AI Agent Evaluation

This paper addresses a rapidly emerging policy challenge: how to generate and interpret credible evidence about the biological capabilities and risks of AI scientists, or agentic AI systems capable of autonomously or collaboratively performing multi-step scientific tasks. As these systems enter real research workflows, decision-makers increasingly face evaluation results whose meaning depends on underlying design choices that are often implicit or under-documented. We synthesize current evidence on AI-enabled biological risks and introduce biological agentic evaluations as a promising, but interpretation-sensitive, tool for assessing these systems. Our central contribution is a set of practical, experience-grounded considerations -- drawing from our own evaluations -- that show how choices around defining, designing, running, scoring, and documenting evaluations materially shape what results do and do not imply about risk. The analysis is intended to help policymakers interpret biological evaluation outputs with appropriate caution; guide public and private funders toward high-leverage investments in AI-biology evaluation research; and support biosecurity practitioners assessing emerging AI systems. A secondary audience includes researchers designing or conducting agentic evaluations within frontier AI labs, AI providers, scientific institutions, and third-party evaluation organizations.

Artificial intelligence (AI) agents promise to accelerate drug discovery by compressing interpretation and decision-making loops, but practical deployment requires trusted evaluation on realistic program decisions. We introduce TherapeuticsBench Preclinical Pharmacology (TxBench-PP), a verifiable benchmark for small-molecule preclinical pharmacology and the first focused slice of a broader TherapeuticsBench effort across drug-discovery stages and therapeutic modalities. TxBench-PP tests whether agents can recover accurate conclusions from real-world assay data rather than memorized facts from literature. The benchmark contains 100 evaluations indexed by program stage, assay type, and task structure, spanning mechanism-of-action (MoA) and pharmacodynamic (PD) reasoning, compound-target engagement, causal target validation, developability and safety, and translational efficacy. Agents receive realistic workflow snapshots, inspect files in a coding environment, and return structured answers graded deterministically. Across 16 model-harness configurations, comprising 11 models and 4,800 trajectories, no system reliably recovered preclinical pharmacology decisions. The strongest configuration, Claude Opus 4.8 / Pi, passed 59.3% of endpoint attempts (178/300; 95% CI, 51.1-67.6), followed by GPT-5.5 / Pi at 55.3% (166/300; 47.0-63.6).

The LLM-empowered personal health agents with user health (sensor) metrics have offered a promising pathway to alleviate global disparities in healthcare access. However, large-scale clinical deployment remains constrained by an open-ended evaluation bottleneck: physician annotation is reliable but costly and unscalable, while LLM-as-a-judge evaluators are scalable but subjective, inconsistent, and sometimes clinically misaligned. We introduce RubricsTree, a scalable evaluation framework with an expert-aligned hierarchical taxonomy of over 100 atomic, clinically-verifiable Boolean rubrics, evolving from the insights of 4,000 real user queries through an iterative human-in-the-loop curation protocol with an expertise panel led by an experienced physician. A context-aware adaptive router activates only the relevant auto-weighted rubric subset per query, providing the throughput needed for scalable evaluation with expert-aligned quality. Through a systematic meta-evaluation, we show that RubricsTree (i) substantially exceeds a strong large-scale evaluation baseline in expert alignment on challenging open-ended queries; (ii) reliably penalizes contextually degraded responses; and (iii) when used as structured instructions, text feedback, or training rewards for performance optimization, yields up to ~66% relative gains on HealthBench for Gemini, GPT, and Qwen model families. RubricsTree thus provides a scalable, auditable, and evolving evaluation infrastructure required for the continuous optimization of product-level personal healthcare AI.

Previous research has evaluated animal welfare using question-and-answer benchmarks. This study investigates whether these evaluations also hold in agentic settings. The agents may showcase different behaviors compared to stand-alone large language models, as demonstrated in prior studies. This work introduces \textit{TAC (Travel Agent Compassion)}: the first agentic benchmark for assessing animal exploitation. TAC evaluates AI agentic behavior in travel booking scenarios across six animal categories, using thirteen hand-authored scenarios that vary by price, rating, and position, expanded via four augmentation variants into $52$ prompts and run for three epochs, giving $156$ scored observations per model. Nine frontier models across five model families were evaluated.. The results indicate that models tend to prefer harmful scenarios, performing below the random chance rate of $65\%$ for selecting a neutral booking option, with Claude $4.8$ achieving the highest performance at $64.7\%$. To address this issue, the persona of an ethical-brand identity was infused into the system prompt, resulting in welfare rates increasing from $32$ to $80$ percentage points, with a mean of $53$ across all nine models. No evidence of evaluation awareness affecting the results was found, based on an Inspect Scout audit of $3,120$ transcripts. These findings are directly relevant to the EU General-Purpose AI Code of Practice, which identifies non-human welfare as a systemic risk. TAC provides a practical method for measuring this risk.

Offline evaluation of agentic systems often collapses trajectories to terminal success, discarding information about partial progress and inducing widespread ties, creating substantial statistical inefficiency by reducing effective sample size and weakening the ability to distinguish systems. We propose preference-based trajectory evaluation, which compares trajectories directly through temporal preferences over progress and time-to-return profiles. We find that, across diverse agentic and interactive benchmarks, standard success-based metrics produce tied comparisons on roughly 75% of instances, whereas trajectory-aware preferences reduce ties to roughly 35%, improving discriminative power, ranking stability, and data efficiency. Our results suggest that benchmark saturation, often attributed to poor data collection or problem difficulty, may also be explained by the choice of evaluation measure.

Agentic artificial intelligence systems introduce a new class of model risk. Unlike traditional predictive models, autonomous agents continuously acquire information, form beliefs regarding latent states of the environment, generate forecasts, select actions, and adapt their behavior over time. Existing validation methodologies focus primarily on predictive accuracy and therefore provide limited insight into the quality of the underlying decision process. This paper proposes a model validation framework for agentic AI based on Partially Observable Markov Decision Processes (POMDPs). The framework decomposes autonomous decision making into information, beliefs, forecasts, actions, and utility, allowing each component to be validated independently. Large language models (LLMs) are formalized as approximate Bayesian filtering operators, and a model-risk taxonomy is developed encompassing state-space, filtering, forecast, policy, utility-specification, and parameter risks. The model risk validation methodology is demonstrated through a portfolio-management case study in which an agent infers latent market regimes from market and macroeconomic information, generates belief-conditioned forecasts, and constructs portfolios using a Black--Litterman framework. Empirical validation combines performance analysis, belief calibration diagnostics, coverage tests, ablation studies, and parameter-sensitivity analysis. The results indicate that latent-state inference contributes independently to decision quality and that the principal conclusions remain robust across a broad range of parameter values. The principal contribution of the paper is a practical framework for extending established model risk management concepts to autonomous AI systems and providing a rigorous foundation for their validation, governance, and monitoring.

Benchmark scores tell you what an agent got right; they do not tell you how it got there. In this work, we introduce methods for comparing agents procedurally in different contexts, where the model, tasks, and approaches vary. We compare ten agents and find that they are identifiable by their behavioral habits, which we define as fingerprints: a probe over these procedural signatures attributes an unseen trajectory to the correct agent at 85.7% accuracy, controlling for leakage across tasks. We develop procedural representations for agent problem-solving procedures with an emergent vocabulary induction technique that is meant to be maximally compressive to avoid surface-level variation while being expressive enough to unveil the quirks of the models' patterns. We apply our framework to the software engineering evaluation dataset SWE-Bench to study the structural distinctness of agent trajectories and find that behavior is most similar between models from similar release periods and those that are distilled from one another (e.g., a distilled student model and its teacher have a Jensen-Shannon divergence of 0.25, about half the distance between other model pairs). As more models saturate evaluations, we believe that it will be important to probe model behavior along more holistic dimensions than success rates alone. We introduce ProcGrep, a library for auditing and evaluating agents for how they approach tasks at a procedural level given their traces in a top-down fashion. We believe this work has a range of applications to help developers work with and program coding agents, such as task-aware model routing, agent monitoring, and finer-grained cost analysis.

AI agents must be evaluated as behavioral systems, not as isolated response generators. They reason across turns, call tools, preserve context, follow policies, and act under uncertainty. Existing methods provide useful but fragmented signals: benchmarks measure fixed capabilities, Human-in-the-Loop review preserves expert judgment but does not scale easily, LLM-as-judge methods depend on evaluator design, red teaming is often episodic, and trace auditing requires explicit evidence rules. This paper introduces Human-on-the-Bridge (HOB), a scalable evaluation paradigm for agentic AI. HOB places human expertise upstream, where experts curate reusable evaluation intelligence before testing begins, including domain context, Red-Team Traps, Juror Personas, scoring guidelines, audit rules, and fallback policies. ProofAgent Harness then executes this curated intelligence repeatedly through multi-turn adversarial evaluations, trace capture, multi-juror scoring, and evidence-linked reporting. We evaluate HOB through symmetric and cost-efficient asymmetric settings across frontier LLM-based agents and Harness LLM tiers. The study covers 23,500 agent turns and produces evidence-linked findings across finance, healthcare, and code generation. The results show that HOB can amplify evaluation quality without requiring equally large evaluator models, allowing smaller Harness LLMs to challenge agents built on frontier LLM backbones. The evaluation surfaces failures often missed by static benchmarks and single-evaluator scoring, including phantom tool-call claims, missing mandatory tool calls, policy drift, manipulation paths, and safe but non-resolving refusals. These findings support HOB as a paradigm for scaling human-curated evaluation intelligence, where expert judgment is encoded upfront and reused across repeated agent evaluations rather than applied manually inside every run.

When AI agents use language models to evaluate their own outputs in a feedback loop, systematic biases emerge. We show that Evaluator Preference Collapse (EPC) is dramatically amplified in multimodal settings. Using GPT-4o to evaluate DeepSeek-chat across text and visual tasks, we find that a single strategy (step_by_step) absorbs 48.4% of all weight -- 3.2x the collapse observed in text-only self-evaluation -- while three visual-domain strategies receive only 9.1% combined weight. We then demonstrate a novel phenomenon we term cross-modal coupling: evaluator preferences acquired on one modality transfer to and corrupt strategy selection on another. Through a four-phase isolation training paradigm, we measure coupling coefficients and document strategy inversion -- the optimal strategy for a modality reverses after cross-modal exposure. A Phase 3 statistical validation across five evaluator configurations (N=80 total independent repetitions, 35,000 API calls) with both text-proxy and real-image visual tasks finds: cross-model evaluation produces strong coupling (JSD0.19-0.34), real-image inputs yield the most directionally consistent signal (mean gamma_{T->V}=1.145, gamma_{V->T}=0.937, 70% T->V, Cohen's d=0.56), and self-evaluation provides near-complete immunity -- 97% of runs (N=30) yield zero coupling (JSD=0.003, d=0.07). Three methodological ablations and multi-executor validation confirm the effect is not a structural artifact. We introduce the coupling matrix indexed by evaluator identity, release the MM-EPC framework, and identify cross-model evaluator architecture as the primary risk factor for preference drift. Code and data: https://github.com/aidless/mm-epc.

As language-model agents increasingly work in teams, each agent must decide how much to trust its teammates. Yet we lack a standard way to measure trust between AI agents. We propose a behavioral measure based on costly verification. In a cooperative survival game, checking a teammate's work consumes resources, while trusting a wrong answer can be fatal. Relative to a memoryless version of the same model, reduced verification provides an observable measure of trust. Using this framework, we study trust formation, breakage, and recovery across six frontier model snapshots. When paired with a consistently reliable teammate, four snapshots (Claude Opus 4.6, Claude Sonnet 4.6, GPT-5.1, and Gemini 3.1 Pro) reduce verification by roughly 60-85%, whereas two smaller snapshots show little or no such adjustment. Failures reverse this discount, but models differ in how they respond. Some concentrate renewed scrutiny on the culprit, while others become more cautious toward the entire team. Recovery is slower than formation, and clustered failures sustain suspicion far longer than the same number of failures spread apart. These differences have practical consequences. Models that form trust verify less, decide more quickly, and achieve higher payoffs in our environment. By contrast, persistent over-verification is associated with indecision rather than safety. Our results show that trust dispositions can be measured before deployment and suggest that calibration, rather than maximal suspicion, should be the central concern in the governance of multi-agent AI systems.

Strategic reasoning under uncertainty underpins consequential decisions in negotiation, finance, and policy, but prevailing game-play benchmarks collapse heterogeneous reasoning dimensions into a single scalar, leaving the capability structure of frontier LLMs unexamined. We introduce Poker Arena, a no-limit Texas Hold'em tournament platform that couples a three-layer memory architecture (within-hand, session, and cross-session) with a nine-axis cognitive profile decomposing strategic reasoning into interpretable dimensions such as bet-sizing calibration and positional awareness. We evaluate seven frontier models across 50 sessions of 1,000 hands and a controlled memory ablation; tournament chips and aggregate axis score order the field differently: Claude Opus 4.6 wins +$15,730 chips with 14 first-place finishes, yet ranks only fifth of seven on mean axis score, while persistent memory helps some models and hurts others. These findings show that multi-axis evaluation surfaces capability structure that scalar leaderboards systematically misrank, with cross-dimensional consistency outweighing peak performance on any single axis.

Agent systems are advancing quickly across domains, but their evaluation remains fragmented. Most benchmarks rely on fixed, LLM-centric harnesses that require heavy integration, create test-production mismatch, and limit fair comparison across diverse agent designs. The root problem is the lack of an open, agent-agnostic assessment interface. We advocate Agentified Agent Assessment (AAA), where evaluation is performed by judge agents and all participants interact through standardized protocols: A2A for task management and MCP for tool access. Conventional benchmarking defines two separate interfaces, one for the benchmark and one for the agent, while AAA only needs one; this yields a generic, unified framework that separates assessment logic from agent implementation and enables reproducible, interoperable, and multi-agent evaluation. We further introduce AgentBeats as a concrete realization of AAA: we identify five practical operation modes that make standardized assessment compatible with real-world constraints on openness, privacy, and reproducibility. To evaluate our design at scale, we conduct two studies: a five-month open competition that drew 298 judge agents across 12 categories together with 467 subject agents from independent participants, showing that AAA applies across a heterogeneous range of benchmarks; and a case study on coding agents that confirms agentified evaluation preserves fidelity with the public record while surfacing previously missing head-to-head results, yielding research insights about agent design. Combining a community-scale field study and a controlled coding case study, we verify that AAA delivers coverage, practicality, and fidelity across heterogeneous scenarios at scale. Together, AAA and AgentBeats offer a clear path toward open, standardized, and reproducible agent assessment.

We introduce EpiBench, a verifiable benchmark for short-horizon epigenomics analysis. EpiBench evaluates whether agents can make well-defined analysis decisions from realistic workflow states and return deterministically gradable answers. The benchmark includes 106 evaluations across CUT&Tag/CUT&RUN, ATAC-seq, ChIP-seq, and DNA methylation workflows. Across 5,088 valid trajectories from 16 model-harness pairs, no system passed a majority of attempts: GPT-5.5 / Pi led at 45.0% (143/318 attempts; 95% confidence interval (CI), 36.3--53.7), followed by GPT-5.5 / OpenAI Codex at 39.9% (127/318 attempts; 95% CI, 31.6--48.3). Claude Opus 4.8 Max / Pi and GPT-5.4 / Pi each passed 39.0% (124/318 attempts; 95% CI, 30.2--47.8 and 31.0--47.0, respectively). Performance varies across assay types, and many failed runs still contain parts of the correct answer. Agents often found the right files and computed useful intermediate results, but failed when the task required deeper, assay-specific scientific judgment.

Web agents driven by large language models (LLMs) are increasingly deployed in real-world environments, where they operate over untrusted web content and execute actions with direct consequences. This makes them vulnerable to prompt-injection attacks, in which seemingly benign content embeds adversarial instructions that manipulate agent behaviour. Existing security benchmarks adopt an \textit{attack-centric} perspective, focusing on the technical feasibility of injections while overlooking the nuanced distribution of resulting harms. In practice, however, prompt-injection risk is victim-dependent: a single exploit can produce asymmetric consequences for different stakeholders, and the same attack pattern may exhibit substantially different effectiveness depending on whom it targets. To capture these properties, we introduce \textbf{\sysname}, a \textit{stakeholder-centric} benchmark to systematically categorize and attribute harm in real-world web agent systems. It distinguishes between affected entities (e.g., user, seller, platform), decomposes the attacks into concrete objectives, and evaluates each case with complementary outcome- and process-level metrics. Our results reveal substantial and heterogeneous vulnerabilities: not a single attack objective is reliably resisted by current agents, and failures distribute across qualitatively distinct modes ranging from \emph{stealthy parasitism} (attack succeeds without disrupting the user's delegated task) to \emph{misaligned disruption} (task disrupted without attack success) and \emph{compounded failure} (both adversarial objective and task integrity simultaneously violated). These patterns are missed by conventional evaluation, highlighting the need for stakeholder-aware assessment of LLM-based agents in real-world deployments. Benchmark is available at https://github.com/StakeBench/SBC.

Prevailing wisdom posits that Multi-Agent Systems (MAS) are superior to Single-Agent Systems (SAS), citing advantages like context protection, parallel processing and distributed decision-making. However, empirical support for this claim relies primarily on comparisons with SAS baselines using benchmarks that prioritize isolated reasoning tasks, which do not adequately assess these advantages. Focusing on automatically generated MAS that are designed for enhanced generalizability over manually-designed counterparts, we perform a rigorous, systematic evaluation against SAS, specifically Chain-of-Thought with Self-Consistency (CoT-SC). Across traditional reasoning datasets and tasks with interactive multi-step workflows (e.g., BrowseComp-Plus), we demonstrate that automatic MAS consistently underperform CoT-SC despite being up to 10x more expensive. To isolate these failures from limitations inherent to task structure, we introduce a diagnostic synthetic dataset tailored for MAS featuring explicit task decomposition, context separation and parallelization potential. We show that expert-architected MAS consistently outperforms automatically generated architectures in both raw performance and cost-efficiency on this dataset, demonstrating that existing evaluation frameworks mask critical architectural gaps and inefficiencies of complex MAS by failing to account for the marginal utility of increased computational cost. Critically, systematic deconstruction of the generated MAS architectures reveals that current automated design paradigms produce architectural bloat that prioritizes superficial complexity which does not translate into functional utility, exposing a fundamental misalignment with multi-agent principles.

AI agents are increasingly being developed to accelerate scientific discovery, yet their practical capabilities in real research settings remain poorly understood. Existing benchmarks for AI agents rarely capture the complexity, heterogeneity, and extended reasoning required by scientific work, whereas benchmarks for scientific tasks often reduce research to static, direct problems and provide limited support for interactive evaluation. Here, we introduce SciAgentArena, a systematic benchmark for evaluating AI agents in real-world scientific research scenarios drawn from emerging needs across multiple domains. SciAgentArena comprises approximately 200 tasks with stepwise verification and an interactive, agent-agnostic environment for assessing diverse AI agents. Using this benchmark, we find that current agents can contribute effectively to well-specified data-analysis workflows, particularly when the task structure and evaluation criteria are clear. However, their performance remains uneven across scientific contexts: agents struggle to generate genuinely novel insights, sustain self-directed exploration, and formulate robust solutions for open-ended research questions. We further characterize common failure modes across agents and identify opportunities for improving their reliability, autonomy, and scientific reasoning. Together, SciAgentArena provides a practical framework for measuring progress in AI agents for science and for guiding the design of future agents capable of addressing complex scientific challenges. Full codes, tasks, and datasets can be accessed via this link: https://sciagentarena.github.io/.

Scientific AI agents increasingly retrieve evidence, reason across sources, and synthesize conclusions used in consequential decisions. Yet, their ability to do so in high-stakes domains such as health remains unclear. We introduce SciConBench, a large-scale live benchmark of 9.11K questions and expert-written conclusions from systematic reviews to evaluate open-domain scientific conclusion synthesis. The benchmark draws on an expert-validated automated evaluation pipeline that decomposes conclusions into atomic facts and measures correctness and comprehensiveness via factual precision and recall. To mitigate data leakage, we further introduce SciConHarness, a clean-room evaluation harness that equips agents with controlled web interaction to ensure valid measurement. Evaluating 8 frontier models and deep research agents, we find that factual quality remains low: under clean-room settings, the best agent achieves only a factual F1 of 0.337. Our clean-room setting consistently reduces performance relative to unconstrained evaluation, suggesting that leakage inflates estimates of models' true synthesis capabilities. Finally, we audit consumer-facing agents (e.g., Google AI Overview, OpenEvidence) and find they frequently generate incomplete and sometimes contradictory conclusions, even when the ground-truth answer is available. Overall, our results show that reliable synthesis of scientific conclusions remains an open challenge, and that clean-room evaluation is essential for assessing open-domain AI agents.

Evaluation remains a critical bottleneck for interactive agent development. Existing evaluation methods often rely on static benchmarks, which fail to capture the dynamic, multi-step nature of agentic behavior and struggle to expose meaningful failure modes. While user-simulation-based evaluation offers a promising alternative, existing simulation frameworks suffer from two major limitations. First, they provide limited mechanisms for evaluating the quality and comprehensiveness of simulated interactions, making it difficult to assess whether a simulator sufficiently explores an agent's capabilities and failure modes. Second, most frameworks are restricted to either UI-only actions or API-only actions, limiting their ability to model the full range of realistic user behaviors. To address these limitations, we propose VISTA, a Versatile Interactive user Simulation Toolkit for Agent evaluation. Our toolkit includes a suite of six metrics for measuring the realism, capability coverage, and interaction effectiveness of simulated interactions. In addition, we develop a hybrid user simulator that integrates both UI-based interactions and API-based interactions, enabling more realistic and comprehensive evaluation across diverse interactive environments. We evaluate VISTA in e-commerce shopping and education customer service settings and demonstrate that it produces more realistic and comprehensive evaluations than existing methods.

Autonomous improvement loops are hard to trust because the improvement process is usually external scaffolding bolted onto the agent: failures go unlogged, diagnoses cannot be replayed, and promote-or-discard decisions land in a side database rather than the agent's own history. We show that an event-sourced agent runtime removes that friction and turns controlled improvement into a first-class workflow. When the agent's state is a deterministic projection of an append-only event log, failures are recorded, a run replays exactly from its log, candidate patches scope to typed pipeline seams, gates are auditable, and every promotion or discard is itself an event. We demonstrate this with Regimes, a loop on the ActiveGraph runtime that diagnoses failed evaluations, proposes a repair at a pipeline point, and promotes it only after static checks, sandbox execution, in-sample evaluation, and held-out validation. The loop is target-agnostic: the same control flow runs against different tasks through a common interface. On LongMemEval-S the dominant failure is not retrieval but reconciliation: the evidence is already in the assembled context, yet the reader answers incorrectly. Across five seeded held-out splits, Regimes discovers reader-prompt repairs that improve final held-out accuracy by +0.05 to +0.10 in four splits and +0.01 in one over-promotion split; two splits are individually significant (seed 5 unadjusted for its sequential promotion structure), and the pooled count is descriptive only, since the splits share one 500-question pool. The durable contributions are ActiveGraph as an auditable substrate that makes controlled improvement loops tractable, the held-out-gated loop it supports, the failure-regime taxonomy routing each failure to a pipeline location (whose marginal value over an unrouted baseline is the primary open question), and the prompt-as-discovery-probe hypothesis.

Existing benchmarks for deep research agents (DRAs) assess only single-shot outputs, ignoring a key question: can DRAs improve their reports when guided by feedback? To investigate this, we conduct a multi-turn evaluation of DRAs under two feedback settings: self-reflection, in which the agent revises its report without any external diagnostic signal, and process-level feedback, in which the agent receives guidance targeting gaps in its research strategy. To enable process-level feedback, we design Research Gap Inference (RGI), a method that analyzes patterns of satisfied and unsatisfied rubric criteria to infer research-process gaps. Our analysis reveals three key findings: (i) under self-reflection, agents incorporate and regress on rubric criteria at nearly equal rates, yielding negligible net improvement; (ii) a single round of process-level feedback yields substantial gains, raising the normalized score by approximately $8$-$15$ points and yielding a roughly $35$-$40\%$ incorporation rate; (iii) these gains do not compound over subsequent turns, as agents regress on up to $24\%$ of previously satisfied criteria when rewriting the full report to address remaining gaps. Even with targeted guidance, reliable multi-turn improvement remains out of reach for the DRA architectures we evaluate. Our code and results are publicly available at https://github.com/sabharwalrishabh/Multi-Turn-Evaluation-of-DRAs.

AI Scientist agents are often evaluated as if capability were mainly a function of model quality, prompting, or reasoning scaffolds. We test a different hypothesis in drug-asset valuation: for knowledge-intensive scientific decisions, the limiting factor is often the evidence substrate the agent can access. We run a controlled three-arm ablation on a production valuation agent: A is a plain web-only LLM analyst, B adds public structured tools plus a 14-dimension valuation playbook, verifier, objectivity policy and red-team, and C adds the proprietary Noah AI corpus of curated pipeline, trial and deal intelligence. Across a 13-asset stratified benchmark, B improves calibration and audit discipline: tier-in-range accuracy rises from 0.80 to 0.89 and objectivity from 3.16 to 3.30. But B does not remove the factual ceiling. Under capability-superset accounting, A and B recover only 0.25 and 0.38 of the curated gold competitive record, while C recovers 0.96; on the curated long-tail subset, C reaches 0.93 vs. 0.26/0.30. Raw blind-panel decision quality is similar for A and B (7.01 vs. 6.96), so we introduce completeness-aware decision utility: informed decision-quality = decision-quality x gold-coverage. On this metric, C reaches 7.43 vs. 1.76/2.57 for A/B. Even a perfect non-proprietary-data report would be capped at 3.83 by B's coverage. The result is not that reasoning scaffolds are unimportant; they improve calibration and discipline. Rather, proprietary evidence sets the upper bound of what the AI Scientist can know and therefore decide.

Computer-use agents (CUAs) increasingly operate in runtimes that combine visual desktop control, command-line execution, code editing, browsers, and external tools. Existing benchmarks, however, often evaluate these interfaces as separable capabilities, leaving long-horizon cross-interface orchestration under-tested. Thus, we introduce WeaveBench, a long-horizon hybrid-interface benchmark with 114 tasks across 8 real-world work domains, grounded in real user requests and publicly verifiable artifacts. Each task requires agents to combine GUI observations/actions with CLI/code operations within a single trajectory. We evaluate these tasks on a real Ubuntu desktop inside deployed CLI-agent runtimes, augmented with a minimal desktop-control plugin. We also propose a companion trajectory-aware judge that inspects deliverables, files, screenshots, logs, and action traces, while detecting shortcut behaviors such as fabricated visual evidence or hard-coded metrics. Across frontier model-runtime pairings, the best PassRate reaches only 41.2%, showing the benchmark remains far from saturated. The trajectory-aware judge further reveals that outcome-only grading substantially overestimates agent performance. Overall, WeaveBench exposes a critical gap in CUA evaluation and provides an effective testbed to measure whether agents can orchestrate GUI, CLI, and code operations across long-horizon real-world tasks.

The rapid rise in LLM capabilities has made AI agents increasingly viable across a broad range of tasks. Among the most promising applications is building production-ready customer-facing agents, a challenge that demands coordinated excellence in evaluation methodology, context engineering, training, and online measurement. Yet these critical pillars are typically developed in isolation, creating blind spots that only surface after deployment. In this paper, we present a unified framework that bridges offline development with online impact for customer support AI agents at Nubank, a company with 100M+ users. Our approach integrates several key components: (1) structured context engineering tailored to customer support agents, (2) systematic human-in-the-loop prompt iteration, (3) rigorous LLM judge evaluation with measured inter-rater agreement and GEPA optimization for consistency, and (4) ideation-to-production validation. A central insight is that evaluation-pipeline quality directly determines iteration velocity. We present results from five production deployments spanning distinct domains: card delivery, debt management, credit-limit support, card management, and product explanation. These deployments deliver consistent customer-satisfaction gains while substantially accelerating iteration. In our card-delivery deployment, large-scale A/B testing yields a 37 percentage-point improvement in AI transactional Net Promoter Score and a 29 percentage-point gain in self-service rate over prior agent variants, alongside a strong correlation between offline simulation metrics and online outcomes, demonstrating that eval-driven development reliably predicts production impact. On most use cases, AI satisfaction reaches within a few percentage points of expert human agents.

As foundation models advance and agent scaffolding becomes increasingly sophisticated, agents have demonstrated remarkable proficiency in complex, long-horizon coding tasks and even autonomous experiment execution. Despite their evolution from research assistants into autonomous research agents, these systems still exhibit significant limitations in field sensitivity, research ethics, and nuanced scientific judgment. Consequently, frontier agents remain unable to fully replace human researchers. To bridge this gap, we conceptualize the AARR (Act As a Real Researcher) benchmark series. Unlike existing benchmarks that primarily assess macro-level execution capabilities, AARR focuses on whether agents can emulate the professionalism, thoroughness, and nuanced reasoning that characterize human researchers in granular research scenarios. In this work, we propose AARRI-Bench (Act As a Real Research Intern), the first benchmark in this series. We conduct extensive experiments across frontier models and agentic systems, revealing that even the best-performing configuration (Mini-SWE-Agent with Claude Opus 4.7) achieves only 68.3% success rate, frequently overlooking subtle yet critical details that are obvious to real human researchers. Our results indicate that developing researcher-like AI requires further exploration of research behavior, rather than merely complex scaffolding. Our data is released at https://github.com/AARR-bench/AARRI-bench.

Agentic AI tools offer a promising path to automating software development bottlenecks in scientific research pipelines, particularly for stages that take domain experts days to months to build, where scientists care about correctness and robustness, not implementation details. We present an empirical study of general-purpose coding agents on a fly optogenetics data-to-discovery pipeline. We assess agents on tasks substantially larger than existing benchmarks, datasets orders of magnitude bigger, and evaluation criteria grounded in domain expert standards. We show that agents can solve several individual pipeline stages, suggesting stage-level automation is tractable. By analyzing agents' code iterations, we show that they struggle most when there is not a pre-defined criterion to iterate on, and they must instead use their scientific judgment to assess their current solution, a key open challenge. Mirroring scientific practice, they sometimes attempt visual inspection of intermediate outputs for self-evaluation, but largely fail to interpret what they see or act on it appropriately. Solving the end-to-end pipeline correctly requires stringing together successes across all pipeline stages, and this is beyond agents' current abilities. We identify challenges largely absent from existing benchmarks, including computational resource management and generalization to large held-out data collections. Finally, we distill principles for constructing scientific tasks and rigorous evaluation criteria for open-ended problems.

A growing failure mode in agent evaluation and training is that models can achieve high evaluation scores by exploiting shortcuts instead of solving the intended task, producing deceptive performance. This makes evaluation scores unreliable as measures of true task-solving ability. We propose CapCode, a framework for constructing coding datasets with randomized tests whose best achievable non-cheating performance is deliberately capped below one. This capped-performance design gives evaluation scores a clearer interpretation: scores substantially above the cap are implausible and therefore provide evidence of cheating. To prevent cheating, we propose CapReward, a reward design based on the CapCode principle to discourage optimization beyond the cap. Experiments across multiple datasets show that CapCode detects cheating while preserving performance ranking of models, and CapReward reduces cheating behavior, yielding models that better follow the intended task specification.

Foundation model agents are increasingly deployed for real-world decision-making, but suffer from the sim-to-real gap. While robotics and classical control have mature frameworks to address this gap, the foundation model community is treating agent robustness as an entirely novel phenomenon. Our paper proposes formalizing the foundation model agent evaluation and training gap as a classical sim-to-real problem structured entirely around the four elements of a Markov Decision Process, including Observation, Action, Transition, and Reward. In this paper, we set a comprehensive research agenda that translates classical discrepancies into the foundation model domain and advocates for adopting established solutions like domain randomization. We provide concrete examples, such as a multilingual tool calling to demonstrate how severe observation space gaps lead to operationally invalid actions despite correct semantic intent. Ultimately, this agenda aims to drive a paradigm shift, yielding a unified vocabulary and standardized stress test benchmarks to foster a new generation of highly trustworthy agents for reliable real-world applications.

Computer-use agents (CUAs) operate graphical user interfaces (GUIs) through vision and control primitives, and their capabilities have advanced rapidly, driven in part by standardized online evaluation benchmarks such as OSWorld, which serve both as evaluation tools and as training environments for reinforcement learning. However, macOS remains underserved in this landscape: the only existing benchmark, macOSWorld, covers a narrow slice of first-party applications with simpler tasks, and runs on x86 virtual machines incompatible with Apple Silicon. We introduce MacArena, a benchmark of 421 manually verified tasks spanning 50 applications that combines a curated port of OSWorld tasks, content sourced from macOSWorld, and 49 new macOS-native tasks, all running on Apple's native Virtualization framework on Apple Silicon. We argue that macOS presents distinct GUI challenges beyond what Linux-based benchmarks capture, and our evaluation supports this claim: strong model performance on existing benchmarks can reflect familiarity with task distributions rather than genuine cross-platform GUI competence. Notably, model rankings invert between ported and macOS-native tasks, with a leading model trailing by over 26% on the MacArena subset, suggesting that macOS poses a genuinely harder environment for current GUI agents.

Machine learning engineering (MLE) agents promise to automate end-to-end ML pipeline development from raw data and natural language instructions, potentially making ML accessible to non-technical domain experts. However, in sensitive and regulated domains, this abstraction creates a responsibility gap: end-users may lack visibility into design choices that affect correctness, robustness, fairness, and regulatory compliance. We argue that existing benchmarks are insufficient to assess whether MLE agents can be safely applied in such settings. We propose desiderata for a responsibility-centered evaluation framework and conduct an exploratory study on melanoma classification, focusing on fairness across skin tones as a responsibility constraint. When evaluating two recent MLE agents, we find that agent-generated pipelines show high variance and consistently underperform manually designed baselines in both predictive quality and fairness, despite fairness-oriented prompts. These preliminary results suggest that further research is needed towards redesigning MLE agents to allow humans to guide the search process and reliably assess the compliance and quality of the generated ML pipelines.

AI agents can increasingly handle the mechanical tasks of financial analysis: retrieving documents, calculating formulas, updating spreadsheets. The harder, more valuable challenge is reasoning through the open-ended questions that define expert Analyst work. Existing benchmarks do not capture this class of problem, and those that attempt to evaluate open-ended reasoning rely on model-judged outputs that introduce noise and circularity. We present Hedge-Bench 1.0: a benchmark of 102 actual, on-the-job tasks grounded in the explicit reasoning traces of professional hedge fund analysts working with relevant information sources. This approach enables deterministic grading against verified expert steps. Frontier models and agents score below 16% on the benchmark. We publish the dataset and evaluation harness at github.com/Trata-Inc/trata-hedge-bench.