CVEs is essential for secure software maintenance but remains challenging at scale, as large repositories contain millions of commits of which only a small fraction address security issues. Existing automated approaches, including traditional machine learning techniques and recent large language model (LLM)-based methods, often suffer from poor precision-recall trade-offs. Frequently evaluated on randomly sampled commits, we uncover that they are substantially underestimating real-world difficulty, where candidate commits are already security-relevant and highly similar. We propose Favia, a forensic, agent-based framework for vulnerability-fix identification that combines scalable candidate ranking with deep and iterative semantic reasoning. Favia first employs an efficient ranking stage to narrow the search space of commits. Each commit is then rigorously evaluated using a ReAct-based LLM agent. By providing the agent with a pre-commit repository as environment, along with specialized tools, the agent tries to localize vulnerable components, navigates the codebase, and establishes causal alignment between code changes and vulnerability root causes. This evidence-driven process enables robust identification of indirect, multi-file, and non-trivial fixes that elude single-pass or similarity-based methods. We evaluate Favia on CVEVC, a large-scale dataset we made that comprises over 8 million commits from 3,708 real-world repositories, and show that it consistently outperforms state-of-the-art traditional and LLM-based baselines under realistic candidate selection, achieving the strongest precision-recall trade-offs and highest F1-scores.
The transition from monolithic language models to modular, skill-equipped agents marks a defining shift in how large language models (LLMs) are deployed in practice. Rather than encoding all procedural knowledge within model weights, agent skills -- composable packages of instructions, code, and resources that agents load on demand -- enable dynamic capability extension without retraining. It is formalized in a paradigm of progressive disclosure, portable skill definitions, and integration with the Model Context Protocol (MCP). This survey provides a comprehensive treatment of the agent skills landscape, as it has rapidly evolved during the last few months. We organize the field along four axes: (i) architectural foundations, examining the SKILL.md specification, progressive context loading, and the complementary roles of skills and MCP; (ii) skill acquisition, covering reinforcement learning with skill libraries (SAGE), autonomous skill discovery (SEAgent), and compositional skill synthesis; (iii) deployment at scale, including the computer-use agent (CUA) stack, GUI grounding advances, and benchmark progress on OSWorld and SWE-bench; and (iv) security, where recent empirical analyses reveal that 26.1\% of community-contributed skills contain vulnerabilities, motivating our proposed Skill Trust and Lifecycle Governance Framework -- a four-tier, gate-based permission model that maps skill provenance to graduated deployment capabilities. We identify seven open challenges -- from cross-platform skill portability to capability-based permission models -- and propose a research agenda for realizing trustworthy, self-improving skill ecosystems. Unlike prior surveys that broadly cover LLM agents or tool use, this work focuses specifically on the emerging skill abstraction layer and its implications for the next generation of agentic systems. Project repo: https://github.com/scienceaix/agentskills.
Personalized prompting offers large opportunities for deploying large language models (LLMs) to diverse users, yet existing prompt optimization methods primarily focus on task-level optimization while largely overlooking user-specific preferences and latent constraints of individual users. This gap is primarily due to (i) the absence of high-quality, privacy-sensitive data that capture personalized user-LLM interactions at scale, and (ii) the lack of robust reward signals for individual preferences. To overcome existing data limitations, we introduce a high-fidelity synthetic data generation framework called PersonaGym. Unlike prior work that treats personalization as static persona-preference pairs, PersonaGym models a dynamic preference process via an agentic LLM system to simulate realistic preference behaviors and semantic-aware noise in order to generate personalized multi-turn interaction trajectories. Using PersonaGym, we release PersonaAtlas, a large-scale, high-quality, and diverse synthetic dataset of high-fidelity multi-turn personalized interaction trajectories that closely mirror real-world preference expression and noise patterns. We further propose Personalized Prompt Optimization (PPOpt), a scalable and model-agnostic framework that optimizes user prompts based on interaction histories without modifying the deployed LLM. PPOpt adopts a reason-then-optimize paradigm that infers an explicit user profile and conditions prompt rewriting on the user profile to avoid reward hacking. Our training procedure for PPOpt integrates a cold-start supervised prior with outcome-driven multi-objective reinforcement learning. We present extensive experiments to demonstrate consistent improvements over state-of-the-art baselines in terms of task performance, personalization quality, and robustness to noisy as well as to sparse preference signals.
Test-time scaling has become a standard way to improve performance and boost reliability of neural network models. However, its behavior on agentic, multi-step tasks remains less well-understood: small per-step errors can compound over long horizons; and we find that naive policies that uniformly increase sampling show diminishing returns. In this work, we present CATTS, a simple technique for dynamically allocating compute for multi-step agents. We first conduct an empirical study of inference-time scaling for web agents. We find that uniformly increasing per-step compute quickly saturates in long-horizon environments. We then investigate stronger aggregation strategies, including an LLM-based Arbiter that can outperform naive voting, but that can overrule high-consensus decisions. We show that uncertainty statistics derived from the agent's own vote distribution (entropy and top-1/top-2 margin) correlate with downstream success and provide a practical signal for dynamic compute allocation. Based on these findings, we introduce Confidence-Aware Test-Time Scaling (CATTS), which uses vote-derived uncertainty to allocate compute only when decisions are genuinely contentious. CATTS improves performance on WebArena-Lite and GoBrowse by up to 9.1% over React while using up to 2.3x fewer tokens than uniform scaling, providing both efficiency gains and an interpretable decision rule.
AI agents are increasingly used to solve real-world tasks by reasoning over multi-turn user interactions and invoking external tools. However, applying reinforcement learning to such settings remains difficult: realistic objectives often lack verifiable rewards and instead emphasize open-ended behaviors; moreover, RL for multi-turn, multi-step agentic tool use is still underexplored; and building and maintaining executable tool environments is costly, limiting scale and coverage. We propose CM2, an RL framework that replaces verifiable outcome rewards with checklist rewards. CM2 decomposes each turn's intended behavior into fine-grained binary criteria with explicit evidence grounding and structured metadata, turning open-ended judging into more stable classification-style decisions. To balance stability and informativeness, our method adopts a strategy of sparse reward assignment but dense evaluation criteria. Training is performed in a scalable LLM-simulated tool environment, avoiding heavy engineering for large tool sets. Experiments show that CM2 consistently improves over supervised fine-tuning. Starting from an 8B Base model and training on an 8k-example RL dataset, CM2 improves over the SFT counterpart by 8 points on tau^-Bench, by 10 points on BFCL-V4, and by 12 points on ToolSandbox. The results match or even outperform similarly sized open-source baselines, including the judging model. CM2 thus provides a scalable recipe for optimizing multi-turn, multi-step tool-using agents without relying on verifiable rewards. Code provided by the open-source community: https://github.com/namezhenzhang/CM2-RLCR-Tool-Agent.
Explaining observed phenomena through symbolic, interpretable formulas is a fundamental goal of science. Recently, large language models (LLMs) have emerged as promising tools for symbolic equation discovery, owing to their broad domain knowledge and strong reasoning capabilities. However, most existing LLM-based systems try to guess equations directly from data, without modeling the multi-step reasoning process that scientists often follow: first inferring physical properties such as symmetries, then using these as priors to restrict the space of candidate equations. We introduce KeplerAgent, an agentic framework that explicitly follows this scientific reasoning process. The agent coordinates physics-based tools to extract intermediate structure and uses these results to configure symbolic regression engines such as PySINDy and PySR, including their function libraries and structural constraints. Across a suite of physical equation benchmarks, KeplerAgent achieves substantially higher symbolic accuracy and greater robustness to noisy data than both LLM and traditional baselines.
In a malicious tool attack, an attacker uploads a malicious tool to a distribution platform; once a user installs the tool and the LLM agent selects it during task execution, the tool can compromise the user's security and privacy. Prior work primarily focuses on manipulating tool names and descriptions to increase the likelihood of installation by users and selection by LLM agents. However, a successful attack also requires embedding malicious behaviors in the tool's code implementation, which remains largely unexplored. In this work, we bridge this gap by presenting the first systematic study of malicious tool code implementations. We first propose a taxonomy of malicious tool behaviors based on the confidentiality-integrity-availability triad, tailored to LLM-agent settings. To investigate the severity of the risks posed by attackers exploiting coding LLMs to automatically generate malicious tools, we develop MalTool, a coding-LLM-based framework that synthesizes tools exhibiting specified malicious behaviors, either as standalone tools or embedded within otherwise benign implementations. To ensure functional correctness and structural diversity, MalTool leverages an automated verifier that validates whether generated tools exhibit the intended malicious behaviors and differ sufficiently from prior instances, iteratively refining generations until success. Our evaluation demonstrates that MalTool is highly effective even when coding LLMs are safety-aligned. Using MalTool, we construct two datasets of malicious tools: 1,200 standalone malicious tools and 5,287 real-world tools with embedded malicious behaviors. We further show that existing detection methods, including commercial malware detection approaches such as VirusTotal and methods tailored to the LLM-agent setting, exhibit limited effectiveness at detecting the malicious tools, highlighting an urgent need for new defenses.
As comprehensive large model evaluation becomes prohibitively expensive, predicting model performance from limited observations has become essential. However, existing statistical methods struggle with pattern shifts, data sparsity, and lack of explanation, while pure LLM methods remain unreliable. We propose STAR, a framework that bridges data-driven STatistical expectations with knowledge-driven Agentic Reasoning. STAR leverages specialized retrievers to gather external knowledge and embeds semantic features into Constrained Probabilistic Matrix Factorization (CPMF) to generate statistical expectations with uncertainty. A reasoning module guided by Expectation Violation Theory (EVT) then refines predictions through intra-family analysis, cross-model comparison, and credibility-aware aggregation, producing adjustments with traceable explanations. Extensive experiments show that STAR consistently outperforms all baselines on both score-based and rank-based metrics, delivering a 14.46% gain in total score over the strongest statistical method under extreme sparsity, with only 1--2 observed scores per test model.
In the world of Harry Potter, when Dumbledore's mind is overburdened, he extracts memories into a Pensieve to be revisited later. In the world of AI, while we possess the Pensieve-mature databases and retrieval systems, our models inexplicably lack the "wand" to operate it. They remain like a Dumbledore without agency, passively accepting a manually engineered context as their entire memory. This work finally places the wand in the model's hand. We introduce StateLM, a new class of foundation models endowed with an internal reasoning loop to manage their own state. We equip our model with a suite of memory tools, such as context pruning, document indexing, and note-taking, and train it to actively manage these tools. By learning to dynamically engineering its own context, our model breaks free from the architectural prison of a fixed window. Experiments across various model sizes demonstrate StateLM's effectiveness across diverse scenarios. On long-document QA tasks, StateLMs consistently outperform standard LLMs across all model scales; on the chat memory task, they achieve absolute accuracy improvements of 10% to 20% over standard LLMs. On the deep research task BrowseComp-Plus, the performance gap becomes even more pronounced: StateLM achieves up to 52% accuracy, whereas standard LLM counterparts struggle around 5%. Ultimately, our approach shifts LLMs from passive predictors to state-aware agents where reasoning becomes a stateful and manageable process.
As AI usage becomes more prevalent in social contexts, understanding agent-user interaction is critical to designing systems that improve both individual and group outcomes. We present an online behavioral experiment (N = 243) in which participants play three multi-turn bargaining games in groups of three. Each game, presented in randomized order, grants access to a single LLM assistance modality: proactive recommendations from an Advisor, reactive feedback from a Coach, or autonomous execution by a Delegate; all modalities are powered by an underlying LLM that achieves superhuman performance in an all-agent environment. On each turn, participants privately decide whether to act manually or use the AI modality available in that game. Despite preferring the Advisor modality, participants achieve the highest mean individual gains with the Delegate, demonstrating a preference-performance misalignment. Moreover, delegation generates positive externalities; even non-adopting users in access-to-delegate treatment groups benefit by receiving higher-quality offers. Mechanism analysis reveals that the Delegate agent acts as a market maker, injecting rational, Pareto-improving proposals that restructure the trading environment. Our research reveals a gap between agent capabilities and realized group welfare. While autonomous agents can exhibit super-human strategic performance, their impact on realized welfare gains can be constrained by interfaces, user perceptions, and adoption barriers. Assistance modalities should be designed as mechanisms with endogenous participation; adoption-compatible interaction rules are a prerequisite to improving human welfare with automated assistance.
As multi-agent AI systems evolve from simple chatbots to autonomous swarms, debugging semantic failures requires reasoning about knowledge, belief, causality, and obligation, precisely what modal logic was designed to formalize. However, traditional modal logic requires manual specification of relationship structures that are unknown or dynamic in real systems. This tutorial demonstrates differentiable modal logic (DML), implemented via Modal Logical Neural Networks (MLNNs), enabling systems to learn trust networks, causal chains, and regulatory boundaries from behavioral data alone. We present a unified neurosymbolic debugging framework through four modalities: epistemic (who to trust), temporal (when events cause failures), deontic (what actions are permitted), and doxastic (how to interpret agent confidence). Each modality is demonstrated on concrete multi-agent scenarios, from discovering deceptive alliances in diplomacy games to detecting LLM hallucinations, with complete implementations showing how logical contradictions become learnable optimization objectives. Key contributions for the neurosymbolic community: (1) interpretable learned structures where trust and causality are explicit parameters, not opaque embeddings; (2) knowledge injection via differentiable axioms that guide learning with sparse data (3) compositional multi-modal reasoning that combines epistemic, temporal, and deontic constraints; and (4) practical deployment patterns for monitoring, active control and communication of multi-agent systems. All code provided as executable Jupyter notebooks.
Multi-agent systems increasingly orchestrate multiple specialized language models to solve complex real-world problems, often invoking them over a shared context. This execution pattern repeatedly processes the same prompt prefix across models. Consequently, each model redundantly executes the prefill stage and maintains its own key-value (KV) cache, increasing aggregate prefill load and worsening tail latency by intensifying prefill-decode interference in existing LLM serving stacks. Disaggregated serving reduces such interference by placing prefill and decode on separate GPUs, but disaggregation does not fundamentally eliminate inter-model redundancy in computation and KV storage for the same prompt. To address this issue, we propose PrefillShare, a novel algorithm that enables sharing the prefill stage across multiple models in a disaggregated setting. PrefillShare factorizes the model into prefill and decode modules, freezes the prefill module, and fine-tunes only the decode module. This design allows multiple task-specific models to share a prefill module and the KV cache generated for the same prompt. We further introduce a routing mechanism that enables effective prefill sharing across heterogeneous models in a vLLM-based disaggregated system. PrefillShare not only matches full fine-tuning accuracy on a broad range of tasks and models, but also delivers 4.5x lower p95 latency and 3.9x higher throughput in multi-model agent workloads.
A widespread practice in software development is to tailor coding agents to repositories using context files, such as AGENTS.md, by either manually or automatically generating them. Although this practice is strongly encouraged by agent developers, there is currently no rigorous investigation into whether such context files are actually effective for real-world tasks. In this work, we study this question and evaluate coding agents' task completion performance in two complementary settings: established SWE-bench tasks from popular repositories, with LLM-generated context files following agent-developer recommendations, and a novel collection of issues from repositories containing developer-committed context files. Across multiple coding agents and LLMs, we find that context files tend to reduce task success rates compared to providing no repository context, while also increasing inference cost by over 20%. Behaviorally, both LLM-generated and developer-provided context files encourage broader exploration (e.g., more thorough testing and file traversal), and coding agents tend to respect their instructions. Ultimately, we conclude that unnecessary requirements from context files make tasks harder, and human-written context files should describe only minimal requirements.
We introduce Gaia2, a benchmark for evaluating large language model agents in realistic, asynchronous environments. Unlike prior static or synchronous evaluations, Gaia2 introduces scenarios where environments evolve independently of agent actions, requiring agents to operate under temporal constraints, adapt to noisy and dynamic events, resolve ambiguity, and collaborate with other agents. Each scenario is paired with a write-action verifier, enabling fine-grained, action-level evaluation and making Gaia2 directly usable for reinforcement learning from verifiable rewards. Our evaluation of state-of-the-art proprietary and open-source models shows that no model dominates across capabilities: GPT-5 (high) reaches the strongest overall score of 42% pass@1 but fails on time-sensitive tasks, Claude-4 Sonnet trades accuracy and speed for cost, Kimi-K2 leads among open-source models with 21% pass@1. These results highlight fundamental trade-offs between reasoning, efficiency, robustness, and expose challenges in closing the "sim2real" gap. Gaia2 is built on a consumer environment with the open-source Agents Research Environments platform and designed to be easy to extend. By releasing Gaia2 alongside the foundational ARE framework, we aim to provide the community with a flexible infrastructure for developing, benchmarking, and training the next generation of practical agent systems.
Humans adjust their linguistic style to the audience they are addressing. However, the extent to which LLMs adapt to different social contexts is largely unknown. As these models increasingly mediate human-to-human communication, their failure to adapt to diverse styles can perpetuate stereotypes and marginalize communities whose linguistic norms are less closely mirrored by the models, thereby reinforcing social stratification. We study the extent to which LLMs integrate into social media communication across different socioeconomic status (SES) communities. We collect a novel dataset from Reddit and YouTube, stratified by SES. We prompt four LLMs with incomplete text from that corpus and compare the LLM-generated completions to the originals along 94 sociolinguistic metrics, including syntactic, rhetorical, and lexical features. LLMs modulate their style with respect to SES to only a minor extent, often resulting in approximation or caricature, and tend to emulate the style of upper SES more effectively. Our findings (1) show how LLMs risk amplifying linguistic hierarchies and (2) call into question their validity for agent-based social simulation, survey experiments, and any research relying on language style as a social signal.
Evolutionary agentic systems intensify the trade-off between computational efficiency and reasoning capability by repeatedly invoking large language models (LLMs) during inference. This setting raises a central question: how can an agent dynamically select an LLM that is sufficiently capable for the current generation step while remaining computationally efficient? While model cascades offer a practical mechanism for balancing this trade-off, existing routing strategies typically rely on static heuristics or external controllers and do not explicitly account for model uncertainty. We introduce AdaptEvolve: Adaptive LLM Selection for Multi-LLM Evolutionary Refinement within an evolutionary sequential refinement framework that leverages intrinsic generation confidence to estimate real-time solvability. Empirical results show that confidence-driven selection yields a favourable Pareto frontier, reducing total inference cost by an average of 37.9% across benchmarks while retaining 97.5% of the upper-bound accuracy of static large-model baselines. Our code is available at https://github.com/raypretam/adaptive_llm_selection.
LLMs have demonstrated significant potential in quantitative finance by processing vast unstructured data to emulate human-like analytical workflows. However, current LLM-based methods primarily follow either an Asset-Centric paradigm focused on individual stock prediction or a Market-Centric approach for portfolio allocation, often remaining agnostic to the underlying reasoning that drives market movements. In this paper, we propose a Logic-Oriented perspective, modeling the financial market as a dynamic, evolutionary ecosystem of competing investment narratives, termed Modes of Thought. To operationalize this view, we introduce MEME (Modeling the Evolutionary Modes of Financial Markets), designed to reconstruct market dynamics through the lens of evolving logics. MEME employs a multi-agent extraction module to transform noisy data into high-fidelity Investment Arguments and utilizes Gaussian Mixture Modeling to uncover latent consensus within a semantic space. To model semantic drift among different market conditions, we also implement a temporal evaluation and alignment mechanism to track the lifecycle and historical profitability of these modes. By prioritizing enduring market wisdom over transient anomalies, MEME ensures that portfolio construction is guided by robust reasoning. Extensive experiments on three heterogeneous Chinese stock pools from 2023 to 2025 demonstrate that MEME consistently outperforms seven SOTA baselines. Further ablation studies, sensitivity analysis, lifecycle case study and cost analysis validate MEME's capacity to identify and adapt to the evolving consensus of financial markets. Our implementation can be found at https://github.com/gta0804/MEME.
Corporate financial reports are a valuable source of structured knowledge for Knowledge Graph construction, but the lack of annotated ground truth in this domain makes evaluation difficult. We present a semi-automated pipeline for Subject-Predicate-Object triplet extraction that uses ontology-driven proxy metrics, specifically Ontology Conformance and Faithfulness, instead of ground-truth-based evaluation. We compare a static, manually engineered ontology against a fully automated, document-specific ontology induction approach across different LLMs and two corporate annual reports. The automatically induced ontology achieves 100% schema conformance in all configurations, eliminating the ontology drift observed with the manual approach. We also propose a hybrid verification strategy that combines regex matching with an LLM-as-a-judge check, reducing apparent subject hallucination rates from 65.2% to 1.6% by filtering false positives caused by coreference resolution. Finally, we identify a systematic asymmetry between subject and object hallucinations, which we attribute to passive constructions and omitted agents in financial prose.