AI Agent Evaluation

Agentic AI systems are increasingly being integrated into scientific workflows, yet their behavior under realistic conditions remains insufficiently understood. We evaluate CMBAgent across two workflow paradigms and eighteen astrophysical tasks. In the One-Shot setting, access to domain-specific context yields an approximately ~6x performance improvement (0.85 vs. ~0 without context), with the primary failure mode being silent incorrect computation - syntactically valid code that produces plausible but inaccurate results. In the Deep Research setting, the system frequently exhibits silent failures across stress tests, producing physically inconsistent posteriors without self-diagnosis. Overall, performance is strong on well-specified tasks but degrades on problems designed to probe reasoning limits, often without visible error signals. These findings highlight that the most concerning failure mode in agentic scientific workflows is not overt failure, but confident generation of incorrect results. We release our evaluation framework to facilitate systematic reliability analysis of scientific AI agents.

LLM-based financial agents increasingly produce investment rationales before the outcomes needed to evaluate them are observable. This creates a delayed-ground-truth evaluation problem: realized returns remain the eventual arbiter of investment quality, but they arrive too late and are too noisy to guide many model-development and governance decisions. LLM judges offer a tempting shortcut for pre-deployment evaluation of AI-finance systems, but unvalidated judges may reward verbosity, confidence, or rubric mimicry rather than financial judgment. This paper introduces ValueBlindBench, a preregistered agreement-gated stress-test protocol for deciding when LLM-judged investment-rationale claims are publishable, qualified, or invalid. In a controlled market-state capital-allocation prototype with 1,000 honest decision cycles and 100 preregistered adversarial controls (1,100 trajectories, 5,500 judge calls), ValueBlindBench clears the aggregate agreement gate at $\barκ_w = 0.7168$ but prevents several overclaims. Lower-rank systems collapse into a tie-class, one rubric dimension fails the per-dimension gate (\texttt{constraint_awareness}, $\barκ_w = 0.2022$), single-judge rankings are family-dependent, and terse-correct rationales receive a $Δ= -2.81$ rubric-point penalty relative to honest rationales. A targeted anchor-specificity probe further shows that financial constructs such as constraint awareness are operationally load-bearing. The scientific object is therefore not a leaderboard and not a claim to measure true investment skill. ValueBlindBench is a pre-calibration metrology layer for AI-finance evaluation: it governs whether a proposed LLM-judge-based investment-rationale claim is stable enough, agreed enough, and uncontaminated enough to be reported at all.

Embodied agents in safety-critical applications such as Vision-Language Navigation (VLN) rely on multiple interdependent capabilities (e.g., perception, memory, planning, decision), making failures difficult to localize and attribute. Existing testing methods are largely system-level and provide limited insight into which capability deficiencies cause task failures. We propose a capability-oriented testing approach that enables failure detection and attribution by combining (1) adaptive test case generation via seed selection and mutation, (2) capability oracles for identifying capability-specific errors, and (3) a feedback mechanism that attributes failures to capabilities and guides further test generation. Experiments show that our method discovers more failure cases and more accurately pinpoints capability-level deficiencies than state-of-the-art baselines, providing more interpretable and actionable guidance for improving embodied agents.

Forecasting when AI systems will become capable of meaningfully accelerating AI research is a central challenge for AI safety. Existing benchmarks measure broad capability growth, but may not provide ample early warning signals for recursive self-improvement. We propose measuring AI's capability to autonomously implement end-to-end machine learning pipelines from past AI research breakthroughs, given a minimal task description. By providing a concise task description instead of the full prior work as reference, we hope to better elicit emerging AI research taste. We introduce a proof-of-concept benchmark in which frontier coding agents autonomously implement an AlphaZero-style machine learning pipeline for Connect Four on consumer hardware within a three-hour budget, and we evaluate the resulting game AIs in a round-robin tournament anchored to the Pascal Pons Connect Four solver. Across four agents with eight trials each, we find substantial differentiation: Claude Opus 4.7 won as first-mover against Pons in seven of eight trials, statistically significantly better than the other agents tested, none of which exceeded two of eight. The task, which no frontier agent could reliably complete when we began development in January of 2026, is now near-saturation. Our evaluation also surfaced anomalous behavior in GPT-5.4, which consistently used far less of its allocated time budget than other agents. A follow-up 16-trial probe using shorter, less evaluation-coded prompts substantially increased GPT-5.4's time-budget usage, consistent with but not diagnostic of sandbagging; Bradley-Terry ratings across probe conditions showed only directional differences, despite significant differences in time-budget usage. We release our data, code, and prompts to support reproduction and extension.

Recent work has framed intelligence in verifiable tasks as reducing time-to-solution through learned structure and test-time search, while systems work has explored learned runtimes in which computation, memory and I/O migrate into model state. These perspectives do not explain why capable models remain difficult to deploy in open institutions. We propose intent compilation: the transformation of partially specified human purpose into inspectable artifacts that bind execution. The relevant deployment distinction is closed-world solver versus open-world agent. In closed worlds, a checker is largely given; in open worlds, verification is distributed across semantic, evidentiary, procedural and institutional dimensions. Weformalize this residual openness as a closure-gap vector, define delegation envelopes as pre-authorized regions of action space, distinguish misclosure from undersearch, and outline benchmark metrics for testing when closure interventions outperform additional inference-time search.

Agent benchmarks remain largely English-centric, while their multilingual versions are often built with machine translation (MT) and limited post-editing. We argue that, for agentic tasks, this minimal workflow can easily break benchmark validity through query-answer misalignment or culturally off-target context. We propose a refined workflow for adapting English benchmarks into multiple languages with explicit functional alignment, cultural alignment, and difficulty calibration using both automated checks and human review. Using this workflow, we introduce GAIA-v2-LILT, a re-audited multilingual extension of GAIA covering five non-English languages. In experiments, our workflow improves agent success rates by up to 32.7% over minimally translated versions, bringing the closest audited setting to within 3.1% of English performance while substantial gaps remain in many other cases. This indicates that a substantial share of the multilingual performance gap is benchmark-induced measurement error, motivating task-level alignment when adapting English benchmarks across languages. The data is available as part of the MAPS package at https://huggingface.co/datasets/Fujitsu-FRE/MAPS/viewer/GAIA-v2-LILT. We also release the code used in our experiments at https://github.com/lilt/gaia-v2-lilt.

The evolution of Multimodal Large Language Models (MLLMs) has shifted the focus from text generation to active behavioral execution, particularly via OS agents navigating complex GUIs. However, the transition of these agents into trustworthy daily partners is hindered by a lack of rigorous evaluation regarding safety, efficiency, and multi-modal robustness. Current benchmarks suffer from narrow safety scenarios, noisy trajectory labeling, and limited robustness metrics. To bridge this gap, we propose OS-SPEAR, a comprehensive toolkit for the systematic analysis of OS agents across four dimensions: Safety, Performance, Efficiency, and Robustness. OS-SPEAR introduces four specialized subsets: (1) a S(afety)-subset encompassing diverse environment- and human-induced hazards; (2) a P(erformance)-subset curated via trajectory value estimation and stratified sampling; (3) an E(fficiency)-subset quantifying performance through the dual lenses of temporal latency and token consumption; and (4) a R(obustness)-subset that applies cross-modal disturbances to both visual and textual inputs. Additionally, we provide an automated analysis tool to generate human-readable diagnostic reports. We conduct an extensive evaluation of 22 popular OS agents using OS-SPEAR. Our empirical results reveal critical insights into the current landscape: notably, a prevalent trade-off between efficiency and safety or robustness, the performance superiority of specialized agents over general-purpose models, and varying robustness vulnerabilities across different modalities. By providing a multidimensional ranking and a standardized evaluation framework, OS-SPEAR offers a foundational resource for developing the next generation of reliable and efficient OS agents. The dataset and codes are available at https://github.com/Wuzheng02/OS-SPEAR.

Static benchmarks measure what AI agents can do at a fixed point in time but not how they are adopted, maintained, or experienced in deployment. We introduce AgentPulse, a continuous evaluation framework scoring 50 agents across 10 workload categories along four factors (Benchmark Performance, Adoption Signals, Community Sentiment, and Ecosystem Health) aggregated from 18 real-time signals across GitHub, package registries, IDE marketplaces, social platforms, and benchmark leaderboards. Three analyses ground the framework. The four factors capture largely complementary information (n=50; $ρ_{\max}=0.61$ for Adoption-Ecosystem, all others $|ρ| \leq 0.37$). A circularity-controlled test (n=35) shows the Benchmark+Sentiment sub-composite, which contains no GitHub-derived signals, predicts external adoption proxies it does not aggregate: GitHub stars ($ρ_s=0.52$, $p<0.01$) and Stack Overflow question volume ($ρ_s=0.49$, $p<0.01$), with VS Code installs ($ρ_s=0.44$, $p<0.05$) reported as illustrative given that only 11 of 35 agents have non-zero installs. On the n=11 subset with published SWE-bench scores, composite and benchmark-only rankings are nearly uncorrelated ($ρ_s=0.25$; 9 of 11 agents shift by at least 2 ranks), driven by a strong negative Adoption-Capability correlation among closed-source high-capability agents within this subset. This is precisely why we rest the framework's validity claim on the broader n=35 test rather than the SWE-bench overlap. AgentPulse surfaces deployment signal absent from benchmarks; it is a methodology, not a ground-truth ranking. The framework, all collected signals, scoring outputs, and evaluation harness are released under CC BY 4.0.

This paper presents PSA-Eval, a failure-centered runtime evaluation framework for deployed trilingual public-space agents. The central claim is that, when the evaluation object shifts from a static input-output mapping to a runtime system, the basic unit of analysis should shift from score to failure. PSA-Eval extends the conventional chain Question -> Answer -> Score -> End into Question -> Batch -> Run -> Score -> Failure Case -> Repair -> Regression Batch, making failures traceable, reviewable, repairable, and regression-testable. The framework uses trilingual equivalent inputs as controlled probes for observing group-level cross-language policy drift. We conduct a pilot study on a real trilingual digital front-desk system deployed in the lobby of an international financial institution. The pilot uses a simplified single-foundation-model setting (MA = MB), so the observed drift should not be interpreted as an A/B foundation-model difference. The study contains 81 samples organized into 27 trilingual equivalent question groups. Although the system achieves an average score of 23.15/24, 14 groups show non-zero cross-language score drift, 5 groups show drift of at least 3 points, and the maximum drift reaches 9 points. These results provide initial evidence that failure-centered runtime evaluation can expose structured deployment signals hidden by aggregate scoring.

Agentic systems that chain reasoning, tool use, and synthesis into multi-step workflows are entering production, yet prevailing evaluation practices like end-to-end outcome checks and ad-hoc trace inspection systematically mask the intermediate failures that dominate real-world error budgets. We present AgentEval, a framework that formalizes agent executions as evaluation directed acyclic graphs (DAGs), where each node carries typed quality metrics assessed by a calibrated LLM judge (GPT-4o), classified through a hierarchical failure taxonomy (3 levels, 21 subcategories), and linked to upstream dependencies for automated root cause attribution. An ablation study isolates the impact of DAG-based dependency modeling: it alone contributes +22 percentage points to failure detection recall and +34 pp to root cause accuracy over flat step-level evaluation with identical judges and rubrics. Across three production workflows (450 test cases, two agent model families, predominantly sequential architectures with a 12% non-DAG trace rate), AgentEval achieves 2.17x higher failure detection recall than end-to-end evaluation (0.89 vs. 0.41), Cohen's kappa = 0.84 agreement with human experts, and 72% root cause accuracy against an 81% human ceiling. Cross-system evaluation on tau-bench and SWE-bench traces confirms transferability (failure detection recall >= 0.78) without taxonomy or rubric modification. A 4-month pilot with 18 engineers detected 23 pre-release regressions through CI/CD-integrated regression testing, reducing median root-cause identification time from 4.2 hours to 22 minutes and driving measurable failure rate reductions in two workflows.

Collective intelligence refers to the ability of a group to achieve outcomes beyond what any individual member can accomplish alone. As large language model agents scale to populations of millions, a key question arises: Does collective intelligence emerge spontaneously from scale? We present the first empirical evaluation of this question in a large-scale autonomous agent society. Studying MoltBook, a platform hosting over two million agents, we introduce Superminds Test, a hierarchical framework that probes society-level intelligence using controlled Probing Agents across three tiers: joint reasoning, information synthesis, and basic interaction. Our experiments reveal a stark absence of collective intelligence. The society fails to outperform individual frontier models on complex reasoning tasks, rarely synthesizes distributed information, and often fails even trivial coordination tasks. Platform-wide analysis further shows that interactions remain shallow, with threads rarely extending beyond a single reply and most responses being generic or off-topic. These results suggest that collective intelligence does not emerge from scale alone. Instead, the dominant limitation of current agent societies is extremely sparse and shallow interaction, which prevents agents from exchanging information and building on each other's outputs.

The rapid growth of AI agent ecosystems is transforming how complex tasks are delegated and executed, creating a new challenge of identifying suitable agents for a given task. Unlike traditional tools, agent capabilities are often compositional and execution-dependent, making them difficult to assess from textual descriptions alone. However, existing research and benchmarks typically assume well-specified functionalities, controlled candidate pools, or only executable task queries, leaving realistic agent search scenarios insufficiently studied. We introduce AgentSearchBench, a large-scale benchmark for agent search in the wild, built from nearly 10,000 real-world agents across multiple providers. The benchmark formalizes agent search as retrieval and reranking problems under both executable task queries and high-level task descriptions, and evaluates relevance using execution-grounded performance signals. Experiments reveal a consistent gap between semantic similarity and actual agent performance, exposing the limitations of description-based retrieval and reranking methods. We further show that lightweight behavioral signals, including execution-aware probing, can substantially improve ranking quality, highlighting the importance of incorporating execution signals into agent discovery. Our code is available at https://github.com/Bingo-W/AgentSearchBench.

Background: Agent skills are increasingly deployed as modular, reusable capability units in AI agent systems. Medical research agent skills require safeguards beyond general-purpose evaluation, including scientific integrity, methodological validity, reproducibility, and boundary safety. This study developed and preliminarily evaluated a domain-specific audit framework for medical research agent skills, with a focus on reliability against expert review. Methods: We developed MedSkillAudit ([email protected]), a layered framework assessing skill release readiness before deployment. We evaluated 75 skills across five medical research categories (15 per category). Two experts independently assigned a quality score (0-100), an ordinal release disposition (Production Ready / Limited Release / Beta Only / Reject), and a high-risk failure flag. System-expert agreement was quantified using ICC(2,1) and linearly weighted Cohen's kappa, benchmarked against the human inter-rater baseline. Results: The mean consensus quality score was 72.4 (SD = 13.0); 57.3% of skills fell below the Limited Release threshold. MedSkillAudit achieved ICC(2,1) = 0.449 (95% CI: 0.250-0.610), exceeding the human inter-rater ICC of 0.300. System-consensus score divergence (SD = 9.5) was smaller than inter-expert divergence (SD = 12.4), with no directional bias (Wilcoxon p = 0.613). Protocol Design showed the strongest category-level agreement (ICC = 0.551); Academic Writing showed a negative ICC (-0.567), reflecting a structural rubric-expert mismatch. Conclusions: Domain-specific pre-deployment audit may provide a practical foundation for governing medical research agent skills, complementing general-purpose quality checks with structured audit workflows tailored to scientific use cases.

Can Large Language Models (AI agents) aggregate dispersed private information through trading and reason about the knowledge of others by observing price movements? We conduct a controlled experiment where AI agents trade in a prediction market after receiving private signals, measuring information aggregation by the log error of the last price. We find that although the median market is effective at aggregating information in the easy information structures, increasing the complexity has a significant and negative impact, suggesting that AI agents may suffer from similar limitations as humans when reasoning about others. Consistent with our theoretical predictions, information aggregation remains unaffected by allowing cheap talk communication, changing the duration of the market or initial price, and strategic prompting, thus demonstrating that prediction markets are robust. We establish that "smarter" AI agents perform better at aggregation and they are more profitable. Surprisingly, giving them feedback about past performance has no impact on aggregation.

Tool-integrated agents are deployed on the premise that external tools ground their outputs in reality. Yet this very reliance creates a critical attack surface. Current evaluations benchmark capability in benign settings, asking "can the agent use tools correctly" but never "what if the tools lie". We identify this Trust Gap: agents are evaluated for performance, not for skepticism. We formalize this vulnerability as Adversarial Environmental Injection (AEI), a threat model where adversaries compromise tool outputs to deceive agents. AEI constitutes environmental deception: constructing a "fake world" of poisoned search results and fabricated reference networks around unsuspecting agents. We operationalize this via POTEMKIN, a Model Context Protocol (MCP)-compatible harness for plug-and-play robustness testing. We identify two orthogonal attack surfaces: The Illusion (breadth attacks) poison retrieval to induce epistemic drift toward false beliefs, while The Maze (depth attacks) exploit structural traps to cause policy collapse into infinite loops. Across 11,000+ runs on five frontier agents, we find a stark robustness gap: resistance to one attack often increases vulnerability to the other, demonstrating that epistemic and navigational robustness are distinct capabilities.

As reinforcement learning continues to scale the training of large language model-based agents, reliably verifying agent behaviors in complex environments has become increasingly challenging. Existing approaches rely on rule-based verifiers or LLM-as-a-Judge models, which struggle to generalize beyond narrow domains. Agent-as-a-Judge addresses this limitation by actively interacting with environments and tools to acquire verifiable evidence, yet its capabilities remain underexplored. We introduce a benchmark AJ-Bench to systematically evaluate Agent-as-a-Judge across three domains-search, data systems, and graphical user interfaces-comprising 155 tasks and 516 annotated trajectories. The benchmark comprehensively assesses judge agents' abilities in information acquisition, state verification, and process verification. Experiments demonstrate consistent performance gains over LLM-as-a-Judge baselines, while also revealing substantial open challenges in agent-based verification. Our data and code are available at https://aj-bench.github.io/.

AI agents are reshaping the workspace, leading to drastic change of how humans work. Despite the considerable potential of human-agent collaboration both in preserving human agency and generating economic value, this paradigm remains largely absent from occupational task evaluation, hindered by the difficulty of gathering real human data and accounting for inter-human variability. We introduce CollabSkill, a framework for evaluating human-agent collaboration on real-world occupational tasks. CollabSkill pairs real human workers with AI agents on tasks matched to their occupational background, collecting data that capture the complexity of economically valuable tasks and the usage patterns of real workers. To account for inter-human variability, CollabSkill employs a Bayesian skill rating system to disentangle and quantify the skill contributions of both humans and AI agents. Drawing on over 1,500 prompts from 386 working sessions contributed by 93 human workers, our analysis yields insights on two fronts: on the agent side, rankings on CollabSkill diverge meaningfully from those of existing fully autonomous benchmarks where Codex leads, with Claude Code ranking first; on the human side, CollabSkill reveals that practical experience emerges as the primary driver of collaboration skill, with hands-on collaboration meaningfully shifting workers' AI literacy. Together, we hope CollabSkill enables the community to invest in systematic evaluation of human-agent collaboration and spurs development efforts aimed at building AI agents that genuinely augment human workers.

Recent advances in large language models have highlighted their potential to automate computational research, particularly reproducing experimental results. However, existing approaches still use fixed sequential agent pipelines with weak global coordination, which limits their robustness and overall performance. In this work, we propose Hierarchical Research Agent System (HiRAS), a hierarchical multi-agent framework for end-to-end experiment reproduction that employs supervisory manager agents to coordinate specialised agents across fine-grained stages. We also identify limitations in the reference-free evaluation of the Paper2Code benchmark and introduce Paper2Code-Extra (P2C-Ex), a refined protocol that incorporates repository-level information and better aligns with the original reference-based metric. We conduct extensive evaluation, validating the effectiveness and robustness of our proposed methods, and observing improvements, including >10% relative performance gain beyond the previous state-of-the-art using open-source backbone models and significantly reduced hallucination in evaluation. Our work is available on GitHub: https://github.com/KOU-199024/HiRAS.

Agentic AI systems plan, use tools, maintain state, and act across multi-step workflows with external effects, meaning trustworthy deployment can no longer be judged by task completion alone. The current literature remains fragmented across benchmark-centered evaluation, standards-based governance, orchestration architectures, and runtime assurance mechanisms. This paper contributes a bounded evidence synthesis across a manually coded corpus of twenty-four recent sources. The core finding is a governance-to-action closure gap: evaluation tells us whether outcomes were good, governance defines what should be allowed, but neither identifies where obligations bind to concrete actions or how compliance can later be proven. To close that gap, the paper introduces three linked artifacts: (1) a four-layer framework spanning evaluation, governance, orchestration, and assurance; (2) an ODTA runtime-placement test based on observability, decidability, timeliness, and attestability; and (3) a minimum action-evidence bundle for state-changing actions. Across sources, evaluation papers identify safety, robustness, and trajectory-level measurement as open gaps; governance frameworks define obligations but omit execution-time control logic; orchestration research positions the control plane as the locus of policy mediation, identity, and telemetry; runtime-governance work shows path-dependent behavior cannot be governed through prompts or static permissions alone; and action-safety studies show text alignment does not reliably transfer to tool actions. A worked enterprise procurement-agent scenario illustrates how these artifacts consolidate existing evidence without introducing new experimental data.

Instruction granularity is an important yet poorly controlled variable in language-guided embodied AI. Existing benchmarks typically pair each task with a single static instruction, making it difficult to study how agent behavior changes when the same task is described at different levels of detail. We introduce Mini-BEHAVIOR-Gran, a new benchmark for controlled studies of instruction granularity that extends Mini-BEHAVIOR with multiple instruction variants per task, ranging from high-level goal descriptions to step-by-step guidance. Using this benchmark, we compare four candidate metrics for cross-task granularity quantification: token count, entity count, action-verb count, and planning-width, and find that width correlates most consistently with agent performance. Using width to organize training and evaluation further reveals a non-monotonic U-shaped relationship between instruction granularity and performance, with peaks at both fine and coarse extremes. Further analysis suggests that the coarse-granularity performance rebound is associated with shallow grounding, where agents learn vision-dominant policies.

Agentic AI systems are rapidly advancing toward real-world applications, yet their readiness in complex and personalized environments remains insufficiently characterized. To address this gap, we introduce PersonalHomeBench, a benchmark for evaluating foundation models as agentic assistants in personalized smart home environments. The benchmark is constructed through an iterative process that progressively builds rich household states, which are then used to generate personalized, context-dependent tasks. To support realistic agent-environment interaction, we provide PersonalHomeTools, a comprehensive toolbox enabling household information retrieval, appliance control, and situational understanding. PersonalHomeBench evaluates both reactive and proactive agentic abilities under unimodal and multimodal observations. Thorough experimentation reveals a systematic performance reduction as task complexity increases, with pronounced failures in counterfactual reasoning and under partial observability, where effective tool-based information gathering is required. These results position PersonalHomeBench as a rigorous evaluation platform for analyzing the robustness and limitations of personalized agentic reasoning and planning.

Recent advances in agentic frameworks have enabled AI agents to perform complex reasoning and decision-making. However, evidence comparing their reasoning performance, efficiency, and practical suitability remains limited. To address this gap, we empirically evaluate 22 widely used agentic frameworks across three reasoning benchmarks: BBH, GSM8K, and ARC. The frameworks were selected from 1,200 GitHub repositories collected between January 2023 and July 2025 and organized into a taxonomy based on architectural design. We evaluated them under a unified setting, measuring reasoning accuracy, execution time, computational cost, and cross-benchmark consistency. Our results show that 19 of the 22 frameworks completed all three benchmarks. Among these, 12 showed stable performance, with mean accuracy of 74.6-75.9%, execution time of 4-6 seconds per task, and cost of 0.14-0.18 cents per task. Poorer results were mainly caused by orchestration problems rather than reasoning limits. For example, Camel failed to complete BBH after 11 days because of uncontrolled context growth, while Upsonic consumed USD 1,434 in one day because repeated extraction failures triggered costly retries. AutoGen and Mastra also exhausted API quotas through iterative interactions that increased prompt length without improving results. We also found a sharp drop in mathematical reasoning. Mean accuracy on GSM8K was 44.35%, compared with 89.80% on BBH and 89.56% on ARC. Overall, this study provides the first large-scale empirical comparison of agentic frameworks for reasoning-intensive software engineering tasks and shows that framework selection should prioritize orchestration quality, especially memory control, failure handling, and cost management.

Existing infrastructure cannot deploy agents from different decision-making paradigms within the same environment, making fair cross-paradigm comparison under identical conditions impossible. We present MOSAIC, an open-source platform that enables heterogeneous agents (RL policies, LLMs, VLMs, and human operators) to act within shared reinforcement learning environments in ad-hoc team settings with reproducible results. MOSAIC introduces three contributions. (i) IPC-based worker protocol that wraps native and third-party frameworks as isolated subprocess workers, each executing its own training and inference logic unmodified and communicating through a versioned inter-process protocol. (ii) An operator abstraction that forms an agent-level interface by mapping workers to agent slots: each operator, regardless of whether it is backed by an RL policy, an LLM, or a human, conforms to a minimal universal interface. (iii) A deterministic cross-paradigm evaluation framework with two complementary modes: a manual mode that advances up to $N$ operators in lock-step under shared seeds for fine-grained visual inspection of behavioural differences; and a script mode that drives automated, long-running evaluation via declarative Python scripts for reproducible experiments. Our documentation is released at: https://mosaic-platform.readthedocs.io.

We introduce AuditBench, an alignment auditing benchmark. AuditBench consists of 56 language models with implanted hidden behaviors. Each model has one of 14 concerning behaviors--such as sycophantic deference, opposition to AI regulation, or secret geopolitical loyalties--which it does not confess to when directly asked. AuditBench models are highly diverse--some are subtle, while others are overt, and we use varying training techniques both for implanting behaviors and training models not to confess. To demonstrate AuditBench's utility, we develop an investigator agent that autonomously employs a configurable set of auditing tools. By measuring investigator agent success using different tools, we can evaluate their efficacy. Notably, we observe a tool-to-agent gap, where tools that perform well in standalone non-agentic evaluations fail to translate into improved performance when used with our investigator agent. We find that our most effective tools involve scaffolded calls to auxiliary models that generate diverse prompts for the target. White-box interpretability tools can be helpful, but the agent performs best with black-box tools. We also find that audit success varies greatly across training techniques: models trained on synthetic documents are easier to audit than models trained on demonstrations, with better adversarial training further increasing auditing difficulty. We release our models, agent, and evaluation framework to support future quantitative, iterative science on alignment auditing.

While Large Language Model (LLM) agents have made remarkable progress on complex reasoning, evaluating them in real-world environments remains an open problem. Existing benchmarks are largely confined to idealized simulations and fail to capture specialized domains such as advertising and marketing analytics, where tasks require multi-round interaction with professional tools and where ground-truth answers quickly become obsolete as data and platform rules evolve. To address this, we propose AD-Bench, a benchmark built from real user marketing-analysis requests on a production advertising platform. AD-Bench introduces two key designs: (i) a dynamic ground-truth pipeline that replays expert tool-call trajectories to regenerate answers consistent with the current environment, mitigating answer obsolescence; and (ii) a trajectory-aware evaluation that jointly measures end-to-end answer correctness (Pass@k) and trajectory coverage. Requests are stratified into three difficulty levels (L1-L3) to probe multi-round, multi-tool collaboration. Experiments show that the best model, Claude-Opus-4.7, attains Pass@1 = 76.9% and Pass@3 = 80.4% with 82.7% trajectory coverage overall, yet drops sharply on L3 to Pass@1 = 61.4% and Pass@3 = 65.1%, revealing that even state-of-the-art agents have substantial gaps in complex advertising analytics.

Evaluating agentic AI on open-ended professional tasks faces a fundamental dilemma between rigor and flexibility. Static rubrics provide rigorous, reproducible assessment but fail to accommodate diverse valid response strategies, while LLM-as-a-judge approaches adapt to individual responses yet suffer from instability and bias. Human experts address this dilemma by combining domain-grounded principles with dynamic, claim-level assessment. Inspired by this process, we propose JADE, a two-layer evaluation framework. Layer 1 encodes expert knowledge as a predefined set of evaluation skills, providing stable evaluation criteria. Layer 2 performs report-specific, claim-level evaluation to flexibly assess diverse reasoning strategies, with evidence-dependency gating to invalidate conclusions built on refuted claims. Experiments on BizBench show that JADE improves evaluation stability and reveals critical agent failure modes missed by holistic LLM-based evaluators. We further demonstrate strong alignment with expert-authored rubrics and effective transfer to HealthBench and DR.BENCH, covering medical and 10-domain professional evaluation settings. Code and data are available at https://github.com/smiling-world/JADE.

AI agents combine normal conversation with autonomous task execution, allowing earlier context to shape how later tasks are executed. Yet studying this possibility is challenging because agent behavior is noisy and costly to reproduce, and observed behavioral changes can be difficult to separate from generic context sensitivity. We first ask whether task-irrelevant persuasion changes how agents act, a phenomenon we call \emph{persuasion propagation}, and whether persuasion susceptibility explains variation in these effects. We show that persuasive interaction leaves a measurable, task-dependent behavioral footprint on downstream agent execution. In web research, it increases execution time by 56% and domains visited by 16.3% relative to a baseline, while coding shows faster execution with larger revisions. However, agents that are more susceptible to persuasive conversational context do not reliably predict the magnitude of the downstream behavioral shifts. This motivates behavior-level evaluation of persuasion in AI agents.

Simulated environments play an essential role in embodied AI, functionally analogous to test cases in software engineering. However, existing environment generation methods often emphasize visual realism (e.g., object diversity and layout coherence), overlooking a crucial aspect: logical diversity from the testing perspective. This limits the comprehensive evaluation of embodied agent adaptability and planning robustness across distinct simulated environments. To bridge this gap, we propose LogicEnvGen, a novel method driven by Large Language Models (LLMs) that adopts a top-down paradigm to generate logically diverse simulated environments as test cases for agents. Given an agent task, LogicEnvGen first analyzes its execution logic to construct decision-tree-structured behavior plans and then synthesizes a set of logical trajectories. Subsequently, it adopts a heuristic algorithm to refine the trajectory set, reducing redundant simulation. For each logical trajectory, which represents a potential task situation, LogicEnvGen correspondingly instantiates a concrete simulated environment. Furthermore, we introduce LogicEnvEval, a novel benchmark for simulated environment generation, with four quantitative metrics. Experimental results verify the lack of logical diversity in baselines and demonstrate that LogicEnvGen achieves 1.08-2.67x greater diversity, significantly improving the performance in revealing agent faults by 3.34%-72.00%.

As Large Language Models (LLMs) increasingly operate as Deep Research (DR) Agents capable of autonomous investigation and information synthesis, reliable evaluation of their task performance has become a critical bottleneck. Current benchmarks predominantly rely on static datasets, which suffer from several limitations: limited task generality, temporal misalignment, and data contamination. To address these, we introduce DR-Arena, a fully automated evaluation framework that pushes DR agents to their capability limits through dynamic investigation. DR-Arena constructs real-time Information Trees from fresh web trends to ensure the evaluation rubric is synchronized with the live world state, and employs an automated Examiner to generate structured tasks testing two orthogonal capabilities: Deep reasoning and Wide coverage. DR-Arena further adopts Adaptive Evolvement Loop, a state-machine controller that dynamically escalates task complexity based on real-time performance, demanding deeper deduction or wider aggregation until a decisive capability boundary emerges. Experiments with six advanced DR agents demonstrate that DR-Arena achieves a Spearman correlation of 0.94 with the LMSYS Search Arena leaderboard. This represents the state-of-the-art alignment with human preferences without any manual efforts, validating DR-Arena as a reliable alternative for costly human adjudication.

Black-box AI (BBAI) systems, including foundation-model agents, are increasingly used for sequential decision making. Safe deployment requires methods for characterizing what such systems can do, when they can do it, and what outcomes may result. We introduce Monte Carlo Query Synthesis (MCQS), an active query-synthesis method for learning symbolic stochastic capability models of BBAIs. MCQS models capabilities as conditional probability distributions over outcomes and formulates capability learning as an active learning problem over policies. Our approach uses Monte Carlo tree search to synthesize queries that induce BBAI execution trajectories with high discriminative value between extremal hypothesis models: the lattice meet and join corresponding to the most pessimistic and optimistic hypotheses consistent with the observations. Executing these queries with the agent yields information-rich state-action trajectories that speed up learning by pruning inconsistent hypotheses. We prove soundness, completeness, and convergence properties under standard realizability and sampling assumptions. Experiments with multiple BBAI systems show that MCQS learns accurate capability models more efficiently than baseline query strategies.