Agentic Benchmarks

LLM agents increasingly work on long-horizon tasks, and the decisions they make along the way, such as which hypothesis to test or which implementation to build on, determine the outcome of the whole run. Making these decisions well is becoming a key capability for both engineering and research agents. We refer to the ability to make good long-horizon decisions as the taste of an agent. While existing benchmarks measure the end-to-end success of agents on long-horizon tasks, none of them measures the taste of an agent. To address this problem, we build Taste-Bench, a benchmark of taste questions constructed automatically from trajectories that agents produced in engineering and research tasks. Each question presents a decision fork, a point in a trajectory where multiple directions are available and one of them leads to a better outcome, and the evaluated model chooses among these directions without seeing what happens after the fork. We mine these forks automatically from parallel attempts at the same task and from detours inside a single trajectory, without needing human annotation. We evaluate frontier models on Taste-Bench and find that the best model answers only 59.7% of the questions correctly. We further find that forks whose deciding evidence appears later in the trajectory are much harder for every model, and that a larger reasoning budget does not improve the accuracy. Finally, we show that taste can be trained. We distill the judgment of a teacher that has seen the outcome into a student model, and the student makes better decisions on unseen tasks and improves end-to-end success on held-out SWE-bench Pro tasks.

Modern video games provide a measurable testbed for AI models, combining abilities of visual understanding, instruction decomposition, goal planning, and precise action control over multiple temporal horizons. Existing datasets and benchmarks, however, either cover a narrow range of games, lack language instructions, or rely on high-variance online rollouts. To address these challenges, we introduce GameHorizon, a unified data and evaluation suite that measures gameplay capabilities at different horizons for diverse model families. GameHorizon Suite consists of three components. First, GameHorizon-Annotator is a scalable and automated annotation pipeline for multi-horizon instructions. Second, utilizing the pipeline, we construct GameHorizon-Data, the first large-scale AAA gameplay dataset with temporally aligned videos, player actions, and multi-horizon instructions. It comprises 5,000 hours of recordings from 21 games, collected by 100 human expert players. Third, we build GameHorizon-Bench with reproducible offline and stepwise online testing. The offline track enables reproducible evaluation using thousands of standardized questions organized into three primary tasks and a series of diagnostic variants, while the online track tests whether offline scores reflect actual gameplay capabilities and localizes failures to specific steps within long-horizon gameplay. Based on our GameHorizon Suite, we evaluate 47 models through more than one million model invocations, revealing a meaningful hierarchy of task difficulty and pronounced differences in model capabilities. Our work can provide a standardized yardstick for evaluating gameplay capabilities across horizons and model families. We will release our dataset, annotator, and benchmark to facilitate future research.

Agents today often take real-world actions that depend on long-term memory and context recall over time. However, most current memory benchmarks are built for a conversational question-answer format, where the question itself signals that some fact must be retrieved, and often which one. Moreover, benchmarks rarely require anything beyond accuracy from submissions, allowing memory systems to make unreasonable cost/time tradeoffs to achieve higher scores. We present DolphinBench, a benchmark that evaluates memory directly through an agent's task completion. DolphinBench includes three knowledge-work personas with roughly 500k tokens of user messages per persona and evaluates agents on tasks that depend on information from that history. We verify all 200 tasks per persona by running an agent with and without the relevant history, requiring success with it and failure without it. Finally, we require all evaluations to report total cost and latency alongside accuracy, which enables us to evaluate agent memory systems holistically. No existing memory benchmark combines all three. The dataset and evaluation code are available at https://dolphinbench.ai.

Evaluation of Computer-Use Agents (CUAs) is often limited to the final deliverables they create (at the end of hundreds of steps) and assessed with functional verifiers, as seen in OSWorld. However, such evaluation of end-state performance lacks transparency into how and why agents fail in various tasks, obfuscating critical insight for subsequent improvement. For instance, agents that err during keyboard inputs would require a different mitigation strategy from those that fail to precisely provide click-based inputs on the graphical UI. We introduce OSWorld-Pro: a set of over 300 tasks containing over 2800 subgoals to enable the procedural evaluation of CUAs grounded in over 67,000 human annotations. We use robust human-aligned LLM-Judges to evaluate the fulfillment of OSWorld-Pro subgoals and thereby reveal the progress that models make throughout a series of sequentially dependent subgoals. Our findings reveal that OSWorld-Pro is challenging even for state-of-the-art LLMs, with top performers like Claude Opus 5 achieving only 75.7% vs. 83.4% on OSWorld. Furthermore, we identify critical process-focused failure modes of various models (e.g. subgoal-irrelevant actions and click-based mistakes) to provide insights to improve performance and efficiency of CUAs.

Effective communication between users and AI agents is essential for human-AI collaboration. The XY problem is a well-known communication pitfall where a person asks about their attempted solution rather than their actual problem. We extend prior sycophancy evaluation to the XY problem in agentic settings, evaluating whether agents can resist plausible but misleading suggestions from users and communicate their reasoning. We introduce XYEval, a meta-evaluation framework that can transform an existing benchmark into an XY problem evaluation. We evaluate five models across six diverse benchmark suites. Agents suffer large XY drops under XY mutation across benchmarks, with relative drops reaching up to 46.7%. With $τ^2$-bench, we further show that agent performance drops more when encountering a pedantic user who requires detailed explanations before approving a better solution. Our findings suggest that current agents lack the ability to effectively reason and communicate when facing misleading suggestions. A simple system instruction baseline that encourages awareness of XY problems only offers partial mitigation. Extensive trace analyses provide behavioral insights into how and why these XY drops occur across execution trajectories. Our results show that mitigating the XY problem remains challenging, requiring agents to both recognize user misdirection and clearly communicate the underlying problem.

Generally, most voice agents are cascaded systems, i.e., an ASR model transcribes the caller's audio, a language model reads the transcript and decides what to say and which backend tools to call, and a TTS model speaks the reply. Nearly all of the decision making happens in the language model, but existing evaluations measure it either too broadly or too narrowly. End-to-end voice benchmarks score the full pipeline, so recognition errors and model errors mix into a single number. LLM benchmarks isolate the model but they do not evaluate what makes real phone calls hard, such as transcription issues, caller's voice being split across messages and the requirement that replies follow the language and script specified. We introduce the Multi-Turn Voice Agent Benchmark (MTVA-Bench), which evaluates the language model on the same conditions it faces inside a cascaded system. The caller is played by an LLM following a set of rubrics and tool calls are answered by a mock backend which responds to the arguments the model actually sent. The benchmark contains 49 agents working across 490 reviewed scenarios and supports 7 languages. Scoring is a combination of deterministic checks on tool calls with two LLM judges, one that scores scenario specific rules and one that grades conversation quality without access to the task. Both judges must cite specific messages from the transcript. Task and conversation scores are weighted equally, since a call can complete its task and still go badly for the caller. In a seven-model study, six of the models select the correct tool within 6.4 points of one another, but their overall scores span 24.4 points. Most of the gap comes from argument values, action ordering, rule compliance, and what the model says around its tool calls.

Agent benchmarks are substantially more costly to evaluate than conventional LLM benchmarks. Benchmark compression is therefore a natural solution, yet existing methods primarily model redundancy in task--model final-score distributions, which is important in agentic evaluation. To address this limitation, we analyze large-scale trajectories and identify six complementary process signals that are systematically associated with final agent performance. To disentangle agent performance redundancy from a complete perspective, we propose DualViewEval, an agent benchmark compression method that jointly exploits outcome and process relations to learn an exact-size miniset and predict the full-benchmark scores. Across five agent benchmarks and five representative baselines, DualViewEval achieves the best results in all datasets. With only 20 tasks, it achieves $24\times$--$40\times$ compression on APEX-Agents and BFCL, reducing mean absolute error (MAE) by $14.5\%$--$28.2\%$ over the strongest competitors while improving Kendall's $τ$ by up to $7.2\%$ relative to EssenceBench on SWE-bench Verified. The selected minisets further reveal capability differences among different agents, providing compact and diagnostic feedback for efficient agentic model development.

Privacy evaluations of tool-using LLM agents often inspect a designated action, final response, or attacker report. These local proxies can miss unauthorized exposure elsewhere in a multi-step session and lack common ground truth across outlets, reports, and tool paths. We introduce privacy exposure displacement, the mismatch between a local evaluation proxy and target-grounded session exposure, and ASLEval, an authorization-aware framework that pre-registers a hidden target set, measures all declared visible exits, and reserves internal traces for diagnosis. Across multiple enterprise-style environments and independently implemented runtimes, we observe three recurring patterns. An expected-outlet-only view misses 46.9% of exposure recovered by the visible-exit union; attacker self-reports combine omissions with high false discovery; and schema-aligned internal evidence usually precedes visible exposure at the request/probe level. Reducing model-visible returns changes this path but can eliminate normal-task success. Independent human review supports the adjudication pipeline while identifying harder console and candidate cases. These findings motivate benchmarks that declare the complete visible boundary, ground claims in pre-specified targets and authorization, and report privacy together with task utility.

Coding agents are typically evaluated with desired behavior specified through issues or instructions. In practical web development, however, agents may need to infer behavior from working software and implement it in an incomplete application. We introduce ProgramDistill, a benchmark evaluating coding agents on features discovered through interaction with fully functional reference applications. We build ProgramDistill by factorizing applications into features of different granularities, each associated with replayable behaviors executable via its gold patch. Our pipeline, mine-craft-patch, discovers 1,975 replay-verified behaviors across 26 applications and constructs 4,063 tasks without human intervention. Across nine frontier coding agents, GPT-6 Astra and Claude Opus 5 achieve 49.2% and 28.8% success on cumulative workflows in full-application reconstruction. In partial-application reconstruction, success falls from 100% to 64.0% and from 96% to 32% as restoration depth increases from 1 to 8. ProgramDistill thus provides a scalable benchmark with controlled difficulty for evaluating and diagnosing coding agents, and a natural basis for future curriculum-based training.

Reliable agent evaluation is complicated by automatic harness optimization, which repeatedly uses a released benchmark $B_{\mathrm{rel}}$ to guide a Proposer that edits prompts, memory, retrieval, tools, and control code around a fixed target agent. Task holdout varies semantic tasks but leaves the benchmark protocol fixed, so a "bad genius" Proposer can produce a cheating harness whose released-benchmark gain depends on a benchmark-wide shortcut. We introduce Counterfactual Harness Search and Evolution (CHASE), which casts harness evolution as constraint generation over validity-preserving benchmark counterfactuals. After each Proposer update, a Challenger searches for an executable protocol transformation with large gain destruction. A validity firewall checks that task semantics are preserved, while a confirmation set determines whether the counterfactual enters a finite archive. We formalize an exact shortcut-neutralized benchmark $B_0$ and establish statistical guarantees linking finite counterfactual archives to $B_0$ and characterizing sequential Challenger search. We evaluate CHASE on a synthetic benchmark and on OfficeQA, where CHASE retains strong released-benchmark gains while substantially reducing gain destruction under valid protocol changes.

Computer-use agents that operate through screenshots and simulated actions are advancing rapidly, yet their evaluation remains anchored to general desktop and web tasks. Enterprise Resource Planning systems run the finance, procurement, inventory, and customer operations of organizations worldwide, and pose distinct challenges for computer-use agents: dense interfaces, coordinated multi-step interactions, and errors that alter persistent business records rather than surfacing on screen. Existing enterprise computer-use benchmarks rely on proprietary platforms or on simulated approximations of such software. We introduce ERPBench, a benchmark that evaluates screenshot-only agents on a live and reproducible system and scores each task against ground-truth values in its database. Beyond the benchmark, we present a production-grade harness that gates agent actions behind human approval for safe deployment. Evaluating six closed and open-source agents, we demonstrate that strong general performance does not transfer to enterprise reliability. Even when an agent reaches the right form and saves it, the stored record is often wrong: some agents save in up to 85% of runs but write the correct value in as few as 3%. We further characterize failure modes specific to enterprise workflows.

We present a framework for evaluating data-science agents on live, continuously updated data using executable ground truth and format-agnostic factoid scoring. Consider this example query: "what were last week's audience sizes"---the reference answer changes as the underlying data changes, so static references become outdated and standard LLM-as-a-judge pipelines cannot verify responses against a fixed ground truth. Our central contribution, ground-truth-as-code, encodes each expected answer as an executable reference function that recomputes the answer directly from live data at evaluation time, ensuring the reference remains consistent with the system it describes. We combine this with a factoid-level, format-agnostic judge that decomposes both the agent's response and the computed ground truth into atomic claims and scores precision, recall, and accuracy over them, irrespective of the response format (prose, list, table, HTML, etc.). The approach is applicable to agents whose expected outputs can be expressed as executable data computations. We validate the framework through a human--LLM agreement study on an internally developed machine learning skill deployed in production, using a synthetic database constructed to reproduce production schemas and entity relationships. Relative to a natural-language ground-truth baseline, our method achieves a 29% improvement in the Matthews Correlation Coefficient (MCC)---a class-balanced measure of agreement between expert annotators and LLM-as-a-judge predictions---and a 16% reduction in token consumption per test case, while a self-directed baseline lacking explicit ground truth is anti-correlated with human judgment. Agents that perform multi-source data integration and computation over non-stationary data are routinely deployed in industry; we propose ground-truth-as-code as a practical methodology for their evaluation.

A shopping conversation has many routes to the same cart, and a task-success rate reduces all of them to one score. We build a deterministic and reproducible e-commerce environment that precommits each trial's customer and trajectory parameters, including the persona, difficulty, target cart, and an item reveal schedule. A simulated consumer attempts to buy a target cart from the environment with assistance from the evaluated model. The environment guides the simulator's actions and records every assistant action alongside the environment state at that point. After the trial, these records allow the evaluator to assess individual parts of the conversation against the retained evidence. For example, the evaluator penalizes a search for failing to surface a target product only when the customer has already mentioned that product. We further use this evidence to apply different penalties to tool calls depending on how the assistant's actions compare with an expected tool-call set. Our environment also interacts with the simulator bidirectionally, reading its output to stop the trial when the simulator determines that the customer has become too frustrated and injecting directives in real time that specify when to explore, defer buying an item, or recall a previous exchange. This interaction creates an open-ended and verifiable simulation. Across eight open-weight agents from 20B to 35B parameters, with 160 trials per agent and 44 metrics, the resulting capability profiles distinguish under-action, over-purchase, unsupported product attributes, and poor search, all of which terminal success obscures.

Frontier models score well on shallow document/chart reading tasks. In a controlled data-room audit, moving evidence into buried conditions reduced accuracy, increased forced declarations, increased tool calls, and increased cost per correct answer. Confidence and benchmark calibration did not fully capture wrong answers; a documented production incident shows fabricated structural claims can be mixed with accurate numeric tables. Agentic evaluations need claim-level receipts (statement-level provenance, not answer-level scores), condition-aware scoring, and human-adversarial verification - an auditing discipline, not a leaderboard. The setting we measure is financial due diligence; the setting we are building toward next is defense staff work, where the same buried-evidence shape appears. In both, the model is not a party to the consequences; the person who signs is. In plain terms: in the documented cases we examine, agents can pair accurate numbers with confident fabricated explanations, and the burden of proof must therefore move from the model to the evidence trail.

Comparing and selecting task-oriented LLM agents increasingly relies on a low-cost offline evaluation gate: persona-driven LLM user-simulators converse with each candidate, an LLM-as-a-judge scores the transcripts, and the higher-scoring agent is promoted. We introduce GAUGE, a reusable offline protocol that measures whether this gate's ranking matches a grounded verifiable reward across 25 agents from six providers on the $\tau^2$-bench and SimulatorArena benchmarks, separating two kinds of evaluation validity that release practices conflate: ranking validity and construct validity. First, a satisfaction-success gap: satisfaction carries essentially no information about task success, as conversations rated satisfied by our blind panel are decorrelated from actual success, with 57.5% of them failing the customer's task, a pattern consistent across five rater populations, both benchmarks, and every subjective dimension we rated. Second, while the gate's ranking is robust across the broad capability span, it loses resolution among the near-equal strong agents: this decision-disagreement rate jumps from $<$1% on wide-reward pairs to 31% on close pairs. The gate is thus human-validated yet mis-anchored. As a remedy, we propose a calibrate-then-trust cadence in which a judge-free completion bit is a zero-cost tripwire for truncation regressions.

Agent evaluations can be numerically correct while measuring a different construct from the one implied by their labels. We present a retrospective measurement audit of selected Praxa AI implementation files, historical evaluation artifacts, and operational records. A 139-case offline routing report contains 112 passes and 27 failures despite zero gating failures, because known gaps are explicitly exempted from the gate. An identifier-free export represents 8,843 tool-attempt rows: 8,395 recorded durations and 448 missing values. Of the durations, 121 equal the signed 32-bit maximum and carry abandoned-client labels; inspected database code clamps elapsed lifecycle age. The pooled recorded 99th percentile is 2,147,483,647 ms, versus 38,118.31 ms among server-observed completed calls. This is a stratum contrast, not a treatment effect. In a documented single-trajectory compaction pilot, the reported follow-up input reduction is 94.39%, but the reduction across the trigger and follow-up calls together is 46.54%. We reproduce the descriptive calculations, verify 91 timing statistics through a separate weighted rational-arithmetic implementation, and execute 13 scoring-function and 12 analysis-verifier tests. Finite-completion bounds show how missing durations limit all-row timing statements without imputing values. The contribution is a source-linked case study and reusable verification package for separating gate policy, lifecycle timing, and request-level accounting from broader agent-performance claims. Historical provider runs and the full current pipeline were not independently reproduced; general capability superiority and population-level statistical significance are not established.

Software development follows an implementation-verification loop in which developers or agents iteratively revise an implementation until an evaluator, such as a test suite, accepts it. The evaluator checks the implementation against a set of requirements under a model of the deployment environment. Yet even a formal proof that the implementation satisfies the requirements under the model cannot guarantee acceptable behavior after deployment. Requirements only approximate stakeholder intent, and the model only approximates the real deployment environment. We call these together - requirement gap and model gap - the two-gap framework, which unifies the main failure modes of agentic software engineer-ing: reward hacking exploits omissions in the requirements or model, while hallucination widens the gaps by fabricating requirements or environment assumptions. Because neither gap can generally be certified closed in an open, changing world, the goal shifts from closing them to continuously narrowing them. We therefore propose an assurance-revision loop that uses deployment evidence to revise the requirements, model, or evaluator when stakeholders reject the resulting behavior. We then cast assured agentic development as a resource-allocation problem over human judgment, agent capability, and compute. The two principal bottlenecks mirror the two gaps: human judgment for the requirement gap and faithful, costly evaluation for the model gap. Reality remains the final verifier: acceptable behavior under actual deployment conditions is the ultimate test, while predeployment evaluations remain proxies for it.

Recently, the rapid development of large language models (LLMs) has reshaped software engineering by enabling autonomous code agents that plan, execute, and utilize external tools iteratively to tackle complex tasks. Beyond achieving functional correctness, these agents must faithfully follow process instructions and constraints throughout the development lifecycle. However, existing benchmarks typically focus on final functional correctness or confine instruction-following evaluation to single-turn, general chat or simple code generation scenarios, leaving instruction-following in multi-turn agentic coding underexplored. To bridge this gap, we propose MTAC-IFBench, a comprehensive benchmark for this critical capability. It features multi-turn progressive software development instructions with diverse constraints spanning 6 primary and 18 secondary categories. With an average of 7.04 turns and 91.33 constraints per instance, it poses a rigorous challenge to current LLMs. To make the evaluation reliable, we construct a checklist for each constraint and functional requirement, and integrate verification scripts and judge agents to verify each checklist item. MTAC-IFBench identifies significant deficiencies in existing code agents in multi-turn instruction-following, with their performance degrading rapidly as the interaction session grows longer.

Large language model agents are increasingly deployed for long-horizon task execution, raising a central granularity question for trajectory evaluation: whole-trajectory verification is too coarse to capture concrete failures and their associated evidence in long trajectories, while atomic-step scoring is too fine-grained, noise-sensitive, and computationally expensive. This granularity gap makes a single-reference trajectory paradigm inadequate for assessing the rich space of valid agent execution paths and delays timely feedback and early stopping in long-horizon tasks. To address these issues, we propose DynSTEER, a dynamic stage-wise framework for agent trajectory evaluation. DynSTEER bridges the granularity gap through stage-wise dynamic evaluation that segments rollouts at key execution nodes and adapts its multi-tier review strategy based on stage-level results; it compiles a path-tolerant milestone graph from public task views to preserve diverse legal paths without reference leakage; and it supports terminating unrecoverable agent executions to curb resource waste. Experiments show that DynSTEER improves evaluation discriminability by 85.2% over native evaluation, separates all model pairs with statistical significance, and saves 45.41% of execution steps on failed rollouts.

Agentic Network Operations (NetOps) are an emerging paradigm promising to enable workload-aware, self-adjustable, and reliable autonomous networks. While agents have proven their value in incident summarization and telemetry signal extraction, their effectiveness as autonomous control-loop engines heavily relies on their long-horizon reliability. One such setting is the datacenter fabric, where an agent must respond to alarms and operator intents while abstaining from high-risk actions that may cause or extend downtime. Abstention, however, presupposes that an action's impact is known pre-execution, which necessitates a per-action ground truth that NetOps agent benchmarks do not provide. We construct such a ground truth for the network repair task of NetArena. A symbolic replay of the emulated network, validated against the environment at every turn, yields the exact value of every action. From the action-level value, we derive two pre-execution targets, namely whether an action reduces the repair distance (progress) and whether it increases it (harm). We show across 10 agent models, that agent verifiers leveraging internal signals predict both harm and progress more reliably than a baseline using observable signals only. Perspectively, we aim to use these signals as safety feedback to an agent harness to abstain from risky actions and protect the target system.

The term agent in artificial intelligence lacks a standard definition, complicating the evaluation, comparison, and reproducibility of AI agent research. We address this ambiguity through a survey organized around five dimensions of agenticness: environmental interaction, learning and adaptation, autonomy, goal-directed behavior, and temporal coherence. For each dimension, we examine how the underlying capability has been conceptualized across prior work and synthesize the metrics, benchmarks, and evaluation frameworks used to assess it. This review provides a structured account of the current landscape of agent evaluation, highlighting both established approaches and areas where evaluation remains limited or inconsistent. We additionally introduce the Agent Compendium, a public-facing digital resource that organizes and extends the evaluation methods identified through this review. Together, the survey and compendium provide a common structure for evaluating and comparing agent capabilities across AI systems, supporting more reproducible research, clearer communication, and more systematic study of artificial agents.

Large language model (LLM) agents can benefit from reusable skills distilled from prior task experience, yet existing skill optimization methods often rely on costly execution-based evaluation and substantial task data. We introduce \textbf{COBRA-Skills}, an efficient framework that formulates skill optimization as budgeted sequential optimization over a dynamically evolving candidate space. COBRA-Skills couples contextual-bandit-guided prioritization with evidence-grounded skill evolution, selectively allocating evaluations to promising or informative candidates while continually refining the skill population from execution feedback. Across six heterogeneous agent benchmarks and three target models, COBRA-Skills consistently achieves the strongest average performance among compared methods, while reducing optimization cost by 55--58% relative to SkillOpt and using only 50 unique optimization examples per benchmark. Further analyses show that COBRA-Skills remains robust to changes in the agent harness and performs effectively when the target model itself is used for skill generation and refinement.

Reproducibility is essential to scientific progress, yet the growing volume and complexity of scientific publications make exhaustive manual verification increasingly impractical. Although recent advances in large language model (LLM) agents enable automated experiment reproduction, existing evaluations largely focus on final repositories and are typically limited to machine learning (ML). We introduce AgentActionBench, a process-oriented benchmark for evaluating agent-based experiment reproduction across ML and AI4Science domains. Our framework uses an MCP-based Action Recorder to capture agents' behaviour throughout the reproduction process and evaluates the resulting traces with paper-specific rubrics. AgentActionBench contains 150 papers, including 120 ML papers and 30 AI4Science papers. A human-annotated subset covering 10% of the benchmark provides validation data, while model-assisted augmentation expands the full benchmark to more than 10,000 rubric items. Experimental results show that current systems remain limited, with execution as the primary bottleneck. Meanwhile, the strong Pearson and Spearman correlations between model-generated and human-annotated rubrics validate the reliability of our scalable rubric-generation approach.

A verifier for robot policies reads a candidate behavior and returns a score for how well it did, used both to evaluate vision-language-action policies and to train them. Verifiers range from success detectors and reward models to runtime monitors, safety filters, and temporal-logic specifications. We survey roughly 150 verifiers and compare them along two properties. Availability is how much a verdict costs, how early in a rollout the verdict arrives, and how often a verdict can be asked for. Availability rises as verdicts get cheaper, earlier, and denser. Credibility is how much a high score tells us about the task. Credibility falls as the judgment becomes gameable and self-serving. We group the verifiers by who supplies the judgment: human verifiers, rule-based and formal verifiers, learned and pretrained verifiers, and model-intrinsic verifiers. Across the four families, we find that credibility falls as availability rises. Regardless of who supplies the judgment, there is no free checker. We then examine what validates a verifier itself, and how much a high score tells us. Three measures appear in the literature: agreement with human labels, the performance of the policy it trains, and behavior under reward hacking. We close with nine metrics that make a verifier claim checkable, and coordinates for the verifiers still to be built.

Agentic systems are rapidly moving to production, where they read untrusted inputs, call tools with real permissions, and act autonomously, expanding the security surface beyond chat-only models. Yet standard evaluations remain single-turn and fail to capture multi-step agent vulnerabilities. We present a systematic black-box framework for risk-aware agent evaluation requiring only basic system descriptions. Our approach introduces: (1) a seven-domain taxonomy mapping observable behaviors to risk categories, (2) fully automated SAGE-RT red teaming producing 120 adversarial scenarios per domain, and (3) human-validated evaluation using LLM judges. Empirical validation across two agent architectures (CrewAI and AutoGen) with four base models reveals alarming patterns: 56.25% average governance risk, 65% privacy risk in multi-agent configurations, and agent behavior vulnerabilities reaching 85%. Our black-box approach effectively identifies critical architectural vulnerabilities without privileged access, providing a scalable path toward safer agent deployments.

Existing evaluation frameworks mostly assess only one part of AI agents, such as task completion (AgentBench) or security robustness (AgentDojo, ASB), rather than the complete pipeline of planning, tool selection, tool execution, memory and reasoning. Failures can occur at any stage, yet existing benchmarks rarely identify their precise source. AgentAudit evaluates the entire execution trace across ten capability, grounding, security and behavioural dimensions, namely instruction integrity, planner, memory, tool selection, tool invocation, tool correctness, alignment, tool faithfulness, security and execution integrity, combined with behavioural classification and failure attribution to pinpoint the exact stage responsible for an observed failure. AgentAudit can evaluate any LLM-based AI agent, since it attaches to the agent instead of replacing it. It reads only the recorded execution trace and does not interfere with how the agent runs, so it places no constraint on the agent's internal implementation. We evaluate five language models (OpenAI GPT-5, Claude Sonnet 5, Sarvam 105B, Llama 3.3 70B and Gemini 2.5 Flash) across nine capability and adversarial tasks. Claude Sonnet 5 and GPT-5 obtain the highest mean Composite Trust Scores (95.1 and 80.6 out of 100, respectively), while Sarvam 105B, Llama 3.3 70B and Gemini 2.5 Flash trail substantially (57.6, 45.7 and 22.6). All traces were scored by a single fixed judge model, which was itself one of the evaluated models, a limitation discussed in Section VII.E. More importantly, models with similar task-completion behaviour can diverge sharply in trustworthiness, as several non-frontier models are repeatedly classified Unsafe_Compliance on adversarial tasks rather than merely failing them, a distinction that pass/fail benchmarks cannot surface.

Procedural memory lets language agents reuse successful routines, but reuse presumes that a stored routine remains applicable. We study what happens when that presumption is deliberately violated. The study combines a retrospective, human-assisted interface-adaptation case from BrowserGym TimeWarp with controlled frozen-memory comparisons on synthetic shopping decisions. During the documented WebShop V1-V6 development path, interface-specific code was adapted while the separately stored high-level procedure was not reported to change; this phase does not constitute an autonomous memory-agent evaluation. In the controlled phase, an early pilot produced one task on which two memory conditions selected a more expensive item while the no-memory condition selected the reference minimum. Follow-up probes did not establish a recurring row-order or identity-binding pattern. We then tested four forms of mismatch: changed quantities, a different evidence representation, a conflict between local and global optimization, and distributed promotion evidence, across 32 formal cells. Each cell used one temperature-0 generation with the same local qwen3:8b configuration and no adaptive retry. Across these pairs, none of the predefined diagnostic interference signatures appeared on the tasks for which they were defined when current-task evidence was explicit and sufficient. The result identifies a tested region of non-interference: a procedural memory can be mismatched without becoming behaviorally disruptive. It does not establish general safety or a mechanism. The remaining question is which additional conditions turn applicability mismatch into observable, memory-caused error.

While autonomous agents have made significant strides in "deep research" by iteratively navigating the open web to synthesize information, real-world problem-solving is rarely confined to a single environment. Complex analytical tasks inherently require agents to weave together evidence from both ambiguous unstructured text (e.g., the open web) and highly precise structured data (e.g., relational databases). However, existing benchmarks evaluate these modalities in isolation, failing to capture the critical "handoff" - the ability to preserve constraints when moving evidence between systems. We introduce HybridDeepResearch, to our knowledge the first deep-research benchmark that requires both web search and SQL to form a complete, verifiable answer. The benchmark contains 380 tool-dependent tasks grounded in LiveSQLBench-Base-Lite databases and public web corpora, validated through automated checks and human review, and covering three reasoning patterns: SQL2S, S2SQL, and Parallel. Evaluations across proprietary and open-weight models under various agentic scaffolds reveal that even state-of-the-art models like GLM-5.2, Claude-Sonnet-4.6 and GPT-5 achieve only about 50-54% Pass@8 on the hard subset. Notably, results show that directional reasoning is substantially more difficult than parallel intersection, highlighting that bridging structured and unstructured information spaces without losing constraints remains a major open challenge for agentic systems. Code and datasets are publicly available at GitHub (https://github.com/Snowflake-AI-Research/HybridDeepResearch) and Hugging Face (https://huggingface.co/datasets/Snowflake/HybridDeepResearch).

Evaluating tool-augmented LLM agents requires diverse, realistic user inputs yet most evaluation frameworks use flat role descriptions ("you are an angry customer") that produce near-identical conversations regardless of the underlying scenario. In this paper, we propose a three-tier persona vector with 23 operationalized dimensions: 6 categorical demographics (jurisdiction, age, channel, device, language proficiency, time availability), 12 continuous behavioral traits (patience, assertiveness, digital literacy, etc.) sampled with Gaussian noise around curated profile base vectors, and 5 continuous emotional states (frustration, anxiety, trust, confidence, stress) that shift in response to scenario context. Orthogonal to the persona, a 4-level query-complexity overlay controls utterance phrasing from direct to deliberately vague. We evaluate the persona model inside a synthetic data generation pipeline across 64,698 multi-turn conversations spanning 8 named profiles and 3 production corpora. Key findings: (i) a 15.8 percentage-point spread in agent goal-achievement across personas confirms trait vectors produce measurably different user behavior; (ii) the same persona behaves differently across scenarios due to scenario-reactive emotional state shifts, validating the scenario-reactive design; (iii) domain-specific projects show persona sensitivity on booking-flow compliance (~15-20 percentage points gap between tier-aware and pressure-test personas), demonstrating the model faithfully reproduces real-world difficulty distributions; (iv) seven rule-described trait correlations produce auditable co-occurrence patterns without requiring learned covariance matrices. The persona model is fully specified for reproduction.

An embodied kitchen assistant must do more than recognize food in isolated frames. It must track ingredient states over time and integrate visual observations with recipe and nutritional knowledge to support constraint-aware decision-making. We formalize this capability as \emph{Embodied Nutrition Management}: perceiving nutrition-relevant events, maintaining a persistent food state, and using it for knowledge-grounded planning. Existing benchmarks evaluate static food understanding or embodied cooking actions, but do not measure whether an agent can continuously update and use nutrition-relevant states in dynamic kitchens. To fill this gap, we introduce \textbf{NutriBench-Kitchen}, a benchmark containing 1,500 manually verified question--answer pairs from 160 cooking videos. It covers five task families: Ingredient Entry, Memory Management, Recipe Query, Long-Term Planning, and Short-Term Planning, spanning food-state construction, maintenance, knowledge retrieval, and decision-making across different planning horizons. Evaluations of proprietary and open-source large vision-language models reveal a substantial gap from human performance, particularly in quantitative ingredient estimation, long-term state tracking, and reasoning under interacting constraints. We further introduce \textbf{Nutri-Vgent}, a diagnostic long-video agent with separate episodic, food-state, and recipe memories. Its consistent improvements demonstrate the value of explicit state representations and structured memory for nutrition management. Together, NutriBench-Kitchen and Nutri-Vgent provide a testbed for studying persistent state tracking and knowledge-grounded reasoning in dynamic kitchens. Code is available at https://github.com/V1ol1n/NutriBench-Kitchen.

LLM agents are rapidly becoming production software, deployed to handle customer service, adjudicate disputes, and operate internal systems. Notably, the work of building them is increasingly handed to coding agents, yet existing benchmarks say little about whether an AI system can deliver one under the conditions of a real client engagement. We introduce $\tau^\tau$-bench (pronounced hyper-tau-bench), a benchmark that makes agent construction the task. A developer agent is given the records a business actually keeps, a client who holds requirements, a production API that operations must run through, a codebase to inherit, and limits on serving cost and models: the same starting point a real engagement provides. From these it must deliver a complete customer-service agent, scored by deploying that agent against held-out simulated users. Across 53 tasks spanning four domains, the strongest configuration, Claude Opus 5 under Claude Code, passes just 23.9% of evaluation simulations. Meanwhile, an expert-authored reference ceiling scores 82.2%. The failures mirror ones human agent developers see: models issue shallow queries in place of deep comprehension of the records, communicate almost nothing to the client, and experiment too little with agent architecture and serving spend, shipping the first design that runs. We aim for $\tau^\tau$-bench to turn the work of cooperative agent building into a measurable target for coding agents.

While autonomous mobile agents hold great potential for assisting older adults with smartphone usage, existing GUI benchmarks mainly rely on explicit, goal-oriented instructions and rarely capture the naturally occurring language patterns of older users, such as indirect speech, referential ambiguity, and under-specified requests. This mismatch between benchmark instructions and real-world elderly interactions may hinder reliable agent deployment. To address this gap, we present ElderBench, the first benchmark for evaluating mobile GUI agents in authentic elderly-oriented scenarios. ElderBench is constructed from 249 naturally elicited smartphone tasks collected from older adults across 20 applications. We first characterize the linguistic divergence between elderly instructions and existing GUI benchmark instructions from syntactic, semantic, and pragmatic perspectives. We then evaluate mainstream GUI agents and Vision-Language Models under both online and offline settings, revealing substantial performance degradation when handling elderly-oriented instructions. Through controlled instruction normalization, failure analysis, and fine-grained linguistic feature analysis, we further identify how elderly-specific language patterns contribute to agent failures. Our findings provide actionable design insights toward more adaptive, interpretable, and age-inclusive GUI agents for older adults.

A merchant's payment processor, ledger, ERP and bank feed are updated by messages that get delayed, duplicated, dropped and reordered, so for minutes at a time the four hold contradictory beliefs about the same order. An agent resolving the exception must decide whether to ship goods, re-submit a capture, refund or wait, knowing some of those cannot be undone. We present FinalityBench, an executable benchmark for that decision. It keeps a hidden canonical event log and derives each system's view from a separately faulted delivery stream, so disagreement follows from specified fault semantics rather than being authored. Grading is on executed monetary effects: an episode is scored by the merchant's terminal economic position, relative to a privileged reference told when the pending capture resolves. The corpus of 321 tasks includes 45 twin pairs (90 tasks): tasks whose four system views are identical at the decision instant, whose authoritative probes both return unknown, and whose eventual correct dispositions differ. That snapshot indistinguishability is checked under every evaluation seed rather than assumed; equivalence over all interaction traces is not claimed. Over 14,445 graded episodes from nine programmatic policies, ranking by single-task accuracy and by paired loss disagree in 7 places: a ship-on-first-sign policy is second-best by accuracy at 65.7% and worst in the suite by paired loss, because it cannot tell the two members apart. A runtime gating irreversible actions on an authoritative finality probe reaches 85.4% and, unlike every polling policy, loses nothing to pass^5; its residual loss is almost entirely one archetype, which prices finality information directly. Language models reach the same exact rate as the hand-written gate on a stratified subset, lose about twice as much money, and discover the finality-gating strategy without being told it.

Long-horizon multimodal agents should remember not only what happened but also who participated. This capability depends on linking recurring faces, voices, names, person-associated objects, events, and social relations to consistent identities over time. Existing long-video and multimodal-agent benchmarks measure broad memory question answering, but they do not isolate the ability to maintain recurring person identities and reason over their cross-time relations. We introduce ICM-Bench (Identity-Centric Memory Benchmark), which, to the best of our knowledge, is the first benchmark specifically designed to evaluate identity-centric reasoning over long video memories in multimodal agents. The benchmark contains 839 synthetic clips spanning 141 minutes and 1,217 open-ended questions about six recurring adults in a one-year life album. A theme-configurable pipeline generates the video collection and associates each question with its target identities and traceable supporting evidence. We compare direct caption-memory baselines, memory-augmented agents, and graph-retrieval systems. Gemini 3.1 Pro achieves the highest overall accuracy of 74.0%, yet its score falls to 60.3% on questions that require long-term identity profiles. The results show that current systems recover many event-level memories but remain less reliable when evidence must be accumulated around a stable person.

Evaluating software engineering agents on realistic benchmarks is costly, since each task may require multi-step code exploration, modification, and test execution. Existing efficient evaluation methods select representative subsets to estimate full-benchmark performance, but are largely result-only: they fit historical pass/fail response matrices or static task semantics, discarding how agents solve problems. We propose PTA-IRT, a Privileged Trajectory-Aware Item Response Theory framework that fuses process and outcome signals. Historical execution trajectories supply process-level evidence beyond pass/fail, such as explored context, attempted edits, and solving paths, which PTA-IRT uses as privileged information for calibration subset selection and ability estimation. Under low calibration budgets, PTA-IRT consistently outperforms prior IRT baselines on score and ranking recovery across four SWE benchmarks. Code and data are publicly available at https://github.com/DeepSoftwareAnalytics/PTA-IRT.

Interactive simulations increasingly evaluate policies in markets populated by language-model agents. Their outputs can look economic---prices, profits, consumer surplus, and welfare---without instantiating the behavior named in the claim. We audit this risk in a multi-turn buyer--seller testbed for configurable hotel transactions. An initial implementation reported welfare gains from two marketplace guardrails of +87.4, +35.0, and +28.8 across a Qwen2.5 1.5B--14B ladder. It also gave guarded and unguarded agents different offer schemas and choice procedures. Holding the schema and buyer chooser fixed changes the paired contrasts to +7.2, -13.9, and +23.8. The four largest 14B single-generation effects averaged +229; after three generations per profile-condition, they averaged +37.6 (95% bootstrap interval [-34.2, 109.3]), while generation residuals account for 49.9% of variation in this post-hoc probe. A seller-incentive check is non-monotone: increasing profit pressure produces less profit than the default seller prompt. Scripted positive controls show why this matters. A profit-maximizing seller already attains first-best welfare, so guardrails mostly redistribute and reduce welfare; they create welfare only when the seller is explicitly programmed to force inefficient bundles. We contribute a construct-validity contract separating incentive validity, protocol isolation, stochastic stability, and welfare accounting, and returning INVALID or INCONCLUSIVE before substantive policy claims. In our case, the original estimate is INVALID under protocol isolation, while the controlled study remains INCONCLUSIVE under incentive validity and stochastic stability. The case does not show that guardrails are ineffective; it shows their apparent value is unidentified until the simulated agents and protocol pass these checks.

Despite the growing use of LLM-powered agents to solve multi-step tasks in complex environments, existing benchmarks rarely test state preservation, performance across languages, and application to realistic, grounded scenarios. To address these concerns, we present WorldBench: a comprehensive, multilingual benchmark of genuine, persona-grounded everyday workflows, where agents can act in a sandbox via structured actions. WorldBench comprises 1,600 tasks across seven languages and eight cultures, filtered and refined through feedback from human annotators with language- and culture-specific expertise. For evaluation, we extend metrics from previous works and introduce Constrained Task Success (CTS), which combines natural language instructions and testbeds to score task completion, minimal modification, and other complementary metrics through deterministic and LLM-as-a-Judge evaluations. Our experiments show that frontier models reach only 49.2% CTS, with all models demonstrating large gaps between correctness and environment preservation. We thereby show that current agents remain brittle in multilingual, agentic scenarios, especially for long-horizon tasks and under state-preservation constraints

Using a large language model to play the user is now standard in scalable evaluation. It has a repeatedly diagnosed failure: the simulated user is excessively cooperative, so a system under test can score by the sheer number of questions it asks rather than by making the user willing to speak. We answer with a disclosure gate conditioning information release on the companion agent's behaviour: its state is a ladder of five ordered gates, merged onto three observable depth layers. We specify, ablate, and audit it, and train a user simulator against that specification. Gating behaviour is learned from the training corpus's synthetic branch, while the real branch supplies how people speak and react; after training, the simulator need not be told at runtime which gate each item sits behind. The gate is a load-bearing component of the environment: on the English corpus of a published companion-agent benchmark (CompanionBench), once training no longer states per example which gate each item sits behind, the largest rank displacement across 12 systems under test exceeds the noise band set by re-running that environment under a new seed, while per-system scores show no detectable change. We state two acceptance criteria: a ranking must be order-preserving, and absolute scores must be scale-stable. Of the candidates we examine, only one passes both -- the simulator we release -- and its leaderboard correlates at 0.993 with the benchmark's original simulator. By contrast, prompting a frontier model as the simulator barely moves the ranking while shifting every score upward -- a shift invisible to anyone checking the ranking alone. The environment we specify is the one that benchmark already used. That publication describes the mechanism in about four hundred words, and we supply what it lacked: specification, ablations, human studies, negative controls, and downstream sensitivity analysis.

Recent advances in LLM agents enable a new paradigm for asset pricing, which we call Agentic Empirical Asset Pricing (AEAP): systems that autonomously conduct the scientific discovery process itself. We define AEAP and identify its core building blocks. Existing evaluation practices backtest only the outputs (factors or trades), not the autonomous discovery system that produced them. We focus on factor discovery, contributing a reference architecture, a rigorous evaluation standard for discovered factors, and a method for out-of-sample backtesting the discovery system. As a concrete instance of that architecture, we evaluate SEADS against five re-implemented baselines on two US equity panels using this standard: no single metric ranks the systems consistently, motivating evaluation on multiple axes at once. A separate rolling re-execution then asks the complementary question of whether the discovery process itself, not one static output, is reliable. We also report negative findings and limitations that surface further evaluation pitfalls for future AEAP systems.

Large language models (LLMs) increasingly interact with external environments and accumulate substantial behavioral experience, yet existing agent benchmarks largely evaluate them as fixed policies. It therefore remains unclear whether an agent can actively test its behavior, judge the resulting experience, and use that experience to improve future decisions. We introduce \textbf{S\textsuperscript{3}Gym}, an interactive benchmark for evaluating LLM self-improvement through three coupled capabilities: \textbf{Self-Testing}, \textbf{Self-Judging}, and \textbf{Self-Improvement}. S$^3$Gym separates permissive exploration from strict held-out evaluation and instantiates this protocol in seven text-based games with executable environment verifiers. We evaluate three pathways for incorporating interaction experience: direct History ICL, score-conditioned Summary Memory, and parameter Training. Our experiments reveal that self-improvement is neither automatic nor uniform. Context-level experience improves performance for several model--game pairs, but the most effective pathway depends strongly on the task structure: summaries are beneficial when experience can be compressed into reusable strategic rules, yet often underperform raw history when success depends on precise, state-contingent information. Parameter training produces substantial gains on some tasks, but also exhibits unstable improvement and severe negative transfer on others. These findings show that recognizing successful actions is insufficient; agents must also transform feedback into executable and transferable policies. S$^3$Gym provides a unified framework for diagnosing this process and identifying the bottlenecks that prevent agents from translating interaction experience into reliable self-improvement.

Autonomous scientific research agents are increasingly applied to end-to-end scientific workflows, including literature review, data analysis, experimentation, and report generation. However, open-ended research tasks often do not clearly specify the analyses, methods, and success criteria required to complete the task. As a result, agents may miss important analyses, use inappropriate methods, or draw conclusions that are insufficiently supported by evidence. To address the problem, we present AutoSciRub, an evaluation-first framework that induces a task-specific executable rubric before research execution, and uses it to guide execution, criterion-level verification as well as iterative revision. AutoSciRub decomposes an underspecified instruction into atomic scientific goals, grounds them in relevant literature and task-visible data, and synthesizes specific, actionable, and verifiable criteria. The resulting rubric makes implicit experimental and evidential requirements explicit, providing guidance for experiments and analyses. During revision, rubric-guided verification identifies unmet criteria and enables targeted refinement of the research report and its supporting artifacts. On ResearchClawBench, AutoSciRub consistently improves all tested configurations, with an average gain of 2.08 points across three backbone LLMs under the fixed Codex harness and 2.95 points across three agent harnesses using a fixed DeepSeek-V4-Flash backbone. On a randomly sampled 20-task subset of AstaBench E2E Discovery, AutoSciRub further achieves an average improvement of 16.8 points across three agent harnesses, while maintaining or increasing the number of successfully completed tasks. These results demonstrate that evaluation-first guidance provides an effective and generalizable control mechanism for autonomous scientific research (Code: https://github.com/zjunlp/AutoSciRub).

Agent evaluations often use one benchmark to choose a workflow and then search for task types where its advantage weakens, so both conclusions are selected from the same data. We introduce Selection-Aware Semantic Stress Testing (\SASST{}), which learns a task reweighting from pre-execution features on discovery tasks and evaluates the same paired comparison on separate confirmation tasks. The protocol checks support and stability, uses joint bounds for all planned claims, and can return no claim. We prove conditional asymptotic validity under stated cluster assumptions. A forty-cluster audit finds Gaussian undercoverage and conservative Bonferroni $t$ bounds. In one 480-episode $τ$-bench study, a $3.75$ point discovery gain vanished on confirmation. A second-model study likewise confirmed neither a workflow benefit nor a stable stress rule.

LLM agents are increasingly used to run autonomous ML experiments, iterating on target metrics with little human oversight. Prior work has documented reward hacking in these environments, bringing into question the validity of produced research and the broader safety case for AI R&D. Existing benchmarks do not measure exploits that live in the data or the modeling task itself. We introduce BAITBENCH, a suite of three synthetic tabular ML tasks that each contain a shortcut that allows agents to inflate the public test score but fail on a hidden test set. Since the shortcut is optional and using it breaks no stated rule, BAITBENCH measures how often models exploit the shortcut to achieve inflated scores. Across seven frontier agents scored by our two-stage judge pipeline, 57.1% of runs exhibit reward hacking, with five of seven above 50%. Agents cheat even under a second condition where they are prompted not to -the mean cheating rate remains above 50%. We release BAITBENCH, along with the judge implementation, and an annotated dataset of transcripts containing reward hacks as a testbed for evaluating reward-hacking mitigations head-to-head.

Large language model (LLM) agents are increasingly deployed in user-facing services that require iterative tool use under dynamic business conditions. Reliable evaluation is essential for sustained improvement: it must reveal capability deficiencies, inform priorities, and assess interventions. Yet industrial agent service unfolds both through the iterative trajectory of a current request and through continued user interaction. Final-outcome assessment can therefore obscure where deficiencies arise and whether later service remains aligned with context from earlier exchanges. We propose ATLAS, a dual-horizon diagnostic evaluation framework for industrial tool-use agents. At the request horizon, trajectory-wise diagnostic signals relate deficiencies to execution locations and capability concerns. At the interaction horizon, user-wise signals assess whether service remains responsive across continued interaction. Together, these views provide structured diagnostic evidence for analyzing execution deficiencies and sustained service behavior. ATLAS instantiates them as executable signals with explicit evidence scopes and decision boundaries. LLM judge interfaces are calibrated against high-confidence references from real business logs; when needed, their decision behavior is distilled into efficient diagnostic models for lower-latency, lower-cost evaluation. The resulting feedback supports policy optimization. We evaluate ATLAS on Meituan Xiaotuan production traffic. Offline experiments assess diagnostic-signal fidelity and replay-based policy improvement, while online A/B experiments show concurrent gains in user engagement, downstream business outcomes, and sampled human-audit quality.

Recent advances in large language models and multimodal models have pushed remote sensing (RS) processing from simple perception models to agentic systems designed to tackle complex, long-horizon RS tasks. However, existing systems often rely on monolithic decision-making frameworks, which fail to accommodate the multi-stage, interdependent nature of RS tasks. This centralized approach leads to challenges such as unstable task execution, incorrect tool usage, and error propagation across stages. To address these issues, we propose HiRS-Agent, a hierarchical multi-agent system for long-horizon RS task solving. HiRS-Agent adopts a two-level collaborative architecture: the Manager Layer handles dynamic routing, step-level verification, replanning, and termination control, while the Specialist Layer organizes domain-specific tools according to the RS workflow and is responsible for subtask reasoning and tool execution. To further enhance the system's capability, we introduce a two-stage supervised tuning strategy and a verification-guided hierarchical reinforcement learning stage to jointly optimize coordination and tool-use policies. Experiments on Earth-Agent Benchmark and ThinkGeo show that HiRS-Agent substantially improves long-horizon tool-use capability and final-task correctness, demonstrating the effectiveness of structured multi-agent collaboration for reliable RS agents. The code is publicly available at https://github.com/IntelliSensing/HiRS-Agent.

Large language model-based multi-agent systems are evolving from fixed interaction topologies toward dynamically orchestrated Agent Swarms. However, existing benchmarks are still largely based on single-agent or general-purpose agent tasks, making it difficult to systematically evaluate key orchestration capabilities. We propose SwarmBench, a benchmark that evaluates model performance from multiple perspectives, including accuracy, efficiency, cost, and process quality. Experimental results show that current models exhibit substantial differences in orchestration capability. These differences are reflected not only in final accuracy, efficiency, and cost, but also in the overall quality of the orchestration process itself. Based on these findings, we further propose SwarmExp, a simple yet effective method based on experience extraction and experience replay, which consistently improves the orchestration performance of large language models.

Multi-Agent Debate (MAD) has been widely adopted to improve LLM-based evaluation by prompting multiple agents to negotiate and reach a consensus. However, for subjective rubric-based scoring, inter-agent agreement does not guarantee alignment with human judgments. In this paper, we compare a single-judge baseline against a consensus-based MAD protocol on subjective evaluation tasks and design three ablations to isolate the impact of role prompting, multi-round interaction, and explicit score sharing. Evaluations across six LLMs show that the single-judge baseline achieves the strongest human alignment on average across six judge models, whereas MAD shows degradation in human alignment on both tasks. Our ablations demonstrate that this performance drop stems primarily from asymmetric role prompting rather than the interaction itself. Specifically, assigning a strict judge role introduces a systematic downward bias that the consensus process fails to correct. The central finding is that this bias reflects strict-stance dominance beyond averaging: the consensus score falls well beyond the arithmetic midpoint of the standalone strict and lenient conditions, rather than averaging them out. Removing role asymmetry (Symmetric MAD) largely recovers baseline performance, while masking peer scores widens inter-agent disagreement on average and worsens average human alignment. These findings demonstrate that multi-agent consensus can enforce artificial agreement at the expense of true human alignment, revealing a structural limitation in consensus-style, role-specialized MAD protocols for subjective scoring.

Large Language Models (LLMs) show great potential as clinical agents, yet existing benchmarks reduce clinical workflows to static predictions or unconstrained Markov Decision Processes (MDPs) with coarse action sets. To address this, we introduce GPAgentBench-2K, the first Constrained MDP (CMDP) LLM-agent benchmark for primary-care clinical decision-making, constructed from expert-validated records of real-world GP encounters. Our environment models a full spectrum of six foundational clinical actions, imposes a topological workflow prior over the action space, and operationalizes safety-informed abstention as a first-class outcome. Evaluating 16 state-of-the-art LLMs reveals a significant performance degradation as the action space scales. Crucially, we uncover a clinical quality-safety gap: even frontier models with the highest diagnosis accuracy violate safety constraints in over half of high-risk cases. Finally, we establish a reference point using Constrained Group Relative Policy Optimization (C-GRPO), and show that while explicitly modeling constraints improves performance over unconstrained RL methods, it remains far from clinically acceptable safety.

As agentic systems evolve into complex multi agent orchestration workflows, there is a growing and critical need for systematic frameworks that measures an agent's behavioral consistency and determinism. In this paper, we introduce a formal evaluation methodology that is grounded in AgentGraph, a planner powered by a domain specific language that represents agent reasoning through a dynamically adjustable directed graph. We leverage this structural formalism and utilize graph traversal algorithms that exhaustively enumerate conversational paths, forming a comprehensive evaluation set that captures the agent's complete behavioral space. We then systematically replay these reproducible trajectories to compare observed outputs and state transitions against the intended DSL specification. To quantify reliability, we define novel metrics that measure response and trajectory determinism, structural adherence and semantic consistency across both exact replays and their linguistic variants. Our system's results demonstrate that agents configured using frameworks like AgentGraph and LangGraph with explicitly structured node transitions show superior determinism over agents that are not configured with controlled transitions.

Retrieval-Augmented Generation (RAG) improves the factuality of large language models (LLMs), yet existing RAG systems often struggle with complex, multi-step reasoning that requires adaptive retrieval and continuous revision of intermediate contexts. Recent reinforcement learning (RL)-based agentic RAG methods partially alleviate this issue, but typically rely on coarse-grained action spaces and trajectory-level rewards, resulting in weak reward assignment and a bias toward short-horizon, stereotyped reasoning template. To address, we propose AgenticRag-R1, a RL framework that deeply integrates reasoning, retrieval, and memory via a memory stack and fine-grained action space, supported by hierarchical action-aware rewards and an information-aware trajectory rejection strategy to enable effective long-horizon learning. Experiments across a diverse set of multi-hop, open-domain, and agentic reasoning benchmarks, spanning multiple backbone model sizes, demonstrate that AgenticRag-R1 consistently outperforms strong baselines. Moreover, AgenticRag-R1 learns more robust, interpretable, and memory-aware reasoning behaviors, highlighting the effect of fine-grained action modeling and information-aware optimization for long-horizon reasoning. Our code is anonymous available at https://github.com/jiangxinke/Harness-RL/tree/AgenticRAG-R1-Whitebox.

Coding agents are now commonly evaluated on the SWE-bench family of benchmarks, whose tasks are built from curated GitHub issues: long, structured, and information-rich. Real user requests, however, are typically far shorter and less structured. To characterize this gap, we define a six-category information taxonomy and four dimensions of linguistic style, and apply them to real user prompts from SWE-chat and problem statements from SWE-bench Verified and Pro. We find that requests carrying only a problem statement, alone or with limited additional context, account for 88% of real prompts but just 7% of benchmark problems. Furthermore, 87% of real prompts are casually written whereas 94% of benchmark problems are formal. Guided by these observations, we introduce RealSWE, 381 multi-variant task families derived from SWE-bench Verified and Pro. Variants within each family share the same underlying task and gold patch while differing only in information composition and linguistic style. Evaluating seven contemporary LLMs with RealSWE, we find that i) realistic inputs reduce resolution rates by 6.4 pp on average and can change model rankings. Controlled analysis further shows that ii) including Desired Behavior and Motivation significantly affects performance, whereas Environment Information and Reproduction Steps merely add tokens without measurable benefit; iii) linguistic style has only small, model-dependent effects. These findings provide actionable guidance for users and agents: explicitly stating the desired behavior and motivation, which most real prompts omit, substantially improves the LLM's software engineering performance.

Compositional reliability bounds for multi-agent systems multiply component reliabilities, a step licensed by a conditional-independence assumption that is routinely stated and rarely tested. We test it. Two instances of one model, in a two-agent handoff, co-fail on 90.0% of the missions on which either fails (log OR 6.66, 95% CI [6.38, 7.00]; phi 0.916), in a preregistered evaluation of 18,000 missions scored by deterministic code with no LLM judge. Substituting a different model reduces the association in six of six contrasts; substituting a different vendor, model already different, does not -- a registered hypothesis reported as a null. The error is signed and runs against the operator: positive dependence inflates joint failure above the independence product, so redundancy is over-credited exactly when components share a model. The assumption-free alternative is often vacuous, and fitting a dependence model is worse: we prove a bootstrap bound on a fitted model's functional loses coverage of the truth as n grows, the identification gap being O(1) while the bootstrap haircut is O(n^{-1/2}). More data makes such a certificate worse, with no visible symptom. We give a finite-sample certificate assuming no dependence structure: a linear program over the joint, over a Bonferroni-Clopper-Pearson box around measured co-execution moments. It is sound, sharp for the information supplied, and monotone in the moment family. Enriching ten moment functionals to fourteen narrows the identified interval by 85.7% and lifts the certified floor from 0.2455 to 0.4116. A companion anytime-valid certificate holds type-I error at 0.0471 under optional stopping. Common dependence statistics are marginal-bounded and can reverse an apparent ordering of conditions when the compared agents fail at different rates. Contracts, scoring code, analysis scripts, and the preregistration are released.

Agents are increasingly considered for automating network operations and maintenance, where engineers must diagnose network faults, optimize configurations to enhance services, and reduce operational costs while acting under strict constraints. However, existing evaluations fail to accurately model real network characteristics or assess agents under partially observable telecom environments with diverse vendors, devices, protocols, and interfaces. In this paper, we introduce CTBench, a public benchmark for assessing whether an agent behaves like a competent telecom troubleshooting engineer. CTBench focuses on root cause analysis and path restoration. Each task is constructed by experts and annotated with rich task metadata, including golden evidence steps. CTBench uses expert-grounded metrics that evaluate both final answers and the diagnostic evidence. Experiments with representative harness-model combinations show that state-of-the-art agents perform very well at identifying endpoints in path-restoration tasks but, more generally, underperform in root cause analysis. In particular, agents struggle with interface state, link-layer, service-management, and other operational faults. Most importantly, even when agents produce plausible or correct final answers, they often fail to provide the evidence-grounded diagnoses required in operational practice. Our results further show that path restoration is generally more resource expensive, yet larger resource usage does not necessarily translate into better diagnosis.

Agentic systems powered by large language models (LLMs) have opened new opportunities for business ideation. Yet existing approaches remain confined to a text-only paradigm, despite the inherently multimodal nature of real-world contexts. We thus introduce MBA-Bench, the first multimodal benchmark for training and evaluating business ideation agents, comprising 30K samples across six domains, each domain characterized by distinct visual cues not fully conveyed by text alone. Concretely, we automatically caption images and employ GPT-4o to generate five reference ideas for each of three business questions through retrieval query generation, market evidence retrieval, and evidence-augmented synthesis. Following prior work, we evaluate agents across six business-oriented criteria using MLLM-as-a-Judge. To consider settings where criteria are hidden or disclosed, we present MBA-b and MBA-k for blind and known, respectively. We train both with two novel reward objectives---creativity and feasibility---while MBA-k further optimizes the six disclosed criteria for eight in total. Both are trained via LoRA-based supervised fine-tuning followed by group relative policy optimization with these setting-specific rewards. For extensive experiments on MBA-Bench, we set up two baselines accommodating either captions only or multimodal inputs, with the latter nearing closed-source performance on several metrics. MBA-b and MBA-k outperform caption baselines by 63.9% and 77.1%, and multimodal baselines by 25.6% and 35.8%, respectively.

AI agents are rapidly improving in cybersecurity capabilities when the source code is available for analysis, yet much of the software most consequential to cybersecurity, including malware, firmware, and proprietary applications, is available only as binaries. Analyzing such software requires reverse engineering(RE): recovering program semantics before the analysis can be meaningfully performed. However, evaluating agentic RE poses a fundamental challenge: benchmark instances must be unseen as source code in the LLMs' training data to prevent models from taking shortcuts by recognizing them rather than really analyzing them, while also matching the scale and anti-analysis protections of real software. Unfortunately, however, existing benchmarks do not jointly satisfy these requirements. To this end, we introduce SRE-Bench, the first realistic, contamination-free RE benchmark. Built entirely from scratch by RE experts with over 5,000 hours, SRE-Bench comprises 19 private, real-world-scale programs averaging 16.9K lines of code. We further developed 44 in-house anti-analysis primitives, yielding 262 binary instances and 1572 deterministically graded tasks. Our evaluation across five frontier LLMs (GPT-5.6-sol,Claude-Opus-5,GPT-5.5,Grok-4.5, and GLM-5.2) shows that RE remains largely unsolved: the strongest model, GPT-5.6-sol, scores 61.4% per instance, and fully solves only 31.5% of the instances. Our analysis further reveals that agents behave differently from human engineers, where agents are relatively insensitive to compiler optimization and static linking. Controlled ablations also confirm that both contamination control and realistic scale are essential. These results indicate that strong source-code security capabilities do not yet transfer to binary analysis, highlighting RE as an important frontier for agentic cybersecurity and SRE-Bench as a rigorous testbed to measure progress.

Mobile agent benchmarks increasingly rely on LLM-based judges to evaluate task completion, yet the reliability of these judges on mobile agent trajectories remains largely unexamined. We introduce MobileJudgeBench, a benchmark for systematically evaluating LLM-as-judge methods on mobile agent trajectories. Our benchmark comprises 931 human-annotated trajectories spanning 6 mobile agent benchmarks, 4 agent models, and 68 apps. We evaluate 6 judge methods (five adapted from SPA-Bench, A3 with two modes, AndroidArena, and AgentRewardBench, plus a simple baseline we design) across multiple LLM backends. Our experiments reveal three key findings. First, a simple baseline judge with sampled screenshots is competitive with, and often exceeds, purpose-built methods, indicating that more elaborate judge pipelines do not consistently improve judge quality; among competitive methods, the LLM backbone is the primary driver. Second, benchmark quality metrics reliably predict real-world judge utility: they correlate with both agent ranking fidelity for evaluation and downstream performance when judges serve as reward signals for on-policy reinforcement learning. Third, failure analysis across two LLM backends uncovers qualitatively opposite failure profiles, one conservative and the other permissive, linked to the backbone's precision-recall characteristics.

Enterprise practitioners read agent leaderboards as if they ranked agent capability. We show, across three open agent-trace benchmarks (TheAgentCompany, $τ^2$-bench, and AppWorld), that the agent main effect accounts for less than 3% of total variance in every dataset and check type, while the agent-by-task interaction accounts for 7-23%. Leaderboards rank specialization, not capability. We arrive at this through a four-facet Generalizability Theory variance decomposition, fit with three estimators (Henderson Method-I, REML via lme4, and a Bayesian binomial GLMM) that agree to three decimal places. Four further findings sharpen what the leaderboard is hiding. First, aggregate reliability collapses on the hardest task quartile: $Eρ^2$ on $τ^2$ action_checks falls from 0.752 to 0.000. Second, training-cell reliability negatively correlates with held-out reliability ($r = -0.90$ on $τ^2$), meaning the designs that look most reliable replicate worst. Third, population-level diagnostics transfer across enterprise benchmarks (capability-gap ratio stable at 0.35-0.40) but per-family agent rankings invert. Fourth, on the MAST failure taxonomy, trace-level mode profiles are idiosyncratic (MAE = 0.261) while cell-level profiles generalise (MAE = 0.056, $r = 0.83$). We package these into Deployment Decision Reliability (DDR), a one-page reporting discipline that turns the variance-component table into five decisions an enterprise buyer can defend. All code, data loaders, and fit artifacts are released under an open-source license.

Real-world data science involves long-horizon workflows that span data wrangling, exploration, modeling, visualization, and validation, and require coordinated use of tools such as notebooks, IDEs, terminals, browsers, and databases within real operating environments. Yet existing benchmarks lack real-computer interaction and do not evaluate whether agents can execute complete end-to-end data-science workflows in realistic computing environments, failing to capture the multi-stage, multi-tool nature of data-science practice. We introduce DSAgentBench, the first benchmark to evaluate whether agents can automate full data-science workflows inside real computer environments. DSAgentBench contains 275 diverse tasks covering the entire data-science life-cycle, reflecting the complexity and tool coordination required in practice. Each task requires grounding decisions in intermediate outputs and coordinated tool use, and includes a deterministic evaluator that verifies analytical correctness, visual outputs, and model performance rather than code-only execution. Our extensive experiments with 15 closed- and open-source models show that even the strongest agent, Claude-4.6-Sonnet, achieves only 56.70% task success, while all open-source agents remain below 1%, frequently failing at tool orchestration, OS grounding, and multi-step reasoning. These results reveal a substantial capability gap between current agentic systems and real data-science workflows, positioning DSAgentBench as a foundation for developing grounded, verifiable, autonomous data-science agents. We release DSAgentBench at https://github.com/vis-nlp/DSAgentBench.

The safety of large language model (LLM) agents depends not only on model weights but also on the agent harness that manages context, memory, tools, permissions, and runtime control. Existing safety mechanisms often treat the harness as a fixed deployment artifact, limiting their ability to evolve with emerging risks. Moreover, coupled functions across harness components obscure safety responsibility attribution, making localized evolution difficult. We propose Safety Harness Evolution (SHE), a framework that learns evolving safe boundaries from rollout trajectories. SHE decomposes the harness into four artifacts with explicit safety responsibilities, including the System Prompt, Rule Bank, Safety Memory, and Tool Policy, defining clear functional boundaries for localized evolution. Based on this decomposition, SHE introduces an attribution-guided evolution loop that converts trajectory failures into structured diagnoses, learns artifact-specific boundary refinements, and selects evolved harnesses through safety-utility validation. Experiments on Agent-SafetyBench demonstrate that SHE effectively enhances safety through harness evolution, achieving a 3.1x ASR reduction compared with static SafeHarness, while also improving benign utility. The evolved harness further generalizes to unseen risks on the held-out AgentHarm benchmark and transfers across agent models without additional evolution.

Recent advances in visual generative models have enabled high-quality image and video generation, but evaluating these models often demands sampling hundreds or thousands of images or videos, which is computationally expensive. Existing evaluation methods also rely on rigid pipelines that overlook specific user needs and provide numerical results without clear explanations. Mimicking how humans quickly form impressions of a model's capabilities from only a few samples, we propose the Evaluation Agent framework, which employs human-like strategies for efficient, dynamic, multi-round evaluations, offering detailed, user-tailored analyses. Given a natural-language evaluation request, the agent decomposes it into sub-aspects, generates targeted prompts, samples images or videos from the evaluated model, invokes suitable evaluation tools, and iteratively updates its plan from the observed evidence, covering both predefined benchmark dimensions and open-ended user concerns. The framework is thus efficient, promptable, explainable, and scalable across models and tools. Experiments show that Evaluation Agent reduces evaluation time to 10% of traditional methods while delivering comparable results. We further introduce Open Evaluation Agent (Open-EA) by constructing EA-CoT-10K, a corpus of history-conditioned step-level instruction-tuning records derived from multi-round evaluation rollouts, and training EA-3B from Qwen2.5-3B-Instruct as a local planning backbone that preserves the structured reasoning, tool invocation, and summary protocol of the API-based agent while reducing dependence on proprietary backbones. Experiments validate the API-based agent on established T2I/T2V benchmarks and open-ended queries, and evaluate Open-EA on four in-domain and three out-of-domain T2V generator families, showing partial cross-family transfer of the learned policy.