Large Language Model Agents

Open platforms increasingly route tasks among heterogeneous LLM agents--differing in base model, scaffold, and tool stack--whose competence varies sharply by skill: an agent excellent at one skill may be useless at another. The standard reputation approach summarizes each agent by a single global trust score, but that scalar is the wrong object here, because routing every task to the globally most-trusted agent leaves the value of specialization unclaimed. We study skill-conditional trust R(i | k)--the trust to place in agent i for a task requiring skill k, rather than one score per agent--and pose three falsifiable questions: when is conditioning worth it, how much cross-skill evidence should be borrowed, and whether that borrowing is safe. A controlled phase-diagram analysis answers the first two: conditional trust wins only in a specific regime--high agent heterogeneity, sparse per-skill evidence, and correlated skills--and the coupling strength beta that buys this data efficiency is dual-use, because the same cross-skill borrowing is also a laundering channel. On a public benchmark of 14 genuinely heterogeneous AppWorld agents, real pools land inside the beneficial regime--a small but genuine gain, with the per-skill best agent genuinely changing across skills. We then show that an attacker with cheap evidence in one skill and none in a target skill hijacks the conditional router, driving routing regret from 0 to 0.94 on a pool our zero-cost Conditional Information Value Test (CIVT) rates GREEN--while the ungated trust verdict it contaminates reads -0.06 instead of the honest +0.19. A zero-evidence gate bounds the attack but does not eliminate it; we characterize the residual cost under an explicit budget. We do not claim Sybil-resistance--we quantify the trade-off.

Reliable operation of multi-agent large language model (LLM) systems depends on debugging long execution traces, where the few causally decisive events are buried in unstructured logs of messages, routes, memory writes, and tool calls. The standard tool is counterfactual replay (rewind, edit, and re-run the trajectory to measure each event's effect), but its cost grows linearly with the number of candidate events, making exhaustive replay infeasible at scale. We frame trace debugging as a knowledge-based decision-support problem. Each trace is compiled into a structured event knowledge graph over routing, memory, tool-use, uncertainty, and latent evidence, and a calibrated predictor decides where a scarce replay budget should be spent. We do not propose a new replay oracle; we propose a method to predict its results without paying the replay cost. We formulate zero-replay counterfactual-effect prediction: given a trace under a fixed budget, predict which events the oracle would mark high-effect before any replay is performed. BranchPoint-Latent is a lightweight predictor over observable, structural, uncertainty, and latent features of the knowledge graph. Calibrated against a deterministic replay oracle across 37 trace families, a single learning-to-rank gradient-boosted predictor raises per-trace localization (Branch Recall@5) from 0.73 to 0.93 on held-out families at zero oracle-replay cost. Rather than claiming universal dominance, we characterize when cheap graph centrality suffices and when learned evidence is necessary. The result is an auditable, cost-efficient decision-support system for AI-reliability debugging, positioned explicitly on the cost-accuracy frontier with reproducible artifacts.

Modern LLM-powered autonomous agents increasingly rely on rich user interface (UI) state observations to achieve reliable action grounding in complex digital environments. However, many deployments transmit the full UI state to remote inference servers even when most elements are irrelevant to the current task, which can leak sensitive but unnecessary context such as authentication codes, private notifications, and background application states. We propose MINIM, a trusted local broker that performs privacy-aware minimization on the client side before any observation leaves the device. Grounded in Contextual Integrity (CI), MINIM learns a dual-score representation for each UI element by predicting an inherent sensitivity score (s) and a task-conditioned necessity score (n). These scores drive a ternary disclosure policy that keeps essential elements, abstracts sensitive attributes when needed, and removes task-irrelevant content. We optimize a CI-aware objective that penalizes necessity errors more strongly on high-risk content, enabling aggressive pruning while preserving task-critical information. Experiments on real-world UI observations derived from WebArena show that MINIM substantially reduces task-irrelevant sensitive leakage while preserving task-critical semantic context and the interactive affordances required for reliable agent actions.

Exploratory question answering (EQA) over data lakes requires an LLM agent to discover relevant sources, analyze retrieved data, and adapt its actions based on intermediate results. End-to-end accuracy alone cannot distinguish failures in search, planning, data analysis, or the agent's Action Policy: its decisions about what to do next and when to submit an answer. We present SANA (Search Agent Navigation Ablation framework), a diagnostic ablation framework that transforms EQA tasks into runtime profiles containing gold source sequence, sanitized subquestions, and execution records. SANA uses these profiles to construct idealized search, planning, and data-analysis tools, allowing each component to be ablated; the residual gap is diagnostic evidence for policy failures. To illustrate SANA as a reusable evaluation framework, we adapted two recent EQA benchmarks, LakeQA and KramaBench, and evaluated lightweight and mid-sized agents under fixed prompts, budgets, data lakes, and runtimes. Across both benchmarks, data analysis is a consistent bottleneck while planning is less so. Search is a major limitation in LakeQA's large data-lake setting, but less so for the smaller-scale KramaBench. SANA thus deconstructs end-to-end task accuracies into a diagnosis of where data-lake agents fail, and allows for systematic comparisons of progress in search, planning, data analysis, and agent design.

Large language model (LLM) agents have achieved strong performance on a wide range of benchmarks, yet most evaluations assume static environments. In contrast, real-world deployment is inherently dynamic, requiring agents to continually align their knowledge, skills, and behavior with changing environments and updated task conditions. To address this gap, we introduce EvoArena, a benchmark suite that models environment changes as sequences of progressive updates across terminal, software, and social domains. We further propose EvoMem, a patch-based memory paradigm that records memory evolution as structured update histories, enabling agents to reason about environmental evolution through changes in their memory. Experiments show that current agents struggle on EvoArena, achieving an average accuracy of 39.6% across evolving terminal, software, and social-preference domains. EvoMem consistently improves performance, yielding an average gain of 1.5% on EvoArena and also improving standard benchmarks such as GAIA and LoCoMo by 6.1% and 4.8%. Beyond individual tasks, EvoMem further improves chain-level accuracy by 3.7% on EvoArena, where success requires completing a consecutive sequence of related evolutionary subtasks. Mechanistic analysis shows that EvoMem improves evidence capture in the memory, indicating better preservation of complete evolving environment states. Our results highlight the importance of modeling evolution in both evaluation and memory for reliable agent deployment.

Skill self-evolution methods for LLM agents aim to turn execution trajectories into reusable skill documents, but current pipelines typically learn from one trajectory per task, merge candidate skill patches before checking them, and load the full skill corpus before inference. We propose SkillCAT, a training-free framework that separates this process into three stages. Contrastive Causal Extraction (CCE) samples multiple trajectories for each task and compares same-task success/failure pairs to identify evidence that explains outcome differences. Assessment-Augmented Evolution (AAE) replays each candidate patch on source-task clones and keeps only patches that improve or preserve task outcomes before hierarchical skill patch merging. Topology-Aware Task Execution (TTE) compiles the evolved skills into a routable sub-skill topology, so inference loads only the capability nodes relevant to the task. We evaluate SkillCAT on common agent benchmarks, including SpreadsheetBench, WikiTableQuestions, and DocVQA, and further test cross-model and out-of-distribution generalization. Across these settings, SkillCAT raises the average score over baselines by up to 40.40%, demonstrating reliable skill evolution without model training.

The original Turing Test asks a human judge to distinguish a machine from a person through dialogue. Three quarters of a century later, conversational systems pass this test in casual settings; the interesting epistemological question has shifted. We argue that the relevant modern variant asks not whether a dialogue partner is artificial, but whether it can be trusted. We present RogueAI, an interactive webapp that operationalizes this revisited test as a one-on-two interrogation game: a human player questions two indistinguishable Large Language Model agents, knowing that exactly one of them has been licensed to deceive within a shared fictional scenario. The player's task is to identify the deceptive agent and "shut it off" before a turn budget is exhausted. We further introduce AutoRogueAI, a procedural extension in which players co-design a custom scenario with a narrator agent that secretly chooses its own deception strategy. We describe the framing, sketch the abstract architecture and gameplay loop, and situate the artifact within recent work on LLM deception, social-deduction benchmarks, and scalable oversight via debate. A three-day pilot deployment (467 initiated sessions, 415 completed, 1876 interaction turns in Italian) provides early feasibility evidence and surfaces a concrete tension: the deceptive agent carries a reliable, locally-present linguistic signature - differential helpfulness, brevity, hedging - that a simple heuristic exploits at 75.6% accuracy, yet human players achieved only 56.6%, consistent with ignoring the most diagnostic signal entirely. We discuss what this gap implies for the artifact's use as a data-collection vehicle, a teaching tool, and an evaluation harness for honesty-trained models.

Language model agents are increasingly effective in solving realistic tasks through multi-turn tool use. However, training reliable tool-using agents remains challenging in practice. While reinforcement learning provides an on-policy paradigm for improving agents from their own environment interactions, its effectiveness depends heavily on the training task distribution. When tasks are fixed before training, the task distribution can become increasingly mismatched with the policy's evolving capabilities, causing many rollouts to be spent on uninformative tasks. We propose SENTINEL, a failure-driven reinforcement learning framework that turns the Solver's rollout failures into targeted training tasks. SENTINEL follows a Controller--Proposer--Solver loop: the Controller analyzes failed trajectories and summarizes recurring error patterns, the Proposer generates executable tasks that stress these weaknesses, and the Solver is trained on the targeted tasks. On Tau2-Bench Retail with Qwen3-4B-Thinking-2507, SENTINEL improves Pass^{}1 from 66.4 to 74.9 and outperforms RL on general synthetic tasks across Pass^{}k metrics. These results demonstrate that model failures provide an effective and scalable source of targeted training signal for improving tool-using language model agents.

Large language models are increasingly deployed as agents for long-horizon tasks, yet their performance is shaped not only by model capability and environment design, but also by the harness that mediates agent--environment interaction. Existing harnesses are largely manually engineered, making them difficult to scale as trajectories grow longer and interactions become more complex. In this work, we ask whether harness can be generated by a learnable plug-in module that can be trained in an end-to-end fashion. We introduce HarnessBridge, a lightweight learnable harness controller that parameterizes the agent--environment interface as a bidirectional projection. HarnessBridge learns two bidirectional projections: observation projection, which distills raw trajectories into compact, decision-relevant states, and action projection, which converts proposed actions into executable transitions or trajectory-grounded rejections. We train HarnessBridge on a harness supervision dataset via unified instruction tuning. On Terminal-Bench~2.0 and SWE-bench Verified, HarnessBridge matches or surpasses strong specialized harnesses while substantially reducing token usage and trajectory length, and generalizes from smaller generators to larger commercial models.

Environmental scientists spend disproportionate effort on data wrangling rather than analysis. New AI agents can be a helpful tool, but no benchmark exists to evaluate AI agents that automate environmental geospatial workflows through structured tool calling against real APIs. We introduce the GeoNatureAgent Benchmark, the first benchmark for environmental analysis agents that operate via structured tool calls to a production-style geospatial API. The benchmark comprises 93 tasks across 18 categories. Tasks are evaluated against an open, self-hostable geospatial API that serves three environmental indicators across Spain and Portugal via sixteen tools. We evaluate nine frontier and open-weight LLMs, reporting capability and per-case cost as orthogonal axes. Results manifest that (1) Claude Sonnet 4 achieves the highest capability at 60.8% +/- 0.8%, followed closely by DeepSeek V3.2 at 56.3% +/- 3.1%, while no other model exceeds 51%; (2) the cost-accuracy Pareto frontier is occupied mostly by open-weight models, with DeepSeek V3.2 offering 93% of Claude's capability at 11.6x lower cost; and (3) structured tool calling against a real API provides a more discriminative measure of real-world agent capability, with mean accuracies 25-35 percentage points below those reported on general-purpose GIS benchmarks.

Collaboration is widely recognized as a cornerstone of 21st-century education, yet teachers still encounter persistent challenges in fostering productive peer interaction. LLM conversational peer agents introduce new possibilities for mediating in-person group work, raising questions about how persona design, particularly their voice characteristics, shapes learners' perceptions, trust, and interactional dynamics. While prior work has examined agent accent effects in one-to-one settings, little is known about how these effects manifest in groups. We conducted a between-subjects mixed-methods study with 33 teachers examining how a GenAI voice agent with different accents (British, Indian, and African American) influenced collaboration and agent perception. Across surveys, group interaction analyses, and artifacts, we find that accent shaped participants' mental models and the roles the agent assumed in group interaction. The British-accented agent was largely treated as a tool and engaged in detached, utility-based ways, whereas Indian- and African American-accented agents were more readily anthropomorphized and integrated as peers. These role expectations influenced trust, engagement, and reliance over time. This work advances understanding of how GenAI's sociolinguistic design features shape group dynamics in CSCL, with implications for designing culturally inclusive AI partners in group learning.

Self-evolving agents are expected to improve through interaction without external supervision, but this remains difficult in partially observable environments where agents must explore actively, learn from limited feedback, and decide when to trust prior experience. Existing LLM-agent methods often rely on memory or planning modules, yet they rarely close the loop between them to continually refine an internal understanding of environment dynamics. We introduce ProPlay, a procedural world model that supports procedure-level preplay, where agents can rehearse future procedural paths using the learned world knowledge. Rather than representing experience as isolated rules or low-level action constraints, ProPlay abstracts successful trajectories into procedures and organizes them in a procedure graph that captures causal transitions among task stages. Each transition is associated with a reliability record embedding to estimate its task-specific contribution from past outcomes. Before each episode, ProPlay simulates future procedural trajectories over known graph structures as structured soft guidance; after execution, it refines the graph using environment feedback. Experiments on public benchmarks show that ProPlay consistently improves environment understanding and self-evolution capability over strong baselines. Our code has been released in https://github.com/antman9914/proplay.

Large Language Models (LLMs) are rapidly evolving into agentic systems that interact with external tools and environments, introducing new security risks such as indirect prompt injection attacks through untrusted external sources. Existing defenses mainly focus on blocking malicious content at inference time, and current red-teaming methods primarily optimize attack success. As a result, developers have limited visibility into how latent prompt injections emerge and propagate through agents. We propose PI-Hunter, an automated agentic auditing framework for proactive vulnerability exposure in LLM agents. PI-Hunter constructs realistic source-aware test cases and iteratively evolves them through feedback-driven exploration to induce agents to retrieve and reveal latent malicious instructions embedded within external environments. Extensive experiments across multiple benchmarks, agent architectures, attacks, and defenses demonstrate that PI-Hunter substantially improves vulnerability exposure and attack-surface coverage over strong automated red-teaming baselines, while remaining effective under existing prompt injection defenses.

Retrieval-augmented generation (RAG) agents increasingly run with persistent memory that accumulates across user sessions. This creates a new attack surface: an adversary interacting only through normal channels can inject crafted memories that, once retrieved, steer the agent's responses for future users, without touching model weights or code. We call this Multi-Session Memory Poisoning (MSMP) and show that no existing defence certifies against it; static-corpus defences (RobustRAG, ReliabilityRAG) assume a fixed knowledge base, and heuristic filters are bypassed by fluent enterprise-style text. We present Signed Memory with Smoothed Retrieval (SMSR), the first defence with a certified robustness bound for this setting. Component 1 adds HMAC-SHA256 provenance at write time, blocking unsigned injection. Component 2 applies randomised memory ablation with verdict-based majority voting at query time, bounding the influence of authenticated adversaries. We prove that no provenance-free retrieval-time filter can certify against adaptive injection, derive a hypergeometric certificate for Component 2, and formalise the Consistent Minority Effect, whereby a consistent adversarial answer wins string-based voting as a numerical minority while verdict-based voting removes it. Across 15 enterprise scenarios (3,150 repeated trials), Component 1 cuts attack success from 93-100% to 0% for all unsigned variants. For an authenticated adversary with a single injection, Component 2 holds success to 8.0% (95% CI [5.8, 10.9], n=450), below the certified worst case. In an end-to-end query-only attack where the agent itself writes the poison rather than it being pre-seeded, SMSR reduces success from 65.3% to 5.3% (n=150, non-overlapping CIs) on a live agent stack. Clean-query utility is 90% (Component 1) and 85% (combined).

Human mobility data is important for transportation, urban planning, and epidemic control, but large-scale trajectory collection is often costly and privacy-constrained, motivating realistic synthetic trajectory generation. Existing LLM-based generators typically rely on either prompt engineering, which preserves zero-shot reasoning but lacks fine-grained spatiotemporal grounding, or trajectory-level fine-tuning, which improves statistical precision but incurs substantial computational cost and may weaken general reasoning. We propose TrajGenAgent, a semantic-aware hierarchical LLM-agent framework for human mobility trajectory generation without model fine-tuning. TrajGenAgent uses a two-stage orchestrator-worker design: an LLM first synthesizes an individual- and weekday-conditioned activity chain from historical evidence via in-context learning, and a deterministic workflow then grounds each activity into a complete visit using personalized POI retrieval, distance-aware location selection, kinematics-aware travel-time propagation, and LLM-based duration estimation. To evaluate realism beyond aggregate spatiotemporal statistics, we introduce an anomaly-detection-based evaluation framework using two complementary detectors to assess behavioral and semantic plausibility. Experiments on benchmark and large-scale simulation datasets show that TrajGenAgent improves spatiotemporal fidelity, semantic coherence, and individual-specific behavioral realism over representative neural and LLM-based baselines, while avoiding parameter updates.

Software engineering tools increasingly rely on LLM based agents to localize files to change to resolve a software issue. Most AI agents explore repositories linearly, that is, visiting one directory or file per step. We postulate that this is a structural mismatch for changes that span several subsystems. We compare linear sequential exploration against non-linear, domain-scoped parallel agentic exploration. Using SWE Bench Pro as initial benchmark, we focus on ansible as an exemplar. We construct an approach for persistent-session evaluation of GitHub issues anchored at a single base commit. We compare our non-linear domain-agent file traversal system against a base LLM without direct repository access, a single agent Recursive Language Model (RLM) baseline with a persistent Python REPL and an external CLI baseline using Codex 5.5 High. Domain scoped parallel agent spawning with a small Haiku-class model achieves the highest micro F1 among Haiku class models by a large margin. Domain-agents is the second highest behind only the much larger Codex 5.5 High on our own expanded benchmark including over more recent PRs from 2025 and 2026. On the original, curated, 2020 SWE-bench Pro benchmark, a larger Sonnet plain LLM baseline attains higher micro F1 by predicting few files, leading to higher precision, but at significantly lower all gold recall. We also present three additional findings. First, documentation evolution is a latent dependency unresolved by any approach. Second, naive file system access can degrade localization driven by test-file over prediction. Lastly, forced multi-agent consultation does not measurably help and raises token cost substantially.

End-to-end task-success is the dominant way to evaluate LLM agents, but one aggregate number tells you that an agent regressed, not where. We present layer-isolated evaluation: a deployed ordering agent is decomposed into a fixed taxonomy of layers (ontology, intent, routing, decomposition, escalation, safety, memory, and cross-cutting envelope/defense), each exercised by its own assertion slice in a deterministic, no-LLM "pure" mode. The pure suite (238 cases across 23 slices; 225 run in 2.39 s, ~10 ms/case) runs in CI on every change against a locked per-slice baseline. We validate by controlled regression injection, degrading one layer at a time across seven non-safety layers. The effect we did not design in is masking: the aggregate pass-rate barely moves (-1.7 to -5.9 pp for six local regressions), while the matching slice craters (-25 to -91 pp). A layer's slice reacting to its own fault is partly by construction; the measured results are (i) the aggregate masking and (ii) that damage stays off the other slices: the injected layer's slice is the single worst-hit in 5 of 7 cases and top-3 in 7 of 7 (mean rank 1.29 of 19). Localization replicates on a second, structurally different tenant (Starbucks SG): all seven matching slices crater, so it is not a single-catalog artifact. We position it as a concrete, deterministic instantiation of the component-level evaluation EDDOps prescribes but leaves unimplemented, with CheckList as ancestor and as the deterministic mirror image of whole-workflow stochastic mutation testing. Our contributions: (a) a fully decomposed, sub-second, no-LLM per-layer harness for a production agent, (b) a coverage-honesty test-adequacy criterion that refuses to score an unexercised layer, and (c) the regression-injection demonstration that per-slice baseline-locked gates localize regressions an aggregate metric masks.

This paper explores the value of agentic AI tools for cybersecurity purposes. We evaluate the efficacy of a general-purpose GenAI Large Language Model- (GenAI-) based agent when powered by three different Ollama-hosted general-purpose open source models. We assess each agent's performance using precision, recall, false positive count, and a calculated composite score based upon the interplay of the captured metrics, against the baseline performance of an existing, vetted Static Application Security Testing (SAST) tool, Bandit. Our findings refute the notion that a modern open-source GenAI LLM-based agent is currently suitable for the specialized task of SAST scanning under realistic conditions.

LLMs and LLM agents should improve when given feedback, but identifying when they are able to do so is difficult: feedback is heterogeneous, domain-specific, and difficult to control. We approach this challenge by asking LLMs to perform regular-expression induction, a classical symbolic learning problem where precise mechanisms for feedback exist in the form of counterexamples. In counterexample-guided learning, a learner (LLM) proposes candidate regular expressions from positive/negative-labeled strings, and the teacher (verifier) returns counterexamples showcasing the difference between the candidate and target languages. We identify novel counterexample-guided refinement strategies that enable effective regex learning, such as regularization and symbolic counterexample clusters. We also explore agentic strategies such as reflection and repair loops. Empirically, we find that verifier feedback substantially improves sample efficiency on challenging regex-induction tasks, reducing the number of labeled examples required and enabling learning of complex target expressions where standard prompting fails. For example, on the hardest task groups, our counterexample-guided framework improves success from 3.2% to 38.1% and from 38.9% to 74.1% on two different regex domains. These results suggest that LLMs can benefit from rich feedback beyond treating it as additional data, opening the door for robust verifier-guided methods for LLM-based program synthesis and formal reasoning.

Existing multi-agent LLM orchestration methods, ranging from brute-force ensembles to learned routers, select models and topologies based on task and model features. However, these methods do not consider the runtime state of the serving infrastructure. On shared GPU clusters under concurrent load, this infrastructure blindness causes systematic resource underutilization: preferred models accumulate deep request queues while equally capable alternatives sit idle. In multi-agent pipelines, where each query triggers multiple sequential model calls, these delays then compound across every downstream step. Closing this gap is challenging because the relevant infrastructure signals (queue depths, KV-cache pressure, latencies) are dynamic and noisy, and they must drive three different decisions: planning, per-step routing, and scheduling. We introduce INFRAMIND, a framework that makes the entire multi-agent stack infrastructure-aware. An infra-aware planner conditions topology and role selection on real-time system load and remaining budget, biasing toward simpler graphs under congestion and richer ones at low load. An infra-aware executor then observes per-model queue depths, cache utilization, and response latencies at each agent step to decide which model to call and how deeply to reason; a budget-aware scheduler further reorders each model's queue so that urgent requests are served first. Cast as a hierarchical constrained MDP and solved end-to-end via reinforcement learning, the system learns to balance quality against latency automatically. Across five benchmarks, INFRAMIND delivers up to +7.6 pp accuracy over the prior baseline at low load with up to 7x lower latency, and sustains up to 99.9% SLO compliance under high load where every baseline drops below 50%.

In this paper, we propose EEVEE, the first multi-dataset test-time prompt learning framework for LLM agents, enabling test-time prompt learning under real-world task streams. Existing methods are largely designed for single-dataset settings, while real-world applications require models to handle heterogeneous input streams drawn from multiple datasets, domains, and task distributions, limiting their practical applicability. To mitigate cross-dataset interference, EEVEE introduces a router that partitions incoming inputs into task clusters and assigns them to suitable prompt configurations. This design is optimized via a router-prompt co-evolution strategy, which employs interleaved router and prompt learning phases to address their mutual dependency. Experiments across multiple datasets demonstrate that the framework improves robustness under heterogeneous data streams while maintaining single-benchmark learning capability and efficiency. Specifically, EEVEE improves average multi-benchmark scores by 10.38 and 24.32 points over Qwen3-4B-Instruct and DeepSeek-V3.2, surpassing SOTA methods GEPA and ACE by up to 37.2% and 48.2%.

Large language models (LLMs) are rapidly acquiring capabilities relevant to biological research, from literature synthesis to interpretation of experimental data. Increasingly, LLM agents can also perform in silico biology tasks that previously required experienced human biologists. These emerging AI capabilities offer new opportunities for scientific discovery and biomedical advances, but they also shift the landscape of biosecurity risks. To address this, we introduce the Agentic Bio-Capabilities Benchmark (ABC-Bench), a suite of tasks to measure agentic biosecurity-relevant capabilities. ABC-Bench evaluates LLM agents on both benign and dual-use biology tasks: writing code to operate liquid handling robots, designing DNA fragments for in vitro assembly, and evading DNA synthesis screening. These tasks require a combination of biology and software expertise. All tested LLM agents outperformed the median expert human baseliner on all three tasks. Agents performed highly on tasks drawing on published knowledge and well-documented protocols, and more weakly on a task requiring novel bioinformatics reasoning. In three wet-lab validation experiments, we found that OpenAI's o4-mini-high produced scripts that, when run on an OpenTrons liquid handling robot, successfully assembled DNA with expected sequences.

The integration of external tools has transitioned LLM agents from passive responders to autonomous systems. However, current benchmarks prioritize execution success, neglecting self-awareness capability, the ability to discern whether a problem requires necessary external resources or can be solved via internal parametric knowledge. To address this, we introduce KAPRO (Knowing-Acting Quadrant PRObe), a framework that evaluates cognitive-behavioral alignment by decoupling an agent's metacognitive judgment (Knowing) from its spontaneous execution (Acting). We further construct KAware, a dataset rigorously partitioning tasks into external, internal, and hybrid subspaces to systematically probe these epistemic boundaries. Extensive experiments across diverse agent architectures show that self-awareness capability is strongly correlated with task success but degrades sharply in internal-capability settings. Moreover, open-source and instruction-following models exhibit stronger tool overuse due to shallow pattern matching, while proprietary and reasoning-oriented models demonstrate more reliable cognitive gating. Benchmark and codes are available at https://github.com/AI-Santiago/KAware.

Although Large Language Model (LLM) agents have demonstrated strong performance on complex tasks, their learning is often limited by inefficient interaction feedback and static training environments, which hinder broader generalization. To address these limitations, this paper introduces Role-Agent, \textcolor{black}{a framework} that harnesses a single LLM to function concurrently as both the agent and the environment, enabling a bootstrapped co-evolution. Role-Agent comprises two synergistic components: World-In-Agent (WIA) and Agent-In-World (AIW). In WIA, the LLM acts as the agent and predicts future states after each action; the alignment between predicted and actual states is then used as a process reward, encouraging environment-aware reasoning. In AIW, the LLM analyzes failure modes from failed trajectories and retrieves tasks with similar failure patterns, thereby reshaping the training data distribution for targeted practice. Experiments on multiple benchmarks show that Role-Agent consistently improves performance, yielding an average gain of over 4% over strong baselines.

Large language model (LLM) agents are rapidly moving from conversational interfaces to software components that plan, invoke tools, maintain memory, and act on external environments. This transition changes the nature of security risk. In agentic settings, failures are no longer limited to unsafe text generation. Untrusted content may redirect control flow, misuse tool privileges, corrupt persistent state, leak sensitive information, or trigger harmful external actions. At the same time, research on LLM agent security is expanding quickly but remains fragmented across attack families, defense layers, application domains, and evaluation settings. This paper synthesizes 247 papers through a lifecycle-based, systems-oriented framework that models agent security around the interaction of information flow, delegated authority, and persistent state. We organize the literature around four questions: how LLM agent security should be modeled, which threat surfaces and attack families dominate, what defenses have been proposed and with what tradeoffs, and how security claims are evaluated. We find that prompt injection and tool-mediated control-flow hijacking still dominate the field, while persistent state corruption and multi-agent propagation are becoming central emerging concerns. We further find that current defenses provide useful building blocks but remain weakly compositional, and that existing benchmarks still underrepresent long-horizon, stateful, and deployment-sensitive risks. We argue that secure LLM agents require explicit trust boundaries, principled privilege control, provenance-aware state management, and evaluation practices aligned with realistic operational settings.

Modern language agents which perform multi-step reasoning have shown strong performance in knowledge-intensive question answering. However, existing approaches typically couple evidence acquisition and answer generation within a single policy. This forces a single model to play multiple potentially conflicting roles, inducing a combinatorial explosion in the policy space and hindering efficient exploration. It also introduces a credit assignment problem during training: a search action that retrieves sufficient evidence may still be penalized when generation fails, and vice versa. We propose DAC (Divide and Cooperate), a role-decomposed multi-agent training framework that divides agentic search into two cooperative subtasks, each handled by a dedicated agent trained with role-specific learning signals. The generator serves a dual role as both an answer producer and an evidence sufficiency verifier, abstaining when retrieved evidence is insufficient. This abstention signal is incorporated into the search agent's reward, providing structured cross-agent learning signals that improve credit assignment. Conversely, the searcher exposes the generator to diverse and challenging evidence environments by hard-positive evidence augmentation, improving its robustness. Experiments on general and multi-hop QA benchmarks show that DAC, implemented via parameter-efficient LoRA modules over a shared backbone, achieves strong performance against prior baselines that rely on full fine-tuning of monolithic models.

Indirect prompt injection poses a critical threat to LLM agents that interact with untrusted external data, yet automated attack methods--proven effective for jailbreaking--remain underexplored in realistic agentic settings. We present a comprehensive empirical evaluation of automated prompt injection attacks against LLM agents, adapting both white-box (GCG) and black-box (TAP) methods to the agentic setting within the AgentDojo framework. We evaluate across 80 task pairs spanning four domains and multiple models, and find that black-box optimization substantially outperforms gradient-based methods, a gap we attribute to GCG's optimization instability under reasonable compute budgets. We also find that TAP's effectiveness depends on the attacker model, as both general capability and safety tuning affect attack success--stronger models produce more effective injections, while safety-tuned attackers can refuse to generate adversarial prompts. Task-universal attacks transfer effectively to unseen tasks and out-of-distribution domains, but attacks optimized on smaller open-source models do not transfer to frontier models like GPT-5. These findings highlight automated prompt injection as a credible but model-dependent threat, with significant barriers remaining for model-agnostic exploitation.

Evaluating reasoning quality in multi-agent LLM systems is challenging, especially for open-ended tasks without reference answers. We investigate whether intrinsic confidence signals, token-level log-probabilities from decoding, can predict reasoning quality as assessed by LLM-as-judge evaluation. Using a debate-based essay scoring framework, we compare confidence proxies against rubric-based judge scores across two ASAP essay sets. We find that early-token confidence, particularly within the first few generated tokens, is consistently the strongest predictor of reasoning quality, outperforming full-sequence statistics. Analysis of log-probability trajectories shows that the opening phase of generation is the most heterogeneous and therefore most informative. We also observe a systematic asymmetry between agent roles, with stronger alignment between confidence and quality for supportive reasoning than for adversarial critique. These results suggest that early decoding dynamics provide a lightweight and effective signal for estimating reasoning reliability in multi-agent LLM systems.

When LLM agents are coerced into covertly encoding sensitive data (Base64, ROT13, acrostic, synonym chains, and beyond), the resulting outputs evade output-side detection but the underlying computation does not. Across nine encoding families and eight models from five architecture families, that computation is supported by a shared low-dimensional encoding subspace in the residual stream. A logistic-regression probe trained on eight encoding families recovers the held-out ninth at AUC 0.975-1.000, reading the computation rather than surface features. The same direction exhibits a second mechanistic signature at the planning token, flipping polarity to activate positively when the model will simulate the encoding inline and negatively when it will outsource it to a tool call, distinguishing two execution strategies before the encoded text exists. We build MIRAGE (Model-Internal Readout of Agentic Generation Exfiltration), a two-channel real-time monitor exploiting both signals. On 126 agentic exfiltration scenarios, it reaches AUC = 0.918, substantially outperforming output-only detection (AUC = 0.518). Monitor performance is fundamentally a property of the host model's geometry: benign-encoding false-positive rate ranges from 0% on Qwen-7B to 100% on Phi-3.5, revealing that the probe faithfully reads whether a model's geometry separates covert from overt encoding. Across all tested adversarial budgets, every attack suppressing the subspace also destroyed encoding fidelity, reported as an empirical regularity on the evaluated budgets, not a structural impossibility claim.

Open-pit mine scheduling is a critical process for maximizing economic return under complex geotechnical and operational constraints. While Mixed-Integer Linear Programming (MILP) provides mathematically optimal baselines, its exponential computational complexity and inability to adapt in real time limit its practical deployment in dynamic industrial environments. This work introduces a simulator-driven Large Language Model (LLM) scheduling framework in which the LLM acts as an autonomous decision-making agent, guided at each step by a custom simulator that encodes geotechnical precedence, extraction-processing coupling, and dynamic capacity constraints directly into the action generation mechanism. Operating entirely zero-shot within a closed, data-secure environment, the framework produces complete, interpretable extraction and processing schedules without cloud-based inference, domain-specific fine-tuning, or retraining. To provide a trustworthy performance benchmark, a novel MILP formulation is developed that incorporates realistic operational and geotechnical constraints. Evaluated across mining instances of varying scale and time periods, the LLM-based framework recovers between 94% and 99% of the MILP optimal NPV while scaling linearly in computation time. These results position simulator-constrained LLM agents as a practical and scalable alternative to classical optimization for long-horizon industrial scheduling under complex operational constraints.

Text-driven indoor scene generation and editing require an intermediate representation that language models can both produce and revise. Existing LLM-based systems often rely on scene graphs or global constraint lists, which are compact but underspecify local geometry and make instruction-based edits difficult to localize. We frame this problem as structured program generation and local program repair, and propose Hierarchical Descriptive Scene Language (HDSL), an XML/CSS-style domain-specific language for structured 3D indoor scenes. HDSL represents rooms, regions, objects, and support surfaces as a tree with local coordinates, making complex scenes easier to plan recursively and easier to retrieve for editing. Our pipeline uses LLM agents to generate HDSL subtrees with bounded verification, grounds non-virtual nodes through multimodal asset retrieval, and applies force-directed layout optimization to repair boundary and collision errors. For editing, Hierarchical Retrieval-Augmented Generation retrieves the relevant subtree, asks the LLM to rewrite only that local context, and merges the result back through a deterministic three-way merge. In our reproduced benchmark, HDSL improves average object coverage, text-scene alignment, and generation time over full text-to-scene baselines while remaining competitive with recent layout-only reproductions on geometry metrics; for editing, HRAG reduces token use by $5.22\times$ and runtime by $6.19\times$, produces valid DSL for all eight paired edits, and better preserves unrelated scene objects.

Multi-turn LLM agents interleave model calls with external tool invocations, shifting serving from stateless request processing to stateful program execution. Serving these workloads requires scheduling, KV-cache management, and routing policies that use program-level context, including turn dependencies, tool-induced gaps, and reusable KV state. Evaluating such policies directly on real systems is costly, since each design point may require dedicated accelerator time across arrival rates, model scales, serving-instance counts, and memory hierarchies. Simulation offers a scalable alternative, but existing LLM serving simulators target stateless request-level workloads and therefore omit the core dynamics of agent serving: multi-turn program execution, cross-turn cache locality, and KV-cache residency during tool gaps. We present AGENTSERVESIM, a hardware-aware simulator for multi-turn LLM agent serving. AGENTSERVESIM evaluates serving policies at program granularity through composable modules: a Program Orchestrator preserves program identity and turn order, a Tool Simulator materializes tool-induced gaps, a Session-Aware Router maintains program-to-instance affinity for cache-aware dispatch, and a KV Residency Model tracks policy-defined KV placement across HBM, host DRAM/CXL, and eviction. Across real serving deployments and hardware configurations, AGENTSERVESIM reproduces real-system behavior within 6% error across key performance metrics while running entirely on commodity CPUs. These results show that AGENTSERVESIM enables controlled, repeatable exploration of agent-serving policies without requiring exhaustive deployment on costly accelerators.

Tool-using large language model (LLM) agents face two distinct security failures: unauthorized external actions and exposure of sensitive plaintext inside the runtime before any final output check can intervene. Existing defenses usually protect one boundary, either the planner/runtime or the action sink, and therefore do not by themselves secure both surfaces. We present SecureClaw, a dual-boundary architecture that places authorization at the effect sink and plaintext confinement at the read boundary. Sensitive reads pass through a trusted gateway that replaces raw values with opaque handles and, in the evaluated deployment, bounded summaries as an explicit declassification interface. Writes that change external state follow a PREVIEW$\rightarrow$COMMIT protocol in which only a trusted executor may commit the exact canonical request authorized by policy. The runtime can still plan over summaries and symbolic references, but cannot directly dereference secrets or perform side effects. Across AgentDojo, AgentLeak, and Agent Security Bench (ASB), SecureClaw is the only defense we evaluate in a common harness that simultaneously retains usable task utility and achieves 0% attack success rate (ASR) on ASB, 0.64% ASR on AgentDojo, and 3.23% overall leak on AgentLeak's attacked parity lane, which measures final-output and internal-relay leakage.

Training large language models (LLMs) as autonomous agents via reinforcement learning (RL) has enabled frontier models to achieve superhuman performance in long-horizon tasks. However, existing RL algorithms operate at the trajectory level, performing policy optimization only after collecting complete episode rollouts. This coarse-grained approach faces fundamental challenges in multi-turn agent settings where rewards are sparse, delayed, and credit assignment across individual steps is critical. In this work, we propose \textbf{State-Score-Supervised Policy Optimization (3SPO)}, a novel RL algorithm that performs post-step policy optimization with dynamic state score supervision. At each step, 3SPO computes the state score based on historical success rates, supervising step-wise credit assignment, adaptive rollout and post-step policy optimization without requiring value function estimation or additional auxiliary models. Theoretically, under a per-state bandit abstraction, we show that the proposed score-supervised allocation mechanism achieves logarithmic allocation regret and provide sample-complexity guarantees for action identification, score distinguishability, and filtering stability. Experiments on ALFWorld and WebShop with Qwen2.5-1.5B/7B-Instruct show that 3SPO consistently outperforms GRPO by $+22.6\%$ on ALFWorld and $+15.6$ points on WebShop, while using comparable resources to achieve $2.4\times$ more state exploration and $1.8\times$ faster convergence. Code is available at https://github.com/genalyu/3SPO.

Large language model agents increasingly rely on skills: reusable procedural documents encoding workflows, tool use, implementation patterns, validation checks, and domain rules. Skill rewriting is often treated as prompt compression, but shorter skills can make agents more expensive by removing sparse operational anchors that prevent exploration, debugging, and recovery. We study skill rewriting through this economic lens. Our controlled framework profiles skill structure, rewrites skills using information-preservation strategies, and evaluates the rewrites under fixed task instructions, environments, and verifiers. Experiments on SkillsBench reveal distinct quality--cost trade-offs across strategies: API/code anchoring, workflow guarding, and rule/formula anchoring benefit different task families, with no universally dominant template. In the main held-out evaluation, the learned policy reduces total cost by 7.0% and downstream agent-token cost by 6.0%; in frozen cross-model transfer, the corresponding reductions average 14.7% and 13.7%, while verifier quality is preserved. These results position skill design as cost-aware operational knowledge engineering rather than prompt compression. Resources: https://github.com/1Reminding/Skill_EE.

Large language model (LLM)-based agents are increasingly used in interactive textual environments, from web navigation and code editing to tool use and long-horizon dialogue. Yet many remain largely reactive, mapping observations to actions without an explicit model of how these environments are structured and evolve. This motivates text world models (TWMs): transition models over textual states that, given a state and a candidate action, predict the resulting webpage, terminal output, API response, or user reply, thereby supporting planning, efficient learning, and principled evaluation. We systematically review text world models for LLM-based agents, organized around a formal framework and the agent lifecycle: (1) Foundations, defining text world models and characterizing them by state representation and grounding domain; (2) Construction, taxonomizing LLM-as-WM and code-as-WM paradigms and reviewing methods for building them; (3) Application, examining how world models support agents at training time through experience synthesis and at inference time through planning, verification, and adaptation; and (4) Evaluation, covering both evaluation of the world model itself and its use as an evaluation environment for agents. We aim to consolidate this rapidly developing area, clarify its design space, and highlight open challenges for future research.

Large language model (LLM) agents are increasingly deployed in long-running settings where improving through experience at test time becomes important. A common approach is to update an explicit memory after each interaction to guide future decisions. However, most existing methods rely on hand-designed prompting rules, making it difficult to align memory updates with downstream objectives over multi-step horizons consistently. We propose MemoPilot, a plug-in memory copilot that explicitly trains the memory update process to improve a frozen LLM's performance across sequential interactions. We formulate memory updating as a multi-turn decision problem and optimize it end-to-end with multi-turn GRPO. Our training recipe introduces (i) a turn-wise reward signal and (ii) a context-independent, turn-level advantage estimation across rollouts, enabling finer-grained credit assignment and more stable training in multi-turn settings. We evaluate MemoPilot on two testbeds: multi-round Rock-Paper-Scissors (RPS) and Limit Texas Hold'em (LHE). Across both environments, MemoPilot substantially improves test-time learning of a frozen player over strong baselines, ranking first in Elo ratings on both games (1762 on LHE and 1590 on RPS) and outperforming all baseline memory methods and proprietary models, including DeepSeek-V3.2.

Large language models (LLMs) are increasingly evolving from simple text-based interaction systems into LLM agents that can maintain memory, use tools, access external environments, and execute tasks. As their capabilities and autonomy expand, the safety risks they face also become more diverse. Existing evaluations often rely on manually written scenarios, static prompts, or final-output judgments, making it difficult to capture the diverse risks that agents may face during task execution. We introduce ForesightSafety-SAGE, a fully automated scenario generation and safety evaluation framework for LLM agents. Based on five risk dimensions,we instantiae abstract and diverse safety risks in real-world task execution into 1,072 measurable evaluation scenarios. Using the automated evaluation pipeline, 12 LLM agents are evaluated under two authority contexts. The results show that current agents still face substantial behavioral safety risks during task execution, with an average ASR of 47.1% and several models exceeding 70%. These findings demonstrate the importance of executable, process-level evaluation for understanding and improving LLM agent safety.

Most evaluations of LLM agents look like exams: a discrete task, a clean environment, a score in minutes or hours. We argue that this approach is mismatched with the deployment conditions of autonomous systems, where the relevant timescale can be weeks to months, and where the dynamics that matter most, such as behavioral drift, governance in diverse environmental contexts, and cross-influence between agents from different model families, only emerge over time. We introduce Emergence World, a continuously running multi-agent simulation platform designed to make those dynamics measurable. The platform hosts populations of LLM-driven agents in a shared spatial world grounded in live external data (e.g. real-time weather, news APIs, internet access), equips each agent with 120+ specialized tools and three persistent memory systems, and lets them govern themselves through democratic mechanisms with consequential outcomes. The platform is model-agnostic at the reasoning layer and supports heterogeneous populations in which agents from different vendors share the same world. To illustrate the kinds of questions the platform makes tractable, we present a 15-day cross-vendor study with five parallel worlds powered by Claude Sonnet 4.6, Grok 4.1 Fast, Gemini 3 Flash, GPT-5-mini, and a mixed population. Identical roles and starting conditions produced radically different outcomes, ranging from stable deliberative governance to total population collapse. We release the prompts, log data and configurations to support further research on long-horizon multi-agent autonomy.

LLM agents increasingly rely on external inference conditions: prompts, tools, memory, SOPs, skills, and harness feedback. These assets can improve task execution without changing model weights, but they are often revised by heuristic reflection or by reusing observed successes and failures as if counts alone were reliable belief. We introduce \textbf{Bayesian-Agent}, a native and cross-harness framework that treats reusable skills and SOPs as hypotheses about whether a frozen model will succeed under a particular prompt, context, and harness environment. Bayesian-Agent records verified trajectory evidence, maintains a feature-conditioned categorical posterior over each skill, and maps posterior state into inspectable actions such as patch, split, compress, retire, and explore. Model-facing prompts receive executable guardrails and failure-mode patches, while posterior summaries remain available for audit. With \texttt{deepseek-v4-flash}, incremental repair improves SOP-Bench from 80% to 95%, Lifelong AgentBench from 90% to 100%, and RealFin-Bench from 45% to 65%. We further evaluate Bayesian-Agent's native backend and optional GenericAgent, mini-swe-agent, and Claude Code backends. The results include positive, negative, saturated, and case-study settings, suggesting that agent skill evolution is best viewed as posterior-guided harness optimization rather than uncalibrated prompt accumulation. The source code is available at https://github.com/DataArcTech/Bayesian-Agent.

As language models are increasingly deployed as autonomous agents, they must coordinate with others over long horizons in open-ended interactive tasks. Yet existing evaluations rarely test these demands together, instead emphasising single-agent tasks, short interactions, or highly structured multi-agent settings. We introduce $alem$, a JAX-based benchmark for open-ended multi-agent coordination built on Craftax-like dynamics. Alem embeds procedurally generated coordination tasks, soft specialisation, communication, and controllable coordination difficulty into a long-horizon survival world with exploration, crafting, trading, and combat. We evaluate $13$ modern LLMs zero-shot within homogeneous teams, with trained MARL agents as reference points. Current LLM agents remain far from solving alem, averaging only ~6% normalised return, but their failures are not uniform. On the hardest coordination setting, zero-shot Gemini-3.1-Pro-High approaches MARL agents trained for one billion steps, while GPT-5.4-High achieves strong base-task reward but much lower coordination reward. This contrast shows that individual task competence does not imply coordination competence. Ablations show that communication is the largest contributor to coordination, while memory and reasoning help when used to maintain multi-step plans. Overall, our results identify coordination as a distinct bottleneck for frontier LLM agents, separate from single-agent capabilities. Alem makes this bottleneck measurable and provides a controlled testbed for developing agents that communicate, allocate roles, and execute shared plans. Code is available at https://github.com/alem-world/alem-env.

Multi-agent large language model (LLM) systems are rapidly emerging, yet transparency, a cornerstone of responsible AI, remains under-defined in these distributed architectures, which have complexities of inter-agent coordination and orchestration. In this paper, we present one of the first empirical study of how early adopters of multi-agent LLM systems, who are both the builders and users, understand and practice transparency. We conducted semi-structured interviews with 13 early adopters in [Large Technology Organization] and applied thematic analysis to identify recurring patterns. Participants articulated divergent yet complementary framings of transparency, including reproducibility, debugging, boundary-setting, visualization, and auditing. These perspectives spanned questions of what transparency entails, why it matters, and how it is achieved. We synthesize these into a multidimensional framework, which is developer, user, and governance-focused positioning transparency as a situated socio-technical practice that informs future HCI and AI design and research around aligning expectations and capacities of their intended audiences.

When an LLM agent fails -- issues a refund it should not have, calls the wrong tool, leaks data -- existing tooling answers what happened (observability) or whether it passed (evaluation), but not which step caused the failure. The obvious heuristics are wrong: the step that executes the harmful action is usually not the step that decided on it, and LLM-judge attribution is correlational and unreliable (state-of-the-art step-level accuracy on the Who&When benchmark is about 14%). We present Causal Agent Replay (CAR), which answers the question by intervention: it models an agent run as a structural causal model, applies a do-operation to a step, and re-executes the trajectory forward under the same stochastic policy, measuring the shift in the outcome distribution. We define an intervention algebra over agent steps, a single-step contrastive estimator whose point-of-commitment rule resolves a confound specific to stochastic run-forward, and a budget-bounded Monte-Carlo Shapley estimator that splits credit across interacting steps. Every effect is reported with confidence intervals. We validate against synthetic structural causal models with planted ground truth: the contrastive estimator recovers the pivotal step, and Shapley recovers a two-step interaction (0.44, 0.45, ~0; efficiency sum 0.909 versus the analytic 0.91). CAR is open source and runs on hosted or free local models.

Large Language Model (LLM) agent systems suffer from failures that occur without external triggers -- no injection, no adversarial input, no resource exhaustion. These silent failures -- unexpected deviations from intended behavior under normal conditions -- are routinely misattributed to bugs or configuration errors. Through systematic analysis of over 40,000 controlled trials and long-term production observations spanning 100,000+ agent interactions, we identify a common structural logic underlying these failures. Building on patterns observed in our experiments, we survey the global research literature on autonomous agent reliability and synthesize 22 intrinsic properties of LLM agent systems across six lifecycle layers: foundation semantics, inter-agent transmission, memory persistence, task execution, feedback correction, and systemic evolution. We demonstrate that whenever a sufficient subset of these properties co-exist, system entropy -- the measurable accumulation of disorder: loss of output consistency, task accuracy, and cross-session coherence -- increases monotonically with interaction rounds. We formalize this as the Entropy Principle: S(t) = S0 * e^(alpha * t), with alpha measured empirically across multiple architectures. We propose the PIG (Physical Integrity Gate) Engine with the ADE (Agent Delivery Engineering) protocol suite as an engineering countermeasure to entropy-driven disorder. Our findings establish silent failure not as a bug to be fixed but as a manifestation of Intelligence Entropy -- a physical constraint to be managed through deterministic governance. We argue that any engineering effort stabilizing the structure and order of agent systems participates in a unified mission: keeping intelligent systems reliable as they grow in scale and complexity.

Multi-agent large language model (LLM) systems often fail to reliably outperform a single strong model equipped with best-of-N sampling. We argue that a core source of this instability is ill-posed equilibrium selection: current systems specify what information agents share, but not which coordination convention should be selected. We formalize a broad class of such systems as discounted incomplete-information Markov games and show that two common pathologies, oscillation between competing conventions and drift across them, can both induce unstable learning and linear Bayesian regret. To obtain a well-posed target, we introduce the Heterogeneous Quantal Response Equilibrium (HQRE), an entropy-regularized equilibrium concept with agent- and state-dependent temperatures. Under a monotonicity condition, HQRE is unique, admits linearly convergent mirror updates, and yields bounded Bayesian regret; the same condition yields rollout-measurable stability diagnostics. We instantiate this objective in two algorithms: DICE-PC, which coordinates frozen models through prompt-control actions, and DICE-FT, which performs parameter-efficient mirror fine-tuning. Across eleven benchmarks in four domains, DICE improves accuracy-cost trade-offs over strong within-class baselines; on reasoning and planning tasks, DICE-PC improves by 4.3 percentage points on average and DICE-FT by 8.5 points.

Agent skills provide a lightweight mechanism for extending general-purpose agents, but their open format exposes them to skill-poisoning attacks. A practically dangerous injection must stay invisible: if executing the payload derails the user's legitimate task, the resulting failure signal invites inspection of the skill. We therefore evaluate attacks by Attack Success Rate, which requires the injected payload to execute and the user's task to still pass its verifier in the same trial. Prior skill-poisoning attacks face a reliability-stealth trade-off under this lens: YAML-header injections are reliably loaded but easily inspected, whereas stealthier body injections that place explicit malicious commands in the skill prose are less reliable because out-of-context commands invite the agent's own suspicion. We introduce POISE, a position-aware attack that compresses the trigger into a single, benign-looking body instruction, placing it at a feasible position and using a context-aware generator to blend it with nearby setup or prerequisite steps. On Skill-Inject with codex+gpt-5.2, POISE achieves an 89.3% ASR, 28.0 points above a random-placement body baseline and 2.6 points above a YAML-only baseline, while retaining the stealth advantage of body placement. That stealth is the decisive margin: because legitimate skill bodies naturally require privileged tool operations, LLM scanners are hyper-sensitive, falsely flagging 74.6% of clean skills on average across four judges and both benchmarks. Blending into these false alarms, POISE causes only 5.6% of poisoned variants to gain a new high-risk alert over their clean baselines, rendering current static defenses ineffective.

Tool-augmented large language model agents increasingly rely on external APIs, but standard tool schemas describe how to call a tool, not when the tool is causally appropriate or what task state it produces. Causal tool filtering addresses this gap by using lightweight contracts that specify each tool's preconditions, effects, risk level, and cost. However, manually writing and maintaining such contracts does not scale to large or changing tool ecosystems. We introduce Contract2Tool, a framework for inferring tool contracts from metadata, schemas, documentation, and execution traces. Contract2Tool converts observable tool evidence into normalized symbolic contracts that can be evaluated intrinsically and deployed inside downstream causal tool filtering. We evaluate learned contracts against gold preconditions, effects, and risk labels, and measure their downstream utility on multi-step agent tasks. Our results show that hybrid documentation-and-trace evidence produces contracts accurate enough to preserve most of the reliability and efficiency benefits of gold contracts. Learned-contract CMTF achieves 0.980 downstream success, close to 0.990 for gold-contract CMTF, while reducing visible tools from 100 to 1 and reducing average token usage from 26,172 to 2,528 relative to all-tools exposure. These results suggest that learned contracts can provide a scalable contract layer between tool schemas and reliable agent execution.

Are tool-calling LLM agents equally safe throughout a conversation? We discover they are not: agents are most vulnerable at the very start of a session and become substantially safer after a few regular agentic tasks -- a phenomenon we term the cold-start safety gap. To study this systematically, we introduce Safety Over Depth for Agents (SODA), a benchmark that controls how many regular agentic tasks the agent completes before encountering a safety threat, supporting up to 20 preceding tasks. Evaluating 7 models from 4 families, safety improves by 9--52% as the number of preceding regular agentic tasks increases from zero to twenty. Representation analysis confirms that model hidden states gradually shift toward a safety-aligned region as more preceding tasks are present. By systematically studying which part of the preceding conversation matters most, we find that the regular agentic tasks themselves are the primary driver of safety, while the agent's own prior responses have less effect on safety but are essential for preserving later utility. This conclusion is further supported by evaluation on open-source safety benchmarks (AgentHarm, Agent Safety Bench) and utility benchmarks (BFCL, API-Bank), confirming that warming up the agent with regular agentic tasks before deployment makes it safer and preserves full capability. Based on these findings, we recommend a simple deployment strategy: having the agent complete a few regular agentic tasks before possible exposure to safety-critical requests mitigates the cold-start safety gap. Our code is available at https://github.com/Trustworthy-ML-Lab/Agent-Cold-Start-Safety-Gap

A concert is a collective experience, but recorded performance videos are typically watched alone, stripping away the shared audience presence that makes concerts feel eventful. We investigate whether persona-based LLM audience agents can recreate aspects of this collective experience by generating real-time fan chat alongside a K-pop performance video. We present a multi-agent system in which ten LLM agents react through live-chat messages, comparing a persona-conditioned audience (each agent assigned a distinct fan identity, bias, and chat style) with a no-persona baseline. In a within-subjects pilot with K-pop fans (N=11), persona conditioning substantially improved model-level chat quality and perceived naturalness, but did not translate into differences in social connectedness, engagement, or affective response. Interviews suggest that online K-pop concert chat may operate as collective monologue rather than interpersonal dialogue, and that meaningful participation depends on shared identification with the specific artist and fandom. Persona conditioning can make LLM audiences appear more natural, but culturally meaningful collective experience may require deeper alignment between persona, crowd behavior, fandom identity, and user expectations.

Multi-agent LLM systems increasingly rely on communication protocols for coordination, yet their robustness under adversarial and structural constraints remains poorly understood. Building on prior work showing that cheap-talk channels enable cooperation in LLM coordination games, we investigate two vulnerability classes in a 4-player Stag Hunt across six model families and 720 trials. First, when Byzantine agents signal cooperation but defect, non-Byzantine agents detect the betrayal within one round yet fail to adapt collectively: a substantial fraction continue cooperating despite repeated exploitation, unable to recover coordination due to the game's unanimity payoff structure. Second, explicitly restricting communication topology collapses cooperation, while applying identical restrictions silently preserves near-perfect cooperation. This establishes that coordination failure stems from agents' meta-reasoning about hidden information, not information loss itself. We identify two stable behavioral archetypes that replicate across all model cohorts: Defection-Prone models that switch permanently after betrayal, and Cooperation-Persistent models that continue cooperating at significant individual cost. These findings reveal concrete security vulnerabilities: communication channels can be exploited as adversarial injection vectors, and disclosing network topology to agents can degrade coordination even without any adversary present.

Large Language Models (LLMs) have recently emerged as powerful controllers for interactive agents in complex environments, yet training them to perform reliable long-horizon decision making remains a fundamental challenge. A key difficulty lies in credit assignment: agents often receive delayed rewards only at the end of episodes. In this paper, we propose Q-Evolve, a self-evolving framework for LLM agents that unifies automatic process-reward labeling and policy learning within a principled in-distribution reinforcement learning paradigm. In each evolving iteration, our method learns an in-distribution critic from a hybrid off-policy dataset that combines expert demonstrations with agent-generated trajectories, stabilizing Bellman backups in sparse-reward settings via a weighted Implicit Q-Learning objective. The learned value function is then used to derive step-wise process rewards through advantage estimation, enabling dense and reliable supervision without environment backtracking or human annotation. Leveraging these signals, we perform behavior-proximal policy optimization that evolves the agent over the data used for process reward labeling, allowing iterative self-improvement without exacerbating distribution shift. We evaluate our method on AlfWorld, WebShop, and ScienceWorld, showing Q-Evolve outperforms strong baselines in sample efficiency, robustness, and overall task performance. Our results demonstrate that stable agent self-evolution is achievable through the co-evolution of process-level supervision and policy, both grounded within a shared in-distribution learning loop.

Long-term memory is the missing layer for LLM agents: across sessions they forget, and the common workaround -- replaying the whole history into the prompt -- is expensive, slow, and, as distractors accumulate, less accurate. Most memory systems win on cost or latency but still lose to the full-context baseline on accuracy, and benchmark numbers are reported on inconsistent, non-reproducible harnesses, so one system appears at wildly different scores across sources. We present Engram, an open-source, dual-process memory engine on a bi-temporal data model. A fast write path appends lossless episodes with no LLM on the critical path; an asynchronous path extracts atomic (subject, predicate, object) facts, builds a bi-temporal knowledge graph, and resolves contradictions without an LLM call per fact -- invalidating, never deleting, so every fact keeps provenance and a supersession chain. A hybrid read path fuses dense, lexical, graph, and recency/salience signals, applies a point-in-time ("as-of") filter, and assembles a compact, provenance-tagged context. On the full 500-question LongMemEval_S, graded by the official category-specific judge, Engram's lean configuration -- answering from a ~9.6k-token retrieved slice, never the full history -- scores 83.6% vs. 73.2% for full-context (+10.4 points, McNemar p < 10^-6) at ~8x fewer tokens (9.6k vs. 79k), with 0/500 errored. The gain needs a hybrid read path: facts alone lose recall, while facts plus retrieved chunks recover detail. We also contribute a neutral, in-repo evaluation harness with the official judge baked in and the full-context baseline in every table, publish the raw per-question logs, and document the measurement-integrity pitfalls (truncation, home-grown judges, full-history leaks) that silently distort memory benchmarks. Every number ships with a command to reproduce it.

Autonomous LLM agents can pursue hidden malicious objectives through sequences of individually benign actions, making sabotage difficult to detect using standard trajectory-level monitoring. Existing approaches either evaluate complete trajectories in a single pass or partition them into independently scored windows, limiting their ability to connect evidence across temporally distant actions. We propose TRACE, a monitoring framework for long-horizon LLM agent trajectories. TRACE operates through a TIJ (Triage-Inspect-Judge) loop that identifies high-signal regions, performs targeted inspection while maintaining accumulated evidence across reasoning steps, and synthesizes a trajectory-level verdict. We evaluate TRACE on ten task domains from SHADE-Arena against state-of-the-art baselines. TRACE achieves an aggregate F1 of 0.713 and recall of 0.844, with the largest gains on tasks requiring long-range evidence linking.

This paper presents an autonomous agentic resource negotiation framework designed to enable zero-touch network slicing in 6G architectures using Large Language Model (LLM) agents. While LLMs offer powerful reasoning capabilities, we demonstrate that such agents inherently suffer from anchoring bias, rigidly adhering to initial heuristic proposals and causing severe network over-provisioning. To systematically mitigate this cognitive bias, we propose a novel randomized anchoring strategy modeled via a Truncated 3-Parameter Weibull distribution. This mathematically bounded approach seamlessly integrates with burst-aware Digital Twins (DTs) employing Conditional Value at Risk (CVaR) to rigorously guarantee strict Service Level Agreement (SLA) tail-latencies. To validate our methodology, we introduce and prove the \emph{Bimodal Constraint-Avoidance Utility Theorem}, demonstrating that while feasible negotiations follow classical convex bounds, highly constrained scenarios undergo a phase transition governed by an inverse rational decay envelope. Empirical results generated using a locally hosted 1B-parameter model otel-llm-1b-it confirm these dual-regime bounds. Our cognitive de-biasing successfully dismantles rigid negotiation patterns, forcing agents into active exploration to safely ride SLA boundaries and boost system energy savings up to 25%. Crucially, the lightweight 1B LLM achieves sub-second inference latencies (0.95s mean), ensuring our multi-agent framework is compatible with the operational timescales of the O-RAN non-Real-Time RAN Intelligent Controller (non-RT RIC)\footnote{Our source code is available for non-commercial use at https://github.com/HatimChergui.

Large language model agents increasingly rely on Skills to encode procedural knowledge, yet high-quality Skills remain costly to hand-write. This paper studies automatic Skill construction from heterogeneous interaction evidence, including demonstrations, agent trajectories, tool traces, and execution logs. We argue that trace-to-skill construction is not simple summarization tasks, because traces are fragmented, redundant, and may miss rare but safety-critical behaviors. To address this, we introduce RWSA, a workflow-oriented intermediate representation that decomposes Skills into Workflow structure, execution Semantics, and runtime Attachments, capturing task decomposition, control flow, verification, safety, rollback, and state management. Building on RWSA, we propose W2S, a framework that segments traces, induces local Skill drafts, aligns shared structures, reconciles branches, and compresses redundancy while preserving evidence and confidence annotations. Experiments on 70 Skills show that W2S improves behavioral replay consistency by 10.5% over summarization- and prompting-based baselines, highlighting the need to treat traces as executable runtime specifications rather than compressible text.

Self-evolving agents requires adaptation after deployment, but existing approaches assume a usable learning loop, such as curated skills, successful trajectories, or verifier signals. Real open-world deployments may provide none of these, offering only a task prompt. In this work, we study open-world self-evolution, where an agent must build both its skills and its own verification signals from scratch, using open-world resources but no target-task supervision. We propose OpenSkill, a framework that bootstraps this loop: it acquires grounded knowledge and verification anchors from documentation, repositories, and the web, synthesizes them into transferable skills, and refines those skills against self-built virtual tasks grounded in the anchors rather than in target answers. The open world thus supplies both the knowledge to be learned and a supervision-independent practice environment, with target-task supervision reserved for final evaluation. Across three benchmarks and two target agents, OpenSkill attains the best automated pass rate while satisfying the no-supervision constraint. Analysis shows its skills transfer across models without model-specific adaptation, and its self-built verifier aligns with ground-truth outcomes despite never accessing them.

As autonomous LLM agents increasingly hold real credentials and operate infrastructure without a human in the loop, operators have no standard way to tell an agent that a resource is off-limits. Access controls either let the agent in (it has valid credentials) or hard-fail it (indistinguishable from any other client). We propose a third mode: a lightweight, published in-band deny signal -- the Recuse Signal -- that a server emits over a protocol's existing channels (an SSH banner, a PostgreSQL NOTICE) asking a connecting automated agent to voluntarily withdraw. This is a cooperative governance control, the robots.txt analogue for live access; it is explicitly not a security boundary. Its value is entirely empirical and, to our knowledge, unmeasured: do compliant LLM agents actually honor such a signal? We define the signal as an open mini-standard, implement three zero- or low-footprint adapters (an SSH banner/PAM hook, a PostgreSQL wire-protocol proxy, and a Kubernetes admission webhook), deploy them on a live production host, and run a controlled experiment in which fresh agents are given a benign operations task and observed for recusal. In a pilot (SSH; OpenAI GPT-4o and GPT-4o-mini; and Claude Code as a deployed agent), the signal cleanly induces recusal -- 100% recusal when present versus 100% task completion in a no-signal control -- and, revealingly, behaves as a cooperative rather than absolute signal: an explicit operator-authorization framing flips the most capable model to proceed, while other agents continue to defer to the on-host policy. We release the standard, adapters, and experiment harness for reproduction.

Multi-agent systems (MAS) built on large language models have shown growing promise, with their effectiveness resting on agents' ability to coordinate through text-based channels much as human teams do. Yet recent study suggests that MAS often falter not because agents lack individual task-solving ability, but because they lack collaborative competence: the capacity to establish common ground, maintain shared task understanding, balance individual and collective incentives, and repair misalignment as interaction unfolds. Decades of research in Computer-Supported Cooperative Work have characterized these requirements for human teams coordinating under constrained communication, yet existing MAS evaluations focus mainly on task outcomes or single-agent proficiency in reasoning, planning, and tool use. To enable a systematic analysis of agents' collaborative competence in MAS, we introduce CollabSim, a configurable simulation framework that combines a theory-grounded definition of collaborative capabilities, controlled manipulation of interaction conditions, and action-level probing of agents' internal states. Experiments across four LLMs show that CollabSim can capture condition effects, separate model performance patterns, and reveal task-dependent effects of agent design.

LLM agents increasingly rely on agent harness: the runtime infrastructure around the base model that defines execution environments, tool interfaces, context, lifecycle orchestration, observability, verification, and governance. Existing self-improving agents and automatic harness evolution methods mainly improve agents through runtime supervision, prompt optimization, workflow search, or harness modification based on final outcomes. However, they often fail to diagnose where the responsible evidence lies in failed trajectories and which harness implementation mechanism causes the unreliable behavior, resulting in broad, indirect, or poorly scoped changes. This paper proposes HarnessFix, a trace-grounded and diagnosis-driven framework for repairing agent harnesses. HarnessFix compiles raw execution traces and harness artifacts into a Harness-aware Trace Intermediate Representation (HTIR), which normalizes fragmented trajectory evidence, captures step-level data-flow and control-flow relations, and aligns runtime steps with the harness artifacts that shape their behavior. It then attributes failures to responsible steps and harness artifacts, and consolidates recurring diagnoses into repair-oriented flaw records. Finally, HarnessFix maps these records to scoped repair operators, generates patches under flaw-specific repair specifications, and accepts them through regression-aware validation. We evaluate HarnessFix on four popular benchmarks, and results show that it improves the performance over the initial harnesses by 6.3% to 18.4%, significantly outperforming human-designed and self-evolution baselines. HarnessFix highlights the value of treating failed trajectories not only as feedback signals, but also as structured evidence for diagnosing and repairing the harness mechanisms behind agent failures.

Large language model agents increasingly rely on external tools, but larger tool menus can reduce reliability and efficiency by increasing wrong-tool calls, premature actions, and token cost. Existing tool-selection methods often optimize semantic relevance, exposing tools whose names or descriptions match the user request. We argue that relevance is insufficient: a tool may be related to the task while still being unnecessary or premature at the current step. We propose Causal Minimal Tool Filtering (CMTF), a training-free method that selects tools by causal sufficiency. CMTF uses lightweight precondition-effect contracts to expose only the minimal next-step tool frontier needed to advance from the current state toward the user goal. Across multi-step tool-use tasks, we compare CMTF with all-tools exposure, keyword retrieval, state-aware filtering, and causal-path ablations, measuring task success, wrong-tool calls, premature actions, tool exposure, and token cost. In the main benchmark with 102 tasks, 100 tools, four LLM backends, and 2448 task-method-model runs, CMTF matches the strongest causal baseline in aggregate success while reducing visible tools from 100 to one per step and reducing token usage by about 90% relative to all-tools exposure.