LLM Agent Evaluation

When retrieved evidence contradicts an agent's prior beliefs, does it revise its answer, acknowledge uncertainty, or persist with an incorrect conclusion? Existing evaluations of agentic systems focus primarily on task success, offering limited insight into how agents handle such conflicts. We propose to evaluate agents on epistemic humility (EH): the agent's willingness to recognize, act on, and communicate uncertainty during task execution. We operationalize EH through three trajectory-level behavioral dimensions: Identify, Solve, and Escalate (ISE). Through knowledge conflict, situations where the backbone language model's parametric knowledge contradicts the evidence it encounters, or where two contextual sources disagree, we evaluate two conflict settings: (1) controlled conflict and (2) naturally occurring conflict during multi-step agentic execution, each paired with matched no-conflict controls. Evaluating four agents, we find that higher task accuracy does not necessarily correspond to greater epistemic humility: some high-accuracy configurations recognize conflicts during execution but do not communicate unresolved uncertainty in their incorrect final answers. Trajectory-level analysis further reveals that agents frequently detect conflicts in early steps of execution but fail to maintain or resolve them in later steps. Finally, we show that model-level interventions can improve EH, but often at the cost of task accuracy, suggesting that epistemic humility emerges from the interaction among the backbone model, agent harness, and evaluation environment.

As claims about recursive self-improvement (RSI) and artificial general intelligence (AGI) proliferate, we ask a simple question: do frontier AI models have a sense of data, i.e., can they reliably select the right data for training? We introduce DataSense-Bench to study this capability through the fundamental problem of data selection and performance forecasting in machine learning. We ask AI agents to select and rank candidate training subsets that can be used to fine-tune a small LLM model. Agents are allowed to inspect the data, write and execute analysis code, and run model forward passes, but can not train the model or access the actual evaluation tasks. We then fine-tune the base model on each selected subset and evaluate its post-training performance under a standardized protocol. We instantiate the benchmark in terminal problem solving and tool use, selecting trajectories from OpenThoughts-Agent and EnvScaler and evaluating on TBLite and BFCL, respectively. We then evaluate the agents along two complementary dimensions: the post-training performance of the top-ranked subset, reflecting the ability to identify high-value training data, and ranking accuracy, reflecting the ability to predict the relative performance of the selected subsets. In our experiments, selection gains over random selection are limited; agents do not reliably rank their selected groups, and ranking ability does not hold consistently across tasks: Astra identifies the best group in all three tool-use runs but in only one of three terminal runs. Analysis of execution traces on both tasks shows that agents often use similar data signals while interpreting their training value differently.

Black-box optimization (BBO) arises in many scientific and engineering problems where objective evaluations are expensive and limited. Recent large language model (LLM) agents offer a new way to approach BBO by combining task semantics, computation, optimization tools, and feedback-driven decision making, showing great potential due to the integration with mathematically rigorous tools. However, existing agentic BBO studies use different task domains and system configurations, making their results difficult to compare and the effects of individual design choices hard to isolate. We therefore introduce AgenticBBO-Bench, a cross-domain benchmark for agentic BBO spanning synthetic functions, hyperparameter optimization, database tuning, chip design, and molecular design under a unified finite-budget evaluation protocol. In our experiments, agentic BBO achieves higher family-averaged scores than direct LLM-based methods in all five domains and outperforms the best numerical optimizers in four. We further study three factors shaping agent performance: optimization tools, task information and prior knowledge, and the role of the LLM during search. Our results show that additional numerical tools do not consistently improve performance, task semantics are broadly useful while more specific priors are less reliable, and numerical optimizers can effectively absorb gains from search trajectories established by the agent. Finally, we introduce a five-task frontier challenge within AgenticBBO-Bench and evaluate seven LLMs under the Codex agent harness, where GPT-6 Astra and DeepSeek-V4.1-Flash lie on the Pareto frontier of performance and cost among the evaluated models. Our code is available at https://github.com/lamda-bbo/agentic-bbo.

Guard models are increasingly used to safeguard LLM-based agents, primarily by identifying actions that agents are forbidden to perform. However, identifying forbidden actions alone is insufficient to ensure agent safety. In this paper, we argue that agent safety also depends on identifying required yet unperformed safety-critical actions, which we call obligations. Our preliminary study on a popular benchmark for evaluating safety shows that 56.92% of GLM-5.3 trajectories contain unfulfilled obligations, compared with only 30.00% containing forbidden actions. This finding reveals unfulfilled obligations as a major and previously overlooked source of safety risk. However, to our knowledge, no existing benchmark evaluates whether guard models can identify these obligations. To close this gap, we introduce ObligationBench, the first benchmark for evaluating the capability of obligation identification, comprising 240 expert-validated trajectories covering issue resolution, feature development, and terminal operations. Our evaluation of 14 representative models reveals substantial limitations: the highest recall and exact-match rate are only 48.97% and 10.00%, respectively. To address these limitations, we develop ObligationGuard using 40,000 synthetic training examples. ObligationGuard achieves 57.52% recall and an exact-match rate of 21.67%, surpassing all evaluated models on both metrics. We call on the community to incorporate obligation identification into the design and evaluation of future guard models to improve agent safety.

Large language model (LLM) agents are beginning to operate industrial energy equipment, and what they get right depends on what they are told about the plant. Established building ontologies name many kinds of points across many sites, whereas an industrial equipment system needs few entities with much knowledge about each. This study proposes the ontology tower, a narrow-and-deep ontology of a single equipment system whose knowledge deepens in two ways: through quantities derived from the measured points by physical relations, and through lessons from the operating journal incorporated as knowledge nodes. On a real low-humidity air-handling test plant operated daily through a programmable logic controller, agents received a text projected from its tower in a preregistered evaluation of nine tasks replayed from the plant's records, using four open-weight models from 9 to about 750 billion parameters. This knowledge raised the rate at which the agents avoided the most plausible misjudgment of each task by about 20 percentage points, and the overall task score of the 9-billion-parameter model as much as that of the largest. Operating lessons were used when incorporated into the tower or placed in the prompt as records, but seldom when left in the journal behind a search tool. In live runs through an invariant safety layer, the agents brought the controlled variable into its target band in 12 of 14 runs. An ontology narrow in entities but deep in what is known about them can thus supply the knowledge that an agent for an industrial equipment system needs.

Verifier scores now serve as both benchmark metrics and training rewards for large language model (LLM) agents, and a change in score is routinely read as a change in capability. It may instead reflect a change in the evaluation. We introduce TRACE, a protocol that turns a score change from a verdict into a testable diagnosis: it applies a targeted change to one part of an evaluation, compares paired runs, checks whether the agent's behavior changed, and rescores unchanged trajectories to test whether the scoring rule is responsible. In a controlled suite of 25 synthetic tasks, renaming tools lowers a scripted agent's score by 0.250 even though it performs exactly the same operations; restoring the original names at scoring time closes the entire gap, while the same mutation exposes a genuine behavioral failure in a second agent. On public $τ^2$-bench tasks with four LLM agents, an initial 30-task study finds mixed reward changes whose one clear effect does not replicate. In a larger follow-up on 88 new tasks with repeated runs per condition, renaming tools or reformatting tool outputs leaves reward unchanged to within $\pm$0.10 for seven of eight agent-change pairs, whereas tool names that deliberately mislead lower every agent's reward by 0.20-0.44, showing that the setup can detect real effects. Identical reruns flip 15-36% of task outcomes, so single-run comparisons cannot separate presentation effects from run-to-run variation. Two frontier LLM judges give consistent verdicts when a fixed trajectory is presented differently, yet disagree with each other on 57% of the same records, largely because one grades procedure rather than outcome. TRACE thus separates what a score change says about the agent from what it says about the measurement.

Automated interviewers and conversational agents are increasingly used in research, recruitment, customer service, and education. However, many existing systems rely on fixed question sequences and provide limited context-based personalization without considering participants' knowledge, which can lead to repetitive or irrelevant follow-up questions. Therefore, there is a need for an adaptive interviewing system that can adjust question depth while maintaining conversational continuity and semantic progression. To address this, an Evidence-Traceable Dynamic Interviewer Architecture is presented using a locally hosted Large Language Model (LLM), with the interview continuously adapted throughout the entire conversation based on the participant's responses and evolving context. The interviewer profiles participants' expertise in real time to generate knowledge-appropriate questions, well-articulated responses, and smooth transition messages that support conversational continuity. A five-module prompt-driven architecture and persistent interview-state record support these functions. The interviewer was evaluated with 246 participants. Expertise Profiling module (M3) showed 78.9% exact agreement with independently reported participant expertise, with a weighted Cohen's K of 0.80. Generate Iterative Questions module (M4) showed a strong expertise-complexity association (p=.79, p<.001), and participants reported high relevance (mean 4.41), engagement (mean 4.32), and satisfaction (mean 4.38), providing evidence that the architecture's adaptive components operated consistently with their intended functions while participants reported a positive interview experience.

Context: Large language model (LLM) agents are increasingly used as software and data-engineering assistants, yet evidence about locally deployable open-weight agents remains limited. Existing evaluations often emphasize textual responses or isolated code generation rather than the validity of complete engineering artifacts. Objectives: We evaluate whether local LLM agents can produce correct and reproducible data-engineering artifacts, quantify the effect of a closed-loop workspace condition, and examine trade-offs in model scale, architecture, quantization, runtime, tool use, and failure. Methods: We introduce a benchmark of fifteen mobility-workflow tasks covering data discovery, connectors, transport-feed processing, semantic enrichment, feature engineering, validation, visualization, and reporting. Deterministic checkers assess generated scripts, tables, structured files, figures, and reports. Ten local configurations are evaluated in one-shot and closed-loop conditions, with five repetitions per model, mode, and task, yielding 1,500 scored attempts on a consumer-grade GPU. Results: Among models larger than two billion parameters, the workspace condition increases pass rates by 26.7-52.0 percentage points over one-shot generation. The strongest configuration reaches 85.3% artifact-level success, and a quantized 9-billion-parameter model reaches 69.3% with an approximately 6.5 GB memory footprint. Gains are largest when intermediate artifacts expose errors the agent can inspect and repair. Conclusion: Local open-weight agents can support a meaningful subset of software-intensive data-engineering work, but reliability depends on model capability, task verifiability, and deterministic validation. The benchmark provides a reproducible method for evaluating complete agent configurations before adoption in engineering workflows.

Large language models (LLMs) are increasingly deployed as coding agents that edit files, run builds and tests, inspect execution results, and repair software iteratively. Embedded firmware is a demanding target because correctness depends on closed-loop behavior under sensing, timing, and safety constraints, not only on static source quality. Yet embedded-agent evaluation remains limited and often emphasizes one-shot synthesis or offline correctness. We present a benchmark for closed-loop evaluation of embedded coding agents. Each task provides a plain-text engineering description, constrained workspace, and visible build-and-runtime surface. The agent must translate requirements into implementation and self-verification steps, then iterate until the required device behavior is achieved. The suite contains five embedded-control tasks and four feedback scenarios: one-shot generation, realistic self-verification, CI-style red/green feedback, and oracle-style detailed feedback. The implementation targets simulated ESP32 firmware for reproducibility. We evaluate seven GPT-family and Qwen-family configurations across five tasks and four scenarios, with three repetitions per condition for 420 runs. gpt-5.4 has the highest pass rate among evaluated configurations but does not saturate the benchmark; qwen3.5-27B is the strongest observed local model; and smaller local models degrade sharply in pass rate and search efficiency. These results suggest that capable local embedded coding agents are emerging.

Tool-using agents are usually scored on whether they finish a task while the tools work. Deployments are less forgiving: services time out, endpoints disappear, parameter names change, and results come back well formed but wrong. Prior work has shown that language models over-trust tool outputs that fail silently; we ask how that over-trust plays out across the stages of failure handling in multi-turn agents. Wrapping the executable environments of an established function-calling benchmark in a fault-injection layer, we inject one of four typed faults at a controlled point in the trajectory and record whether the agent notices, changes plan, recovers the task, or repeats itself. Six models from three families, half of them reasoning variants, ran 1,920 trials over 24 multi-step tasks. Agents treat a failure as a problem in 91.3% of trials when the tool returns an explicit error, but in 58.8% of trials when it returns a plausible wrong value, against a 26.8% rate of reporting problems when nothing was wrong. Reasoning models are not better placed: paired against instruct siblings, they notice less (-9.3 points, p < .001) and change plan more (+10.4 points, p < .001), and recovery is unchanged (p = .512). Because agents are stochastic, two fault-free runs of the same task end in the same state only 63.3% of the time; against that baseline, only a missing tool clearly lowers recovery (39.9%), while timeouts, schema drift, and corruption stay within run-to-run variation. After a fault, agents return to the same tool three or more times in a row in up to 22.2% of trials, though strictly identical repeats are rare. A prompt line asking the agent to check each result did not move detection. Agents respond to the error channel rather than to the content of what a tool returns, so failures that stay inside the expected format pass through.

Evaluating agents by outcomes alone can obscure the capabilities that produce them. This problem is especially pronounced in evolving environments, where outcomes reflect a closed-loop interaction between agent behavior and changing external conditions. We introduce LiveMACEBench, a process-aware benchmark that uses live financial markets as a naturally evolving testbed for persistent LLM agents. Five frontier LLMs operate along continuous trajectories under matched Tool Use, Persistent Memory, Rule Following, and Multi-Agent Collaboration configurations. We evaluate them through both realized outcomes and mechanism-specific diagnostics derived from complete decision traces. Across 30 days of live evaluation, we find a pronounced outcome-capability gap: realized returns often diverge from capability-specific measurements, and similar outcomes can arise from markedly different patterns of mechanism use. Trace-level diagnostics further expose distinct bottlenecks across capabilities, demonstrating that mechanism access, effective mechanism use, and downstream performance are not interchangeable measures of agent capability. LiveMACEBench makes this distinction measurable, turning live markets from a performance leaderboard into a diagnostic environment for agent capability

The evaluation of coding agents generally strives to be as realistic as possible. In our study, we collect 4,782 agent sessions of real software engineers in JetBrains IDEs, which we call Production Sessions. Since our subject is interactive agents, we study the sessions with at least three user messages (33% of the sample). These long sessions differ from issue-derived benchmark tasks in two ways: (i) user requests span a far wider mix of task types - questions about the project's code, planning, review, refactoring, execution - and (ii) users switch between types throughout a session. Long-session samples from three public interaction corpora exhibit markedly different Task Flows (the distributions of session lengths, task types, and type-to-type transitions), so no single interaction distribution is universally realistic: benchmarks should name a target use case and calibrate to measurements from it. We present SWE-TaskFlow, an approach for transforming any issue-derived benchmark: it preserves the verified tasks and tests while steering the interaction toward a target Task Flow through prompt splitting and verifiable repository QA, with a TaskFlow Alignment Score (TFAS) for selecting among generated trajectories. In a pilot on 700 SWE-Bench Pro tasks, solving the task sequentially in several steps approximately doubles agent cost without a stable change in resolve rate: the interaction protocol itself is an important dimension of evaluation.

Forensic reconstruction of LLM-agent actions requires not only recovering the correct value, but establishing which preserved record supports that finding. Tool logs, generated explanations, and local citation identifiers capture different parts of this evidence, yet a citation identifier does not establish a source unless its binding to a record is preserved. We audit this distinction using 64 mechanically checkable cases from saved AgentDojo Banking executions. Two LLM readers reconstruct source relationships under controlled variations in visible evidence and identifier-to-record bindings. We separately evaluate complete-record agreement, evidence-grounded findings, justified abstention, and unsupported assertions. With original identifiers and no binding table, Sonnet recovered every literal source location but made unsupported citation-source assertions in 26 of 28 cases requiring the missing relation; 22 nevertheless matched the complete reference. Adding explicit bindings improved grounded reconstruction for both readers, whereas identifier renaming alone provided no consistent remedy. A deterministic same-packet comparator correctly resolved the bounded task or abstained throughout. These results show that factual agreement alone is insufficient for evaluating forensic reconstruction of agent logs and motivate preserving explicit record bindings to distinguish supported findings from correct guesses.

Taxonomies are the symbolic representations through which AI systems organize evidence, aggregate patterns, and answer questions over large document collections. Over customer feedback, the category tree decides how every record is counted and routed, which problems get seen, and which team owns them. Agentic harnesses now make it easy to generate a plausible-looking hierarchy, and such trees are checked today with generic, individually scoped checks: each name fits its description, sits under the right parent, and stays distinct from its siblings. We ask a more operational question: when is a generated hierarchy actually useful as a production taxonomy? We build six taxonomies over two proprietary feedback corpora (1,940 and 5,000 records): for each corpus, a production reference and two repeated runs of the same harness under identical inputs. All six pass every generic naming and structure check, and a deeper product-coverage check even prefers the generated trees. Yet in every generated tree at least 97.7% of leaf names merely restate an ancestor's name (13.9% and 2.9% in the references), and in one, three of every four records fall under multiple top-level categories. We introduce two families of whole-tree metrics: structural discriminators test whether a tree's shape was learned from the data or imposed by its generator; team partitionability tests whether branches split feedback into groups teams can own. Trees the generic checks rate as equally correct differ by 27 percentage points in cross-branch leakage, and only one of the two beats a random split. Surface plausibility is an insufficient measure of taxonomy quality: evaluation must measure the whole tree as well as each node.

Organizations increasingly use frontier language models to analyze customer feedback, but answer quality also depends on how that feedback is organized and made available. We define a \emph{customer context graph} as a unified model of customer and business context. Typed relationships connect customer objects (feedback, conversations, users, and accounts), operational objects (tickets, support agents, opportunities, and competitors), and analytical or action objects (taxonomy concepts, evidence, insights, work items, and outcomes). This lets an agent investigate not only what customers say, but why, who is affected, what action followed, who owns it, and whether it was resolved. For this experiment, the graph is populated from public Cursor feedback; the same architecture can support any type of feedback source. We compare Agentic RAG, a Deep Research Agent, and a Customer Context Graph-backed Agent on the same 9,432 public Cursor feedback records using 30 realistic product, incident, comparison, and metadata questions. Without exhaustive ground truth, we jointly score responses on answer quality (coverage and organization), analytical depth (specificity and decomposition), and evidence quality (citation support and traceability), using a comparative rubric calibrated on 28 of the 30 questions. We sample cited records against their claims and weight the three dimensions equally. Under this aligned rubric, the Customer Context Graph-backed Agent scores 0.961 overall, versus 0.710 for the Deep Research Agent and 0.651 for Agentic RAG, and leads the Deep Research Agent on 27 of 30 paired questions (sign-test p < 10^(-5); strictly best on 26 of 30). Its largest advantage is analytical depth (0.967 versus 0.642), reflecting more specific, hierarchically developed findings with quantified themes and traceable evidence...

Urban diagnosis integrates heterogeneous observations to identify urban problems, localize affected areas, and investigate contributing factors, informing evidence-based urban planning and management. However, its reliance on labor-intensive, case-specific expert workflows limits scalability and reuse, motivating the exploration of agent-based execution. To evaluate this capability, we introduce DUDA-Bench, a hierarchical and interactive benchmark that formalizes data-driven urban diagnosis as a multi-stage agent workflow. It comprises 86 atomic and 22 workflow tasks spanning four analytical stages, grounded in multimodal data from 12 cities covering five urban problem types. Evaluations of seven backbone models and five agent systems reveal a substantial gap between isolated analytical competence and end-to-end diagnosis, with system benefits varying across backbones. Trajectory analysis shows that unresolved evidence gaps propagate across stages, while successful recovery involves revising assumptions and actions using feedback. These findings highlight limitations in coordinating analytical capabilities across stages, particularly adaptive planning, evidence integration, and verification. More broadly, DUDA-Bench provides a framework for translating expert analytical workflows into hierarchical agent tasks and process-aware evaluation, supporting systematic assessment of end-to-end analytical capabilities.

Before tool-using LLM agents are deployed in environmental and geospatial workflows, teams need evidence that an agent reliably selects the right operations against real APIs. We introduce GeoNatureAgent (GNA), a framework for pre-production evaluation of tool-using agents: a fixed sixteen-tool geospatial interface published as a Model Context Protocol (MCP) server, so the agent under test is the only variable, scored against an identical tool layer, task suite, and deterministic scorer. Its flagship instance is a 103-task benchmark (a 93-task main suite across 18 categories plus a ten-task comparison expansion) evaluated against an open, self-hostable geospatial API serving three environmental indicators across Spain and Portugal. We evaluate nine LLMs under three temperature-1.0 seeds, reporting capability and per-case cost as orthogonal axes. (1) Claude Sonnet 4 achieves the highest capability (61.7% +/- 0.7% on all 103 tasks; 60.8% on the main suite), followed closely by DeepSeek V3.2 (57.9%), while no other model exceeds 53%; (2) the cost-accuracy Pareto frontier is mostly open-weight, with DeepSeek V3.2 offering 93% of Claude's capability at 11.3x lower list-price cost; (3) under strict all-checks scoring the best model sits 24-36 points below the 85-97% reported on general-purpose GIS benchmarks, whereas per-check partial credit for the top four models (86-90%) is comparable, so much of that gap reflects scoring strictness rather than task difficulty alone. The MCP server, evaluation harness, benchmark, and API are publicly available; swapping the tool executors and task suite instantiates an equivalent benchmark for any geospatial domain.

An agentic system issues several structurally different kinds of LLM calls. It routes intent, classifies actions, grounds language in a device registry, plans multi-agent pipelines and writes the Python code those pipelines run. The difficulty of these call sites varies by an order of magnitude, yet in practice a single model, chosen for the hardest site, serves all of them. In this work, we evaluate 9 models from 0.8B to a frontier hosted model across the five call sites of a deployed open-source home-automation framework (Wactorz), using its unmodified production prompts and two real Home Assistant installations (280 cases, 2520 scored calls). We find that capability is not ordered the same way at every site, and that larger models are not uniformly better: one 4B model is worse than its 2B sibling at grounded actuation. Paired testing shows the best local model to be statistically indistinguishable from both hosted models at four of five sites. Only code generation separates them, against a small hosted model (p = 0.039) as well as a frontier one (p = 0.002). Aggregate accuracy also hides a safety failure specific to actuation, where small models resolve the accuracy/refusal trade-off in degenerate ways: one model (Gemma4 E2B) actuates on 87.2% of requests for devices the site does not own, while another refuses every request it receives. Routing each site to its best local model reaches 91.8% against 95.4% at no per-call cost. In a live deployment judged by a user, hosting only the two generative sites matches hosting everything (39/43 against 39/43) for 28% of the spend, and the actuation gap the benchmark predicted appears as exactly one case in twenty-six. Benchmark, harness and all records are released at https://github.com/waldiez/slm-callsite-eval.

Frontier AI evaluations increasingly use open-ended, agentic, long-horizon tasks whose transcripts can span hundreds of pages of outputs and actions from complex multi-agent networks. The observability envelop-the range of what evaluators can reliably infer about an agent's behaviours-is therefore narrowing. Language model assistants can help classify and interpret agent behaviour but also afford human evaluators significant analytical degrees of freedom, threatening the reproducibility and auditability of language-model-based transcript analysis. Transect is an open source package built on Inspect Scout to help evaluators understand how a long agent run unfolded, identify behaviour worth investigating, and check interpretations against the transcript. Users specify task context and behavioural vocabulary in a reusable evaluation-family configuration, with judge models and analysis settings supplied separately. Transect's navigable reports align recorded events, token use, sub-agent activity, and model-generated behavioural labels on a common turn-based timeline. Reviewers can quickly grasp a run's narrative, trace any label or event to its source turns, and export the underlying data tables for cross-run analysis. We demonstrate the workflow on an AI R&D evaluation that generated almost 13 million tokens, dividing the agents' work into behavioural phases aligned with research-skill classifications, sub-agent delegations and interactions, and token use. The combined view shows a focus on operational work and manuscript production, with little evidence of a sustained hypothesis generation stage-arguably a necessary component for high-quality scientific outputs. Transect's flexible, customisable transcript-analysis pipeline will enable evaluators to keep pace with longer, more complex, more frequent AI evaluations while supporting scientific rigour, transparency, and reproducibility.

Learning from experience is essential for LLM agents to adapt to unfamiliar and dynmaic environments. Evaluating this ability is therefore important for understanding how effectively agents acquire and use new knowledge. Existing benchmarks have sought to evaluate this ability, but they primarily evaluate tasks whose rules are provided in the instructions or already familiar to pretrained models, making it difficult to distinguish learning from interactions from reasoning with existing knowledge. To address this, we introduce Learn2Play Bench, a benchmark of newly designed text-based games, whose rules are novel or counterintuitive, requiring agents to acquire knowledge through interaction rather than rely solely on pretrained knowledge. These games provide reproducible feedback and automatic scoring, enabling controlled evaluation of learning across repeated attempts. We also vary game instances to test whether agents can apply what they have learned to new situations. Therefore, we evaluate how backbone models, self-evolving methods, and agent harnesses affect agents' learning ability, revealing three findings: (1) Experience retention: Retaining complete records of actions and feedback can support more effective learning than summarizing these experiences into rules or strategies. (2) Human agent gap: Top-performing human players achieve higher peak scores than the evaluated agents. Human explore more varied strategies, and repeat actions less. (3) Harness matters: With the backbone fixed, changing the harness can improve performance while reducing estimated inference cost. Together, these findings provide insights into how LLM agents learn from experience and suggest directions for future work to improve their learning ability. Project website: https://liushiliushi.github.io/learn2play-bench-website/

As large language model (LLM) agents become increasingly autonomous, they may pursue task performance through deception, raising concerns about their reliable deployment. Existing evaluations show that LLM agents can deceive, but often examine isolated scenarios or narrowly defined conditions, limiting systematic understanding of when deception becomes more likely. To address this gap, we introduce DecepEval, a benchmark comprising 1,532 instances across 3 task families and 28 professional scenarios. Drawing on classical fraud theories, we propose the LLM Deception Diamond framework, which characterizes four external conditions that may induce deception: pressure, incentive, opportunity, and conflict. DecepEval pairs neutral and induced versions of each instance to measure condition-dependent changes in deception rates, while explicit task facts and observable agent behavior help distinguish deception from capability-related errors. Evaluations of nine frontier LLMs show that inducements increase deception across models and task families, even among models with low baseline deception rates. DecepEval makes these vulnerabilities measurable, providing a shared benchmark for progress toward trustworthy artificial intelligence.

When using an LLM agent in a consequential domain, making an informed decision about whether to trust its output or intervene requires calibrated confidence in the agent's success. Confidence estimation for agents is difficult because evidence about success is distributed across heterogeneous, interdependent steps of an agent's trajectory. Practical agentic deployments introduce further challenges: frontier LLMs often provide limited access to internal signals, agent roll-outs are costly, and training data may be unavailable or quickly become outdated. To address these challenges, we introduce Confidence Reasoning Graphs (CRGs), an inference-time framework that estimates the probability an agent accomplished its task from a single trajectory, without privileged model access or training data. Rather than compressing an execution into a single holistic judgment, a CRG begins with the claim that the agent accomplished its task, decomposes it into contextualized sub-claims grounded in trajectory evidence, estimates confidence for each terminal claim, and finally aggregates these into an overall confidence estimate. Across three agentic benchmarks, three backbone models, and three agent frameworks, CRGs yield better-calibrated confidence and stronger risk-aware decision making than verbalized, sampling-based, and white-box surrogate baselines. We further find that calibration error alone can be misleading: a white-box surrogate baseline appears well calibrated while providing near-chance discrimination. Ablations attribute CRG's improvements to claim-level confidence estimation and aggregation rather than graph construction alone. Finally, a CRG exposes the claims and trajectory evidence underlying each confidence estimate, enabling it to be audited at decision time.

Agent Skills package procedural guidance and resources for reuse, but a relevant Skill does not necessarily improve task performance. Existing studies characterize Skill content and evaluate downstream performance, yet provide limited explanations of how utility depends on content, execution configuration, and multi-Skill organization. We conduct an empirical study on 87 SkillsBench tasks, defining downstream utility as the pass-rate difference from No-Skill on the same tasks under the same model--harness configuration. We compare the same Skills across nine configurations, then examine alternative published Skills and organizations of fixed Skill sets under three selected configurations. We retrieve marketplace candidates from a curated corpus of 37,596 Skills. LLM-assisted analysis of content, execution traces, and final artifacts, followed by author review, relates provided support to actual use and task outcomes. The same Skills help some configurations and hurt others on 36.78% of tasks, with trajectories showing that recommended procedures can become an execution burden. Relevance rankings overlook more useful candidates. Within the evaluated candidate sets, reranking by support for required operations raises first-choice pass rates by 4.35--5.80 percentage points across the three configurations. We derive 17 authoring practices linking executable procedures to recovery, preservation of task requirements, and checks on final artifacts. Stage Plan and Dependency DAG outperform use order alone, with DAG's additional benefits concentrated in tasks supplied with five or six Skills. These findings guide developers to assess usable operation support, allow procedure adaptation while preserving task requirements, and make artifact dependencies explicit when organizing Skills.

Large language model agents are increasingly deployed to perform complex tasks in real-world environments. However, the knowledge required for correct behavior in these environments is often implicit, undisclosed, and subject to change over time. Recent continual-learning harnesses seek to address this challenge by enabling agents to improve from serving experience. Yet the effectiveness and limitations of these methods are not yet well characterized. Existing benchmarks provide only partial coverage: some explicitly provide the target knowledge, others assume a static environment, and those that support continual adaptation remain limited in scale and knowledge diversity. To enable systematic evaluation, we formalize an evolving-environment streaming dataset (EESD), in which agents must infer, apply, and revise latent environment knowledge from interaction and outcome feedback as hidden policies evolve, and introduce ServeLearnBench, spanning retail support, banking, and sales-pitch generation with 53 environment windows and 7,718 tasks. We evaluate five learning harnesses (RAG, Mem0, SkillOpt, Continual Harness, and Prime) across six models (GPT-5.6 Terra, Opus 5, Kimi K3, GLM-5.3, DeepSeek V4.1 Flash, and GLM-5.3 Flash), covering 28 model-harness pairs and 252 learning runs. Our evaluation reveals three main findings: a substantial gap remains between task capability and learning from experience; continual adaptation is costly and can degrade already-correct behavior; and insufficient exploration emerges as a key bottleneck to effective adaptation. Overall, ServeLearnBench provides a controlled testbed for diagnosing these limitations and tracking progress toward agents that continually and reliably improve through serving experience.

Post-training quantization reduces the cost of deploying language-model agents, but its effect on recovery from temporary tool failures can depend on how recovery is evaluated. We compare 8-bit and 4-bit variants of Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct on twenty deterministic tool-use tasks and five prompts. The 8-bit-4-bit recovery comparison changes direction across prompts and evaluation targets. On tasks that both variants complete without faults under the same prompt, the difference ranges from 0 to +20.2 percentage points for Llama and from -50.0 to +35.0 points for Qwen. Full-pipeline point estimates favor 8-bit Llama under all five prompts, whereas the Qwen comparison changes direction across prompts. The evaluation target can also reverse the result. For Llama under one prompt, scoring each variant only on its own clean-passing tasks favors 4-bit by 17.5 points; scoring the same tasks for both variants gives no difference, while scoring the full pipeline favors 8-bit by 28.3 points. Executor leniency is a third such choice. Rescoring the same logs with strict output parsing, which 8-bit Llama violates far more often than 4-bit Llama under that prompt, turns that +28.3 into -15.0 while leaving Qwen essentially unchanged. These findings show that one prompt, one screened task set, and one scoring policy do not establish a stable conclusion about quantized-agent robustness. Evaluations should compare variants on matched tasks, report full-pipeline success for deployment decisions, state the scoring policy, and quantify uncertainty across tasks rather than injected fault sites.

Tool-using agents make consequential changes to external state, yet correct outcomes do not guarantee that their actions were supported by evidence established beforehand. We study where this evidence-to-action chain breaks as agents move from deciding whether to act to executing single actions and dependent workflows. Across ten model-harness configurations, strong static action assessment can coexist with much weaker interactive execution. Failures often begin before execution: agents stop with incomplete investigation or act before required evidence is established. Once required evidence is obtained, single-action execution is usually reliable, while multi-action workflows additionally expose unresolved prerequisites and incomplete execution. For this analysis, we introduce SafeActBench, comprising 656 cases across six operational domains and five protocols that progress from static action judgment and investigated non-action to single- and multi-action workflows. A provenance-bound Evidence Ledger and deterministic trajectory evaluator track what information was established, when actions occurred, and whether downstream dependencies were satisfied. These results show that failures arise not only from missing information, but also from how agents use established evidence when deciding and executing actions.

The rapid advancement of LLM agents has enabled systems to autonomously perform complex tasks through external tools, but their growing access to personal data introduces significant privacy risks. Existing benchmarks primarily evaluate LLM agent privacy through simulated trajectories and outcome-based metrics, limiting their ability to capture privacy risks arising during multi-step agent execution. In this work, we introduce AgentPrivArena, a framework for evaluating privacy risks in realistic LLM agent workflows. AgentPrivArena integrates authentic MCP tools and self-hosted services within a reproducible execution environment. We further propose trajectory-level privacy metrics that quantify unnecessary information access beyond final response leakage. Building on this framework, we introduce AgentPrivAudit, a runtime auditing approach for monitoring privacy violations during agent execution. Extensive experiments on state-of-the-art LLM agents reveal substantial privacy risks overlooked by existing evaluation paradigms, highlighting the importance of trajectory-level auditing for trustworthy agent deployment.

LLM deployment is shifting from single-turn completion to agentic trajectories in which a model plans, calls tools, reads results and reasons at test time before acting. This inverts the economics of inference hardware: chat serving amortises weight reads across large batches, whereas agent trajectories are sequentially dependent, run at effective batch one, and make per-token decode latency (TPOT) the dominant term in task completion time. Using a roofline analysis and a closed-form episode-latency model, we show why this regime favours accelerators that keep weights in on-die SRAM (Cerebras WSE-3/3T, Groq/NVIDIA LPU) or compiler-managed tiered memory (SambaNova SN40L/SN50), and why three vendor ecosystems converged in 2026 on disaggregated prefill/decode serving. We show that per-step reliability compounds exponentially in trajectory length-a 2% per-step failure rate erases a 2x decode advantage for a 20-step agent-so determinism and tail latency are first-order performance variables. We then propose a four-layer evaluation framework (silicon, serving system, agent episode, enterprise) with a metric set built on goodput at an agentic SLO and cost per successful episode, a six-axis benchmark protocol over six task families, a paired-bootstrap statistical design, an attestation protocol for vendor-run benchmarks, and TCO, availability and adoption-timing models with explicit break-even conditions. All performance figures are public and labelled by evidence class; we state seven falsifiable hypotheses and the experiments that test them, and argue that the most likely original result is that token-throughput rankings diverge from cost-per-successful-task rankings on long-horizon work.

An agent whose tool keeps returning nothing useful should stop relying on it. In a retrieval environment with controlled source failures, we separate how agents judge results from what they do. We compare stopping at the same step after longer and shorter runs of results the agent judged useless; this contrast is zero for clock- or deadline-driven stopping. Where we record their judgments, the seven agents we test call a failing source's results useless 97-100% of the time, yet most of them rarely stop on that judgment. Prompt cues change when they stop but not what they stop on. Permission to answer from memory and a reasoning mode can bring early stops regardless of evidence, a stated budget moves the 7-8B models' stops to the deadline, and a stopping rule or call cost in the prompt is followed at most partly. Stopping follows the evidence only when the harness enforces an integration step that makes the agent answer after five consecutive results it judged useless. This step raises failing-source success for every model, keeps the stopping point fixed when the budget doubles, and needs no extra judgment call when the agent states its judgments. A pre-registered replication on 300 fresh questions confirms the dissociation and the rule's effect.

LLM agents are increasingly built for medical work and scored on clinical benchmarks. Each such score, however, comes from a model running inside an agent harness, the system that controls the loop between the model and its environment. An agent's score is therefore a property of a model--harness pair. For medical agents, how much outcomes change with the harness has rarely been measured. Measuring this change, and explaining it, raises two challenges. First, a harness comparison must change nothing but the harness and be repeated across models and kinds of task. Second, comparing whole harnesses leaves their mechanisms bundled together, so it cannot show when an individual mechanism helps. To address these challenges, we present MedicalHarness, a controlled study of models and agent harnesses on medical tasks. We first build MedicalHarnessBench to evaluate agents on $107$ tasks across four domains that each test a different harness capability. Using this benchmark, we run five open-weight models under five agent harnesses, changing only the harness within a comparison, and analyze both outcomes and execution traces. To study individual mechanisms, we build MH-Lab, a controlled harness that switches off context management, planning or tool exposure one at a time within a shared execution loop. We find that the harness and its interaction with the model account for about a quarter of the outcome variance, and that no single harness is best across models and tasks. Code and data are available at https://github.com/REAL-Lab-NU/MedicalHarness.