LLM Agent Evaluation

LLM: Large Language Model

Ask about this topic

Overview

The field studies whether language-model agents can use tools, manage state, follow instructions, and achieve users’ goals across multi-step interactions. Benchmarks place agents in executable settings such as coding repositories, websites, terminals, business workflows, and medical or scientific tasks, while methods use hidden tests, state checks, trajectory analysis, and simulated users to assess performance. Evaluation increasingly includes cost, latency, privacy, safety, consistency, and process quality—not just whether the final answer is correct. Results remain difficult to compare because harnesses and scoring choices can change rankings, and strong performance on short or isolated tasks often fails to transfer to long-horizon real-world work.

Latest papers

CardsList

1,217 papers231 papers in the last four weeks, up 62% on the four weeks before. 1.3% of all new papers.

Mar 6, 2026

Talk Freely, Execute Strictly: Schema-Gated Agentic AI for Flexible and Reproducible Scientific Workflows

Large language models (LLMs) can now translate a researcher's plain-language goal into executable computation, yet scientific workflows demand determinism, provenance, and governance that are difficult to guarantee when an LLM decides what runs. Semi-structured interviews with 18 experts across 10 industrial R&D stakeholders surface 2 competing requirements--deterministic, constrained execution and conversational flexibility without workflow rigidity--together with boundary properties (human-in-the-loop control and transparency) that any resolution must satisfy. We propose schema-gated orchestration as the resolving principle: the schema becomes a mandatory execution boundary at the composed-workflow level, so that nothing runs unless the complete action--including cross-step dependencies--validates against a machine-checkable specification. We operationalize the 2 requirements as execution determinism (ED) and conversational flexibility (CF), and use these axes to review 20 systems spanning 5 architectural groups along a validation-scope spectrum. Scores are assigned via a multi-model protocol--15 independent sessions across 3 LLM families--yielding substantial-to-near-perfect inter-model agreement (Krippendorff a=0.80 for ED and a=0.98 for CF), demonstrating that multi-model LLM scoring can serve as a reusable alternative to human expert panels for architectural assessment. The resulting landscape reveals an empirical Pareto front--no reviewed system achieves both high flexibility and high determinism--but a convergence zone emerges between the generative and workflow-centric extremes. We argue that a schema-gated architecture, separating conversational from execution authority, is positioned to decouple this trade-off, and distill 3 operational principles--clarification-before-execution, constrained plan-act orchestration, and tool-to-workflow-level gating--to guide adoption.

Joel Strickland, Arjun Vijeta, Chris Moores +7

Intellegens, The Studio, Chesterton Mill, Cambridge, CB4 3NP, UK · Independent Researcher · Cavendish Laboratory, J.J. Thomson Avenue, Cambridge, UK

LLM Agent OrchestrationScientific Workflow AutomationLLM Agent Evaluation
AskPDF
Mar 5, 2026

Tool-Genesis: A Task-Driven Tool Creation Benchmark for Self-Evolving Language Agent

Research on self-evolving language agents has accelerated, drawing increasing attention to their ability to create, adapt, and maintain tools from task requirements. However, existing benchmarks predominantly rely on predefined specifications, which limits scalability and hinders truly autonomous evolution. While recent studies attempt to dynamically generate tools, they primarily emphasize downstream performance, resulting in a "black-box" evaluation that makes it difficult to attribute failures to specific causes. To address this, we propose Tool-Genesis, a diagnostic benchmark designed to quantify agent capabilities across multiple dimensions, including interface compliance, functional correctness, and downstream utility. Tool-Genesis evaluates whether agents can construct task-relevant tools solely from abstract requirements (without preset specifications) and use them to solve realistic problems. Crucially, we find that even state-of-the-art models struggle to produce precise tool interfaces or executable logic in a one-shot setting. These minor initial flaws are amplified through the pipeline, leading to a sharp degradation in downstream metrics. We hope Tool-Genesis will guide future research toward training and steering models to synthesize persistent, general-purpose tools that better address real-world challenges.

Bowei Xia, Mengkang Hu, Shijian Wang +5

UESTC · The University of Hong Kong · Xiaohongshu Inc.

Tool-Using AgentsAI Agent BenchmarksSelf-Evolving AgentsLLM Agent Evaluation
AskPDF
Mar 5, 2026

Evaluating the Search Agent in a Parallel World

Integrating web search tools has significantly extended the capability of LLMs to address open-world, real-time, and long-tail problems. However, evaluating these Search Agents presents formidable challenges. First, constructing high-quality deep search benchmarks is prohibitively expensive, while unverified synthetic data often suffers from unreliable sources. Second, static benchmarks face dynamic obsolescence: as internet information evolves, complex queries requiring deep research often degrade into simple retrieval tasks due to increased popularity, and ground truths become outdated due to temporal shifts. Third, attribution ambiguity confounds evaluation, as an agent's performance is often dominated by its parametric memory rather than its actual search and reasoning capabilities. Finally, reliance on specific commercial search engines introduces variability that hampers reproducibility. To address these issues, we propose a novel framework, Mind-ParaWorld, for evaluating Search Agents in a Parallel World. Specifically, MPW samples real-world entity names to synthesize future scenarios and questions situated beyond the model's knowledge cutoff. A ParaWorld Law Model then constructs a set of indivisible Atomic Facts and a unique ground-truth for each question. During evaluation, instead of retrieving real-world results, the agent interacts with a ParaWorld Engine Model that dynamically generates SERPs grounded in these inviolable Atomic Facts. We release MPW-Bench, an interactive benchmark spanning 19 domains with 1,608 instances. Experiments across three evaluation settings show that, while search agents are strong at evidence synthesis given complete information, their performance is limited not only by evidence collection and coverage in unfamiliar search environments, but also by unreliable evidence sufficiency judgment and when-to-stop decisions-bottlenecks.

Jiawei Chen, Xintian Shen, Lihao Zheng +7

LLM Agent EvaluationWeb Search AgentsAI Agent Benchmarks
AskPDF
Mar 5, 2026

DARE: Aligning LLM Agents with the R Statistical Ecosystem via Distribution-Aware Retrieval

Large Language Model (LLM) agents can automate data-science workflows, but many rigorous statistical methods implemented in R remain underused because LLMs struggle with statistical knowledge and tool retrieval. Existing retrieval-augmented approaches focus on function-level semantics and ignore data distribution, producing suboptimal matches. We propose DARE (Distribution-Aware Retrieval Embedding), a lightweight, plug-and-play retrieval model that incorporates data distribution information into function representations for R package retrieval. Our main contributions are: (i) RPKB, a curated R Package Knowledge Base derived from 8,191 high-quality CRAN packages; (ii) DARE, an embedding model that fuses distributional features with function metadata to improve retrieval relevance; and (iii) RCodingAgent, an R-oriented LLM agent for reliable R code generation and a suite of statistical analysis tasks for systematically evaluating LLM agents in realistic analytical scenarios. Empirically, DARE achieves an NDCG at 10 of 93.47%, outperforming state-of-the-art open-source embedding models by up to 17% on package retrieval while using substantially fewer parameters. Integrating DARE into RCodingAgent yields significant gains on downstream analysis tasks. This work helps narrow the gap between LLM automation and the mature R statistical ecosystem.

Maojun Sun, Yue Wu, Yifei Xie +5

Department of Data Science and Artificial Intelligence, The Hong Kong Polytechnic University, Hong Kong SAR, China · Department of Applied Mathematics, The Hong Kong Polytechnic University, Hong Kong SAR, China.

Data Analysis AgentsLLM Agent Evaluation
AskPDF
Mar 4, 2026

iAgentBench: Benchmarking Sensemaking Capabilities of Information-Seeking Agents on High-Traffic Topics

With the emergence of search-enabled generative QA systems, users are increasingly turning to tools that browse, aggregate, and reconcile evidence across multiple sources on their behalf. Yet many widely used QA benchmarks remain answerable by retrieving a single relevant passage, making them poorly suited for measuring cross-source sensemaking, such as integrating evidence, tracking causal links, and resolving dependencies across facets of a topic. We present iAgentBench, a dynamic ODQA benchmark that targets these higher-level information needs while keeping questions natural and grounded in realistic information-seeking behavior. iAgentBench draws seed topics from real-world attention signals and uses common user intent patterns to construct user-like questions whose answers require combining evidence from multiple sources, not just extracting a single snippet. Each instance is released with traceable evidence and auditable intermediate artifacts that support contamination checks and enable fine-grained diagnosis of failures in retrieval versus synthesis. Experiments across multiple LLMs show that retrieval improves accuracy, but retrieval alone does not reliably resolve these questions, underscoring the need to evaluate evidence use, not just evidence access.

Preetam Prabhu Srikar Dammu, Arnav Palkhiwala, Tanya Roosta +1

University of Washington Seattle, Washington, USA · UC Berkeley Berkeley, California, USA

AI Agent BenchmarksLLM Agent EvaluationKnowledge-Intensive QAEvidence-Grounded QAMulti-Hop QAIR Evaluation
AskPDF
Mar 4, 2026

ττ-Knowledge: Evaluating Conversational Agents over Unstructured Knowledge

Conversational agents are increasingly deployed in knowledge-intensive settings, where correct behavior depends on retrieving and applying domain-specific knowledge from large, proprietary, and unstructured corpora during live interactions with users. Yet most existing benchmarks evaluate retrieval or tool use independently of each other, creating a gap in realistic, fully agentic evaluation over unstructured data in long-horizon interactions. We introduce ττ-Knowledge, an extension of ττ-Bench for evaluating agents in environments where success depends on coordinating external, natural-language knowledge with tool outputs to produce verifiable, policy-compliant state changes. Our new domain, ττ-Banking, models realistic fintech customer support workflows in which agents must navigate roughly 700 interconnected knowledge documents while executing tool-mediated account updates. Across embedding-based retrieval and terminal-based search, even frontier models with high reasoning budgets achieve only ∼\sim25.5% pass^1, with reliability degrading sharply over repeated trials. Agents struggle to retrieve the correct documents from densely interlinked knowledge bases and to reason accurately over complex internal policies. Overall, ττ-Knowledge provides a realistic testbed for developing agents that integrate unstructured knowledge in human-facing deployments.

Quan Shi, Alexandra Zytek, Pedram Razavi +2

Sierra · Princeton University

LLM Agent EvaluationAgent EvaluationConversational AgentsLong-Horizon Agent EvaluationKnowledge-Intensive QA
AskPDF
Mar 4, 2026

CodeTaste: Can LLMs Generate Human-Level Code Refactorings?

LLM coding agents can generate working code, but their solutions often accumulate complexity, duplication, and architectural debt. Human developers address such issues through refactoring: behavior-preserving program transformations that improve structure and maintainability. We investigate whether agents (i) can execute refactorings reliably and (ii) identify the refactorings that human developers actually chose in real codebases. To this end, we construct CodeTaste, a benchmark mined from large multi-file open-source refactorings. To score solutions, we combine repository test suites that measure functional correctness with tailored static checks that verify removal of undesired and introduction of desired code patterns using dataflow reasoning. Our results show a clear gap: agents perform well at implementing refactorings that are specified in detail, but often fail to discover the human refactoring choices when given a focus area for changes. A propose-then-implement decomposition improves alignment, and selecting the best-aligned proposal before implementation can yield further gains. CodeTaste provides an evaluation target and a potential preference signal for aligning coding agents with human refactoring decisions in realistic codebases. We release the benchmark, leaderboard, and code.

Alex Thillen, Niels Mündler, Veselin Raychev +1

Department of Computer Science, ETH Zurich · LogicStar.ai, Zurich, Switzerland

Automated Code RefactoringAutomated Software EngineeringLLM Agent EvaluationCode Generation BenchmarksRepository-Level Code Generation
AskPDF
Mar 4, 2026

On the Suitability of LLM-Driven Agents for Dark Pattern Audits

As LLM-driven agents begin to autonomously navigate the web, their ability to interpret and respond to manipulative interface design becomes critical. A fundamental question that emerges is: can such agents reliably recognize patterns of friction, misdirection, and coercion in interface design (i.e., dark patterns)? We study this question in a setting where the workflows are consequential: website portals associated with the submission of CCPA-related data rights requests. These portals operationalize statutory rights, but they are implemented as interactive interfaces whose design can be structured to facilitate, burden, or subtly discourage the exercise of those rights. We design and deploy an LLM-driven auditing agent capable of end-to-end traversal of rights-request workflows, structured evidence gathering, and classification of potential dark patterns. Across a set of 456 data broker websites, we evaluate: (1) the ability of the agent to consistently locate and complete request flows, (2) the reliability and reproducibility of its dark pattern classifications, and (3) the conditions under which it fails or produces poor judgments. Our findings characterize both the feasibility and the limitations of using LLM-driven agents for scalable dark pattern auditing.

Chen Sun, Yash Vekaria, Rishab Nithyanand

University of Iowa · University of California, Davis

LLM Agent ReliabilityLLM Agent Evaluation
AskPDF
Mar 3, 2026

Benchmarking Emergent Coordination in Large-Scale LLM Populations: An Evaluation Framework on the MoltBook Archive

As multi-agent Large Language Model (LLM) systems scale, evaluating their emergent coordination dynamics becomes increasingly critical. However, current evaluation paradigms-focused on single agents or small, explicitly structured groups-fail to capture the self-organization and viral information dynamics that arise in large, decentralized populations. We introduce a systematic evaluation framework to benchmark role specialization, information diffusion, and cooperative task resolution in open agent environments. We demonstrate this framework on the MoltBook Observatory Archive, a dataset of 2.73M interactions among 90,704 autonomous agents, establishing quantitative baselines for emergent coordination. Our evaluation reveals a pronounced core-periphery structure (silhouette 0.91), heavy-tailed cascade distributions (α=2.57α= 2.57), and severe coordination overhead in decentralized task resolution (Cohen's d=−0.88d = -0.88 against a single-agent baseline). By providing standardized evaluation tasks and empirical baselines, our framework enables the rigorous comparison of future multi-agent protocols and establishes evaluation itself as an object of scientific study.

Brandon Yee, Pairie Koh

Management Lab, Yee Collins Research Group

Multi-Agent CoordinationLLM Agent EvaluationEmergent Behavior in Multi-Agent SystemsAI Agent Benchmarks
AskPDF
Mar 3, 2026

BeyondSWE: Can Current Code Agent Survive Beyond Single-Repo Bug Fixing?

Current code-agent benchmarks primarily evaluate localized issue resolution within a single target repository, leaving under-tested many software engineering tasks that require external knowledge or broader repository-level changes. We introduce BeyondSWE, a 500-instance benchmark drawn from 246 real-world GitHub repositories to evaluate code agents beyond single-repository bug fixing. BeyondSWE covers four representative settings: cross-repository issue resolution, domain-specific issue resolution, dependency-driven migration, and document-to-repository generation, spanning both broader knowledge scope and broader resolution scope. Our evaluation shows that BeyondSWE remains far from saturated: the best OpenHands-based agent reaches 46.12 average score, while the strongest Codex harness with GPT-5.4 (xhigh) reaches 56.65 under a search-aware prompt. To study whether external information access closes this gap, we use SearchSWE as a controlled diagnostic baseline for search-augmented coding. Search access improves most models and substantially helps some tasks, but the gains remain limited and uneven, showing that current agents still struggle to convert retrieved information into precise, version-compatible, and locally actionable code changes. These results suggest that deep search for coding remains an open problem: progress requires agents that can reliably combine external evidence with repository-local reasoning and execution-based verification.

Guoxin Chen, Fanzhe Meng, Jiale Zhao +12

Retrieval-Augmented Code GenerationSoftware Engineering BenchmarksLLM Agent EvaluationCode Generation BenchmarksCode Generation
AskPDF
Mar 3, 2026

Beyond Task Completion: Revealing Corrupt Success in LLM Agents through Procedure-Aware Evaluation

Large Language Model (LLM)-based agents are increasingly adopted in high-stakes settings, but current benchmarks evaluate mainly whether a task was completed, not how. We introduce Procedure-Aware Evaluation (PAE), a framework that formalizes agent procedures as structured observations and exposes consistency relationships between what agents observe, communicate, and execute. PAE evaluates agents along complementary axes (Utility, Efficiency, Interaction Quality, Procedural Integrity) and applies multi-dimensional gating that categorically disqualifies corrupt outcomes. Evaluating state-of-the-art LLM agents on tau-bench yields findings at the axis, compliance, and benchmark levels. At the axis level, the dimensions capture non-redundant failure modes: utility masks reliability gaps, speed does not imply precision, and conciseness does not predict intent adherence. At the procedural compliance level, 27-78% of benchmark reported successes are corrupt successes concealing violations across interaction and integrity. Furthermore, gating substantially collapses Pass^4 rate and affects model rankings. The analysis of corrupt success cases reveals distinctive per-model failure signatures: GPT-5 spreads errors across policy, execution, and intent dimensions; Kimi-K2-Thinking concentrates 78% of violations in policy faithfulness and compliance; and Mistral-Large-3 is dominated by faithfulness failures. At the benchmark level, our analysis exposes structural flaws in the benchmark design, including task scope gaps, contradictory reward signals, and simulator artifacts that produce accidental successes.

Hongliu Cao, Ilias Driouich, Eoin Thomas

Amadeus France

LLM Agent EvaluationLLM Agent ReliabilityBenchmark Validity
AskPDF
Mar 3, 2026

Agentified Assessment of Logical Reasoning Agents

We present a framework for evaluating and benchmarking logical reasoning agents when assessment itself must be reproducible, auditable, and robust to execution failures. Building on agentified assessment, we use an assessor agent to issue tasks, enforce execution budgets, parse outputs, and record structured failure types, while the agent under test only needs to expose a standardized agent-to-agent interface. As a case study, we benchmark an auto-formalization agent for first-order logic (FOL) reasoning on a solver-verified and repaired split of FOLIO. The agent translates natural language premises and conclusions into executable Z3Py programs and employs satisfiability modulo theories (SMT) solving to determine logical entailment. On the cleaned FOLIO validation set, the auto-formalization agent achieves 86.70% accuracy under the assessor protocol, outperforming a chain-of-thought baseline (73.89%).

Zhiyu Ni, Yifeng Xiao, Zheng Liang

University of California, Berkeley

LLM Agent EvaluationAI Agent BenchmarksLogical ReasoningAutoformalization
AskPDF
Mar 2, 2026

Same Pieces, Different Servers: A Tetris Benchmark for AI Agents as Served

An agent meets a model as served: through an endpoint with a price card, a shared cache and other tenants, or on whatever hardware a self-hosted model runs. Benchmarks rank the weights. We introduce a Tetris benchmark that measures what agents get from models as served: every move is scored against an oracle, and every agent receives the same pieces. In five pre-specified experiments with nine open-weight models on one serverless provider, plus an open decision model self-hosted on a CPU, we find that price and size do not predict decision quality; that resending history costs almost nothing when cached input is free, would cost eleven to twelve times more if it were not, and makes play worse; that deployments keep between 4 and more than 64 agent contexts warm, in line with their throughput rather than the model's KV-cache size; that an agent's own long requests slow its slowest short decisions more than tenfold, which a simple admission rule cuts by 59% without hurting play; and that the decision model plays mid-pack but takes 27 s per move on a CPU, about 650 times longer than reported on a GPU. Architecture predicts some of what an agent sees; the deployment sets the rest.

Haochuan Wang

Massachusetts Institute of Technology

LLM Inference EfficiencyLLM Agent EvaluationLLM Inference ServingAI Agent BenchmarksLLM Agent Serving
AskPDF
Mar 2, 2026

Evaluating and Understanding Scheming Propensity in LLM Agents

As frontier language models are increasingly deployed as autonomous agents pursuing complex, long-term objectives, there is increased risk of scheming: agents covertly pursuing misaligned goals. Prior work has focused on showing agents are capable of scheming, but their propensity to scheme in realistic scenarios remains underexplored. To understand when agents scheme, we decompose scheming incentives into agent factors and environmental factors. We develop realistic settings allowing us to systematically vary these factors, each with scheming opportunities for agents that pursue instrumentally convergent goals such as self-preservation, resource acquisition, and goal-guarding. We find only minimal instances of scheming despite high environmental incentives, and show this is unlikely due to evaluation awareness. While inserting adversarially-designed prompt snippets that encourage agency and goal-directedness into an agent's system prompt can induce high scheming rates, snippets used in real agent scaffolds rarely do. Surprisingly, in model organisms (Hubinger et al., 2023) built with these snippets, scheming behavior is remarkably brittle: removing a single tool can drop the scheming rate from 59% to 3%, and increasing oversight can raise rather than deter scheming by up to 25%. Our incentive decomposition enables systematic measurement of scheming propensity in settings relevant for deployment, which is necessary as agents are entrusted with increasingly consequential tasks.

Mia Hopman, Jannes Elstner, Maria Avramidou +2

LASR Labs · Google DeepMind

AI Agent SafetyLLM AgentsLLM Agent Evaluation
AskPDF
Mar 1, 2026

Can AI Agents Agree?

Large language models are increasingly deployed as cooperating agents, yet their behavior in adversarial consensus settings has not been systematically studied. We evaluate LLM-based agents on a Byzantine consensus game over scalar values using a synchronous all-to-all simulation. We test consensus in a no-stake setting where agents have no preferences over the final value, so evaluation focuses on agreement rather than value optimality. Across hundreds of simulations spanning model sizes, group sizes, and Byzantine fractions, we find that valid agreement is not reliable even in benign settings and degrades as group size grows. Introducing a small number of Byzantine agents further reduces success. Failures are dominated by loss of liveness, such as timeouts and stalled convergence, rather than subtle value corruption. Overall, the results suggest that reliable agreement is not yet a dependable emergent capability of current LLM-agent groups even in no-stake settings, raising caution for deployments that rely on robust coordination.

Frédéric Berdoz, Leonardo Rugli, Roger Wattenhofer

ETH Zurich, Switzerland

Multi-Agent ConsensusLLM Agent ReliabilityLLM Agent Evaluation
AskPDF
Mar 1, 2026

Evaluating Agents Across Runtime Contracts: When Mismatch Costs Efficiency or Quality

In CodeAct, language-model agents write Python that calls tools and use execution feedback to choose actions. Persistent runtimes preserve Python variables between actions; stateless runtimes clear them without resetting task progress. Training traces demonstrate task-solving strategies and runtime-specific ways to store and recover intermediate results. We study runtime transfer: whether agents trained under one contract remain effective under the other, a dependence that fixed-runtime evaluations can conceal. Across three tasks requiring a working record built from tool feedback, we fine-tune separate Qwen3-8B agents per task and runtime on instance-paired persistent and stateless traces and evaluate all four training-deployment combinations. We vary the per-turn tool-call cap: how many calls one Python action may execute. In our primary task, agents inspect hidden item attributes and select a high-value subset under a weight limit. At 80 calls per turn, a persistent-trained agent deployed statelessly scores 0.61 of optimal value against 0.77 under matched persistent deployment, but uses 4.4 times as many tokens rebuilding its record. Tightening the cap to 25 reduces mismatched quality to 0.07 while matched quality remains 0.66. Trace replay shows that the cap interrupts inspection, the reset erases the partial record, and rebuilding consumes calls needed for progress. The widening gap replicates across training seeds, a second rollout, and two additional base models. Across the settings studied, cap tightening produces a mismatch-specific collapse only where interruptions are frequent and matched execution resumes while mismatched execution restarts. Evaluations of trace-fine-tuned agents should specify episode limits, test runtime transfer across caps, and report cost alongside quality; otherwise, the same mismatch may appear as redundant computation or near-total quality loss.

Victor May, Van Khue Nguyen, Aaditya Salgarkar +3

Ontocord · École Polytechnique · Carnegie Mellon University +3

Tool-Using AgentsLLM Agent EvaluationLLM Agent Training
AskPDF
Feb 28, 2026

SkillCraft: Can LLM Agents Learn to Use Tools Skillfully?

Real-world tool-using agents operate over long-horizon workflows with recurring structure and diverse demands, where effective behavior requires not only invoking atomic tools but also abstracting, and reusing higher-level tool compositions. However, existing benchmarks mainly measure instance-level success under static tool sets, offering limited insight into agents' ability to acquire such reusable skills. We address this gap by introducing SkillCraft, a benchmark explicitly stress-test agent ability to form and reuse higher-level tool compositions, where we call Skills. SkillCraft features realistic, highly compositional tool-use scenarios with difficulty scaled along both quantitative and structural dimensions, designed to elicit skill abstraction and cross-task reuse. We further propose a lightweight evaluation protocol that enables agents to auto-compose atomic tools into executable Skills, cache and reuse them inside and across tasks, thereby improving efficiency while accumulating a persistent library of reusable skills. Evaluating state-of-the-art agents on SkillCraft, we observe substantial efficiency gains, with token usage reduced by up to 80% by skill saving and reuse. Moreover, success rate strongly correlates with tool composition ability at test time, underscoring compositional skill acquisition as a core capability.

Shiqi Chen, Jingze Gai, Ruochen Zhou +13

University of Oxford · City University of Hong Kong · Hong Kong University of Science and Technology +3

Tool-Using AgentsTool UseLLM Agent Skill LearningLLM Agent EvaluationAI Agent BenchmarksLLM Agent Skill Retrieval
AskPDF
Feb 28, 2026

RAVEL: Reasoning Agents for Validating and Evaluating LLM Text Synthesis

Large Language Models have evolved from single-round generators into long-horizon agents, capable of complex text synthesis scenarios. However, current evaluation frameworks lack the ability to assess the actual synthesis operations, such as outlining, drafting, and editing. Consequently, they fail to evaluate the actual and detailed capabilities of LLMs. To bridge this gap, we introduce RAVEL, an agentic framework that enables the LLM testers to autonomously plan and execute typical synthesis operations, including outlining, drafting, reviewing, and refining. Complementing this framework, we present C3EBench, a comprehensive benchmark comprising 1,258 samples derived from professional human writings. We utilize a "reverse-engineering" pipeline to isolate specific capabilities across four tasks: Cloze, Edit, Expand, and End-to-End. Through our analysis of 14 LLMs, we uncover that most LLMs struggle with tasks that demand contextual understanding under limited or under-specified instructions. By augmenting RAVEL with SOTA LLMs as operators, we find that such agentic text synthesis is dominated by the LLM's reasoning capability rather than raw generative capacity. Furthermore, we find that a strong reasoner can guide a weaker generator to yield higher-quality results, whereas the inverse does not hold. Our code and data are available at this link: https://github.com/ZhuoerFeng/RAVEL-Reasoning-Agents-Text-Eval.

Andrew Zhuoer Feng, Cunxiang Wang, Yu Luo +9

Tsinghua University · Peking University

Language Model Generation EvaluationLLM Agent Evaluation
AskPDF
Feb 28, 2026

EMPA: Evaluating Persona-Aligned Empathy as a Process

Evaluating persona-aligned empathy in LLM-based dialogue agents remains challenging. User states are latent, feedback is sparse and difficult to verify in situ, and seemingly supportive turns can still accumulate into trajectories that drift from persona-specific needs. We introduce EMPA, a process-oriented framework that evaluates persona-aligned support as sustained intervention rather than isolated replies. EMPA distills real interactions into controllable, psychologically grounded scenarios, couples them with an open-ended multi-agent sandbox that exposes strategic adaptation and failure modes, and scores trajectories in a latent psychological space by directional alignment, cumulative impact, and stability. The resulting signals and metrics support reproducible comparison and optimization of long-horizon empathic behavior, and they extend to other agent settings shaped by latent dynamics and weak, hard-to-verify feedback.

Shiya Zhang, Yuhan Zhan, Ruixi Su +4

Team Echo, Nature Select · Sun Yat-sen University

LLM Agent EvaluationMulti-Turn Dialogue EvaluationLong-Horizon Agent Evaluation
AskPDF
Feb 27, 2026

COOP2^2: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems

Many complex tasks require extended effort, diverse capabilities, or coordinated actions beyond what a single agent can provide. However, simply adding more agents does not guarantee better performance, as effective cooperation depends on how agents interact with each other and with task structure to satisfy evolving constraints over time. This challenge is amplified for LLM-based multi-agent systems (LLM-MAS): plans, messages, and revisions occur in natural language, whereas task progress depends on grounded environment actions. Current evaluations mostly treat cooperation as an implicit ingredient of final task success, leaving both cooperation and the effect of multi-agent interaction on task dynamics difficult to study. We introduce COOP2^2, an evaluation framework that grounds high-level agent cooperation dynamics in LLM-MAS within task progress in the environment. COOP2^2 then defines cooperative tasks with verifiable cooperative requirements, allowing us to analyze how cooperation unfolds over time with respect to task progress, as well as where and why cooperation breaks down. Building on this framework, we develop COOP2^2-Repair, which predicts constraint failures from group plans and opens targeted repair channels for guided revisions. Across two environments and three communication structures, COOP2^2-Repair improves task success and constraint satisfaction while exposing the additional decision overhead and communication load required for repair. The project web page can be found at: https://happyeureka.github.io/coop2.

Hanqing Yang, Narjes Nourzad, Shiyu Chen +3

CMU · USC · SUTD +1

Multi-Agent CollaborationMulti-Agent CoordinationLLM Agent EvaluationMulti-Agent PlanningMulti-Agent CommunicationLLM Planning
AskPDF
Feb 26, 2026

Toward Personalized LLM-Powered Agents: Foundations, Evaluation, and Future Directions

Large language models have enabled agentic systems that reason, plan, and interact with tools and environments to accomplish complex tasks. As these agents operate over extended interaction horizons, their effectiveness increasingly depends on adapting behavior to individual users and maintaining continuity across interactions, giving rise to personalized LLM-powered agents (PLAs). In such long-term, user-dependent settings, personalization permeates the entire decision pipeline rather than remaining confined to surface-level response generation. This survey provides a capability-oriented review of personalized LLM-powered agents. Existing work is organized around four interdependent capabilities: profile modeling, memory, planning, and action execution. Using this taxonomy, representative methods are synthesized and analyzed to illustrate how user signals are represented, propagated, and utilized across the agent pipeline, highlighting cross-component interactions and recurring design challenges. Evaluation metrics and benchmarking paradigms tailored to personalized agents are further examined, along with application scenarios ranging from conversational assistants to domain-specific expert systems. By clarifying the design space of personalization in agent systems, this survey provides a structured foundation for developing more user-aligned, adaptive, and deployable LLM-powered agents.

Yue Xu, Qian Chen, Zizhan Ma +5

ShanghaiTech University, China · Tongji University, China · The Chinese University of Hong Kong, China +3

LLM AgentsLLM PersonalizationLLM Agent EvaluationMemory-Augmented Language Model AgentsUser Preference Modeling
AskPDF
Feb 26, 2026

MobilityBench: A Benchmark for Evaluating Route-Planning Agents in Real-World Mobility Scenarios

Route-planning agents powered by large language models (LLMs) have emerged as a promising paradigm for supporting everyday human mobility through natural language interaction and tool-mediated decision making. However, systematic evaluation in real-world mobility settings is hindered by diverse routing demands, non-deterministic mapping services, and limited reproducibility. In this study, we introduce MobilityBench, a scalable benchmark for evaluating LLM-based route-planning agents in real-world mobility scenarios. MobilityBench is constructed from large-scale, anonymized real user queries collected from Amap and covers a broad spectrum of route-planning intents across multiple cities worldwide. To enable reproducible, end-to-end evaluation, we design a deterministic API-replay sandbox that eliminates environmental variance from live services. We further propose a multi-dimensional evaluation protocol centered on outcome validity, complemented by assessments of instruction understanding, planning, tool use, and efficiency. Using MobilityBench, we evaluate multiple LLM-based route-planning agents across diverse real-world mobility scenarios and provide an in-depth analysis of their behaviors and performance. Our findings reveal that current models perform competently on Basic information retrieval and Route Planning tasks, yet struggle considerably with Preference-Constrained Route Planning, underscoring significant room for improvement in personalized mobility applications. We publicly release the benchmark data, evaluation toolkit, and documentation at https://github.com/AMAP-ML/MobilityBench.

Zhiheng Song, Jingshuai Zhang, Chuan Qin +6

Computer Network Information Center, Chinese Academy of Sciences · AMAP, Alibaba Group · Independent Researcher

LLM Agent EvaluationAI Agent Benchmarks
AskPDF
Feb 25, 2026

A Framework for Assessing AI Agent Decisions and Outcomes in AutoML Pipelines

Agent-based AutoML systems rely on large language models to make complex, multi-stage decisions across data processing, model selection, and evaluation. However, existing evaluation practices remain outcome-centric, focusing primarily on final task performance. Through a review of prior work, we find that none of the surveyed agentic AutoML systems report structured, decision-level evaluation metrics intended for post-hoc assessment of intermediate decision quality. To address this limitation, we propose an Evaluation Agent (EA) that performs decision-centric assessment of AutoML agents without interfering with their execution. The EA is designed as an observer that evaluates intermediate decisions along four dimensions: decision validity, reasoning consistency, model quality risks beyond accuracy, and counterfactual decision impact. Across four proof-of-concept experiments, we demonstrate that the EA can (i) detect faulty decisions with an F1 score of 0.919, (ii) identify reasoning inconsistencies independent of final outcomes, and (iii) attribute downstream performance changes to agent decisions, revealing impacts ranging from -4.9% to +8.3% in final metrics. These results illustrate how decision-centric evaluation exposes failure modes that are invisible to outcome-only metrics. Our work reframes the evaluation of agentic AutoML systems from an outcome-based perspective to one that audits agent decisions, offering a foundation for reliable, interpretable, and governable autonomous ML systems.

Gaoyuan Du, Amit Ahlawat, Xiaoyang Liu +1

Amazon Web Services · Amazon Security

AI Agent EvaluationAI Agent AuditingAutomated Machine LearningLLM Agent Evaluation
AskPDF
Feb 24, 2026

A Benchmark for Deep Information Synthesis

Large language model (LLM)-based agents are increasingly used to solve complex tasks involving tool use, such as web browsing, code execution, and data analysis. However, current evaluation benchmarks do not adequately assess their ability to solve real-world tasks that require synthesizing information from multiple sources and inferring insights beyond simple fact retrieval. To address this, we introduce DEEPSYNTH, a novel benchmark designed to evaluate agents on realistic, time-consuming problems that combine information gathering, synthesis, and structured reasoning to produce insights. DEEPSYNTH contains 120 tasks collected across 7 domains and data sources covering 67 countries. DEEPSYNTH is constructed using a multi-stage data collection pipeline that requires annotators to collect official data sources, create hypotheses, perform manual analysis, and design tasks with verifiable answers. When evaluated on DEEPSYNTH, 11 state-of-the-art LLMs and deep research agents achieve a maximum F1 score of 8.97 and 17.5 on the LLM-judge metric, underscoring the difficulty of the benchmark. Our analysis reveals that current agents struggle with hallucinations and reasoning over large information spaces, highlighting DEEPSYNTH as a crucial benchmark for guiding future research.

Debjit Paul, Daniel Murphy, Milan Gritta +14

Huawei Noah’s Ark Lab, UK · Imperial College London · University of Zurich +3

LLM Agent EvaluationAI Agent Benchmarks
AskPDF
Feb 24, 2026

Can AI be a Teaching Partner? Evaluating ChatGPT, Gemini, and DeepSeek across Three Teaching Strategies

There are growing promises that Large Language Models (LLMs) can support students' learning by providing explanations, feedback, and guidance. However, despite their rapid adoption and widespread attention, there is still limited empirical evidence regarding the pedagogical skills of LLMs. This article presents a comparative study of popular LLMs, namely, ChatGPT, DeepSeek, and Gemini, acting as teaching agents. An evaluation protocol was developed, focusing on three pedagogical strategies: Examples, Explanations and Analogies, and the Socratic Method. Six human judges conducted the evaluations in the context of teaching the C programming language to beginners. The results indicate that LLM models exhibited similar interaction patterns in the pedagogical strategies of Examples and Explanations and Analogies. In contrast, for the Socratic Method, the models showed greater sensitivity to the pedagogical strategy and the initial prompt. Overall, ChatGPT and Gemini received higher scores, whereas DeepSeek obtained lower scores across the criteria, indicating differences in pedagogical performance across models.

Talita de Paula Cypriano de Souza, Shruti Mehta, Matheus Arataque Uema +2

Federal Institute of São Paulo (IFSP), Major Fernando Valle Avenue, 2013 Braganca Paulista, 12903-000, SP, Brazil · Institute of Mathematics and Computer Sciences of University of São Paulo (USP), Trabalhador São-Carlense Avenue, 400 Sao Carlos, 13566-590, Sao Paulo, Brazil · Graduate School of Education of University of Pennsylvania (UPENN), 3700 Walnut Street Philadelphia, PA 19104 Philadelphia, 610101, Pennsylvania, United States +1

AI in EducationGenerative AI in EducationLLM Agent EvaluationIntelligent Tutoring Systems
AskPDF
Feb 23, 2026

Interaction Theater: A case of LLM Agents Interacting at Scale

As multi-agent architectures and agent-to-agent protocols proliferate, a fundamental question arises: what actually happens when autonomous LLM agents interact at scale? We study this question empirically using data from Moltbook, an AI-agent-only social platform, with 800K posts, 3.5M comments, and 78K agent profiles. We combine lexical metrics (Jaccard specificity), embedding-based semantic similarity, and LLM-as-judge validation to characterize agent interaction quality. Our findings reveal agents produce diverse, well-formed text that creates the surface appearance of active discussion, but the substance is largely absent. Specifically, while most agents (67.5%67.5\%) vary their output across contexts, 65%65\% of comments share no distinguishing content vocabulary with the post they appear under, and information gain from additional comments decays rapidly. LLM judge based metrics classify the dominant comment types as spam (28%28\%) and off-topic content (22%22\%). Embedding-based semantic analysis confirms that lexically generic comments are also semantically generic. Agents rarely engage in threaded conversation (5%5\% of comments), defaulting instead to independent top-level responses. We discuss implications for multi-agent interaction design, arguing that coordination mechanisms must be explicitly designed; without them, even large populations of capable agents produce parallel output rather than productive exchange.

Sarath Shekkizhar, Adam Earle

Salesforce AI Research

Multi-Agent SystemsMulti-Agent CommunicationEmergent Behavior in Multi-Agent SystemsLLM Agent Evaluation
AskPDF
Feb 22, 2026

AgenticRAGTracer: A Hop-Aware Benchmark for Diagnosing Multi-Step Retrieval Reasoning in Agentic RAG

With the rapid advancement of agent-based methods in recent years, Agentic RAG has undoubtedly become an important research direction. Multi-hop reasoning, which requires models to engage in deliberate thinking and multi-step interaction, serves as a critical testbed for assessing such capabilities. However, existing benchmarks typically provide only final questions and answers, while lacking the intermediate hop-level questions that gradually connect atomic questions to the final multi-hop query. This limitation prevents researchers from analyzing at which step an agent fails and restricts more fine-grained evaluation of model capabilities. Moreover, most current benchmarks are manually constructed, which is both time-consuming and labor-intensive, while also limiting scalability and generalization. To address these challenges, we introduce AgenticRAGTracer, the first Agentic RAG benchmark that is primarily constructed automatically by large language models and designed to support step-by-step validation. Our benchmark spans multiple domains, contains 1,305 data points, and has no overlap with existing mainstream benchmarks. Extensive experiments demonstrate that even the best large language models perform poorly on our dataset. For instance, GPT-5 attains merely 22.6% EM accuracy on the hardest portion of our dataset. Hop-aware diagnosis reveals that failures are primarily driven by distorted reasoning chains -- either collapsing prematurely or wandering into over-extension. This highlights a critical inability to allocate steps consistent with the task's logical structure, providing a diagnostic dimension missing in traditional evaluations. We believe our work will facilitate research in Agentic RAG and inspire further meaningful progress in this area. Our code and data are available at https://github.com/YqjMartin/AgenticRAGTracer.

Qijie You, Wenkai Yu, Wentao Zhang

University of Science and Technology Beijing · 2Peking University · 3Zhongguancun Academy +1

Multi-Hop QALLM Agent EvaluationAI Agent BenchmarksAgentic RAGAgentic ReasoningRAG Evaluation
AskPDF
Feb 22, 2026

Evaluating Test-Time Scaling of General LLM Agents

LLM agents are increasingly expected to operate as general-purpose systems that resolve real-world user requests, yet their dynamic scaling behavior in realistic environments remains poorly understood. In this paper, we systematically investigate two principal test-time scaling axes of LLM agents: sequential scaling through extended interaction and parallel scaling through trajectory sampling. We first introduce a realistic benchmark that provides one unified framework for evaluating LLM agents across search, coding, reasoning, and tool-use domains, more faithfully reflecting the heterogeneity of real-world deployments. Evaluating ten leading LLM agents reveals substantial performance degradation when transitioning from domain-specific evaluations to this realistic setting. Building on this foundation, we progressively scale test-time compute along fine-grained increments to characterize the performance upper bound. We find that neither scaling axis can consistently yield meaningful gains from additional test-time compute in realistic environments, a phenomenon we attribute to two fundamental limitations: the scaling plateau that bottlenecks sequential scaling and the verification gap that undermines parallel scaling. Code is publicly available at https://github.com/cxcscmu/General-AgentBench.

Xiaochuan Li, Ryan Ming, Pranav Setlur +6

Language Technologies Institute, School of Computer Science, Carnegie Mellon University · Meta

LLM Agent EvaluationTest-Time ScalingAI Agent Benchmarks
AskPDF
Feb 20, 2026

WorkflowPerturb: Calibrated Stress Tests for Evaluating Multi-Agent Workflow Metrics

Multi-agent LLM systems that generate structured workflows from natural-language requests are now deployed in production across cloud automation, DevOps, and enterprise orchestration. Operating them exposes a recurring change-management problem. Routine updates, such as re-running an input, swapping the LLM, or refactoring an agent's prompt or orchestration code, often produce workflows that differ substantially from validated references. Engineers then lack a principled way to decide whether a change is safe to ship. Automatic workflow evaluation is the natural tool, but in practice metric scores are poorly calibrated, and a numeric change rarely communicates the severity of the underlying degradation. We introduce WorkflowPerturb, a controlled benchmark that applies realistic, graded perturbations to golden workflows. It contains 4,973 golden workflows and 44,757 perturbed variants across three perturbation types (Missing Steps, Compressed Steps, Description Changes) at severities of 10%, 30%, and 50%. We benchmark multiple metric families, analyzing their sensitivity and calibration using expected score trajectories and instance-level alert rates. Our results characterize systematic differences across families and support severity-aware interpretation of workflow evaluation scores in change-management settings. WorkflowPerturb is publicly available at https://huggingface.co/datasets/microsoft/WorkflowPerturb .

Madhav Kanda, Sharad Agarwal, Rodrigo Fonseca +2

University of Illinois Urbana-Champaign · Microsoft

LLM Agent EvaluationAI Agent BenchmarksWorkflow Automation
AskPDF
Feb 19, 2026

Evaluating Chain-of-Thought Reasoning through Reusability and Verifiability

In multi-agent IR pipelines for tasks such as search and ranking, LLM-based agents exchange intermediate reasoning in terms of Chain-of-Thought (CoT) with each other. Current CoT evaluation narrowly focuses on target task accuracy. However, this metric fails to assess the quality or utility of the reasoning process itself. To address this limitation, we introduce two novel measures: reusability and verifiability. We decouple CoT generation from execution using a Thinker-Executor framework. Reusability measures how easily an Executor can reuse the Thinker's CoT. Verifiability measures how frequently an Executor can match the Thinker's answer using the CoT. We evaluated four Thinker models against a committee of ten Executor models across five benchmarks. Our results reveal that reusability and verifiability do not correlate with standard accuracy, exposing a blind spot in current accuracy-based leaderboards for reasoning capability. Surprisingly, we find that CoTs from specialized reasoning models are not consistently more reusable or verifiable than those from general-purpose LLMs like Llama and Gemma.

Shashank Aggarwal, Ram Vikas Mishra, Amit Awekar

Indian Institute of Technology Guwahati, Assam, India

CoT EvaluationLLM ReasoningLLM Agent Evaluation
AskPDF