LLM Agent Evaluation

LLM-based agents are increasingly applied to the "last mile" of Electronic Design Automation (EDA): repairing residual sign-off Design Rule Check (DRC) violations and converging Power-Performance-Area (PPA) targets after tool runs. Existing EDA-LLM benchmarks, however, omit DRC fixing entirely and rely on flat hierarchies tied to a single toolchain. We introduce PostEDA-Bench, a hierarchical benchmark with 145 tasks across DRC-Essential, DRC-Reasoning, PPA-Mono, and PPA-Multi, supported by EDA toolchains with machine-checkable evaluation. Across eight commercial and open-source LLMs under multiple agent scaffolds, we find that agents handle synthetic DRC-Essential and single-objective PPA-Mono reasonably well but degrade sharply on the more practical DRC-Reasoning, where the best success rate is 36.66%, and PPA-Multi, where the best success rate is 20.00%; vision augmentation consistently enhances DRC-Bench; and trade-off reasoning, rather than knob knowledge, is the dominant PPA-Multi bottleneck.

Large language models (LLMs) power deep research agents that synthesize information from hundreds of web sources into cited reports, yet these citations cannot be reliably verified. Current approaches either trust models to self-cite accurately, risking bias, or employ retrieval-augmented generation (RAG) that does not validate source accessibility, relevance, or factual consistency. We introduce the first source attribution evaluation framework that uses a reproducible AST parser to extract and evaluate inline citations from LLM-generated Markdown reports at scale. Unlike methods that verify claims in isolation, our framework closes the loop by retrieving the actual cited content, enabling human or model evaluators to judge each citation against its source. Citations are evaluated along three dimensions. (1) Link Works verifies URL accessibility, (2) Relevant Content measures topical alignment, and (3) Fact Check validates factual accuracy against source content. We benchmark 14 closed-source and open-source LLMs across three evaluation dimensions using rubric-based LLM-as-a-judge evaluators calibrated through human review. Our results reveal that even the strongest frontier models maintain link validity above 94% and relevance above 80%, yet achieve only 39-77% factual accuracy, while fewer than half of open-source models successfully generate cited reports in a one-shot setting. Ablation studies on research depth show that Fact Check accuracy drops by approximately 42% on average across two frontier models as tool calls scale from 2 to 150, demonstrating that more retrieval does not produce more accurate citations. These findings reveal a critical disconnect between surface-level citation quality and factual reliability, and our framework provides the evaluation infrastructure to assess the disconnect.

Large Language Model (LLM) agents are increasingly expected to maintain coherent, long-term personalized memory, yet current benchmarks primarily measure static fact retrieval, overlooking the ability to revise stored beliefs when new evidence emerges. We identify a critical and underexplored failure mode, Implicit Conflict: a later observation invalidates an earlier memory without explicit negation, requiring contextual inference and commonsense reasoning to detect. To rigorously evaluate this capability, we introduce STALE, a benchmark of 400 expert-validated conflict scenarios (1,200 evaluation queries across three probing dimensions) spanning over 100 everyday topics with contexts up to 150K tokens. We propose a three-dimensional probing framework that tests State Resolution (detecting that a prior belief is outdated), Premise Resistance (rejecting queries that falsely presuppose a stale state), and Implicit Policy Adaptation (proactively applying updated states in downstream behavior). A systematic evaluation of frontier LLMs and specialized memory frameworks reveals a pervasive gap between retrieving updated evidence and acting on it, with even the best evaluated model achieving only 55.2% overall accuracy. Models often accept outdated assumptions embedded in a user's query, and they struggle to recognize when a change in one aspect of the user's state should invalidate related memories. To establish an initial baseline for state-aware memory, we further present CUPMem, a prototype that strengthens write-time revision through structured state consolidation and propagation-aware search, suggesting that explicit state adjudication is a promising direction for robust agentic memory.

LLM-based multi-agent systems are increasingly deployed for payment workflows, yet prevailing metrics, Task Success Rate (TSR) and Agent Handoff F1-Score (HF1), capture only final outcomes or unordered routing decisions. We introduce the Agentic Success Rate (ASR), a trajectory-fidelity metric that compares observed and expected agent execution sequences at the transition level, decomposing performance into Transition Recall and Transition Precision. Applied to the Hierarchical Multi-Agent System for Payments (HMASP) across 18 LLMs and 90,000 task instances, ASR reveals that 10 of 18 models systematically skip a confirmation checkpoint during payment checkout, a deviation invisible to both TSR and HF1, while 8 models enforce the checkpoint perfectly. Notably, GPT-4.1 exhibits hidden workflow shortcuts despite achieving perfect TSR and HF1, while GPT-5.2 achieves perfect ASR. Prompt refinements and deterministic routing guards guided by ASR diagnostics yield substantial TSR improvements, with gains up to +93.8 percentage points for previously struggling models, demonstrating that trajectory-level evaluation is essential in regulated domains.

This position paper argues that, for long-horizon tasks evaluated across models with comparable frontier capability, the agent execution harness, namely the infrastructure layer that governs context construction, tool interaction, orchestration, and verification around a language model, is often a stronger determinant of agent performance than the model it wraps. We formalize and defend the Binding Constraint Thesis: in this regime, performance variance is governed more by harness configuration than by model choice, and current evaluation protocols therefore systematically misattribute harness-level gains to model improvements. We support this thesis along three lines. First, a control-theoretic formalization treats the harness as the controller of a closed-loop dynamical system and the LLM as the stochastic policy it governs, which explains why small harness changes can produce performance shifts that exceed those obtained by substituting one model for another. Second, published benchmarks, industry deployments, and a controlled variance decomposition show that harness-induced variance can substantially exceed model-induced variance, including cases of model ranking reversal. Third, we propose a harness-aware evaluation framework with a disclosure standard and a variance decomposition protocol. Until harness specifications are disclosed, leaderboard comparisons for long-horizon agents should be treated as incomplete and potentially misleading.

Most coding-agent benchmarks ask whether generated code behaves correctly. That remains essential, but repository-level engineering is increasingly agent-managed: one agent writes a repository, and later agents inspect, audit, or extend it as working context. In that setting, a generated repository is not only an answer to a task but also a communication artifact for future work. Even when strong agents nearly satisfy the visible behavioral objective, repositories can differ in how clearly they expose the intended behavior and design choices behind that behavior. We introduce BUILD-AND-FIND, a protocol for evaluating whether downstream agents can recover those intended choices from generated repositories, and how much inspection that recovery requires. For each task, a builder sees a hidden repository specification and creates a codebase; a finder sees only the codebase and a specification-traced multiple-choice question bank. The protocol separates behavioral correctness from artifact-side recovery and reports recovery accuracy, repeatability, implementation coverage, and inspection effort. Accuracy and stability act as gates: effort is interpreted only when recovery succeeds reliably. Among artifacts from which the same intent can be recovered, lower effort by the same finder suggests that the artifact makes that intent easier to locate. Question-only and spec-only controls quantify generic priors and specification access, while audits separate omitted claims from finder failures and check whether correct answers cite artifact evidence. In the released high-prior task pack, recovery accuracy is near saturation, so inspection effort and finder-specific effects provide the main panel-local comparison.

LLM agent systems are built by stacking scaffolding components (planning, tools, memory, self-reflection, retrieval) assuming more is better. We study cross-component interference (CCI): degradation when components interact destructively. We run a full factorial experiment over all 2^5=32 subsets of five components on HotpotQA and GSM8K with Llama-3.1-8B/70B (96 conditions, up to 10 seeds). The All-In system is consistently suboptimal: on HotpotQA, a single-tool agent surpasses All-In by 32% (F1 0.233 vs 0.177, p=0.023); on GSM8K, a 3-component subset beats All-In by 79% (0.43 vs 0.24, p=0.010). Optimal component count is task-dependent (k*=1-4) and scale-sensitive: at 70B, combinations that hurt at 8B provide gains, though All-In still trails the best subset. We fit a main-effects regression (R^2=0.916, adj-R^2=0.899, LOOCV=0.872), compute exact Shapley values, and find 183/325 submodularity violations (56.3%), showing greedy selection is unreliable. A three-body synergy among Tool Use, Self-Reflection, and Retrieval (INT_3=+0.175, 95% CI [+0.003,+0.351]) is reported as exploratory. CCI replicates across model families (Qwen2.5) and is robust to prompt paraphrasing. Our findings suggest maximally-equipped agent defaults should be replaced by task-specific subset selection via interaction-aware analysis.

As Large Language Models (LLMs) evolve into interactive agents, understanding their behavioral alignment within human social dynamics becomes essential. While behavioral game theory offers a framework to study these interactions, previous work has predominantly relied on outcome-based metrics such as average scores. This focus overlooks the mechanisms that facilitate sustainable cooperation, as identical scores can be derived from vastly different strategies. To bridge this gap, we introduce SODE (Social Dynamics Evaluation), a framework that evaluates LLM agents across three evolutionary dimensions: Direct Reciprocity for strategy adaptation, Indirect Reciprocity for reputation sensitivity, and Group Dynamics for cooperative resilience. Applying SODE reveals systematic divergences: instruction-tuned models often exhibit "passive compliance" that renders them vulnerable to exploitation, while reasoning models prioritize short-horizon optimization, destabilizing long-term cooperation. Notably, we demonstrate that a "long-horizon framing" can unlock reciprocal capabilities in reasoning models. Thus, SODE offers a systematic, mechanism-grounded benchmark for aligning AI agents with complex human social dynamics.

Static capabilities benchmarks suffer from saturation and contamination, making it difficult to track capabilities progress over time. We introduce Agent Island, a multiplayer simulation environment in which language-model agents compete in a game of interagent cooperation, conflict, and persuasion. The environment yields a dynamic benchmark designed to mitigate both saturation and contamination; new models can always outperform the current leading player in this winner-take-all game, and agents compete against other adaptive agents rather than face a fixed task set. We rank players with a Bayesian Plackett-Luce model, allowing us to quantify uncertainty in player skill. In 999 games involving 49 unique models, openai/gpt-5.5 dominates its peers with a posterior mean skill of 5.64, compared with 3.10 for the second-ranked model, openai/gpt-5.2, and 2.86 for the third-ranked model, openai/gpt-5.3-codex. We release the game logs as a dataset for analyses of model behavior. As an example, we investigate same-provider preference in final-round votes and find that models are 8.3 p.p. more likely to support a same-provider finalist than finalists from other providers. This preference is not uniform across providers: among separately estimated providers, the effect is strongest for OpenAI models and weakest for Anthropic models.

Language models excel at diagnostic assessments on curated medical case-studies and vignettes, performing on par with, or better than, clinical professionals. However, existing studies focus on complex scenarios with rich context making it difficult to draw conclusions about how these systems perform for patients reporting symptoms in everyday life. We deployed SymptomAI, a set of conversational AI agents for end-to-end patient interviewing and differential diagnosis (DDx), via the Fitbit app in a study that randomized participants (N=13,917) to interact with five AI agents. This corpus captures diverse communication and a realistic distribution of illnesses from a real world population. A subset of 1,228 participants reported a clinician-provided diagnosis, and 517 of these were further evaluated by a panel of clinicians during over 250 hours of annotation. SymptomAI DDx were significantly more accurate (OR = 2.56, p < 0.001) than those from independent clinicians given the same dialogue in a blinded randomized comparison. Moreover, agentic strategies which conduct a dedicated symptom interview that elicit additional symptom information before providing a diagnosis, perform substantially better than baseline, user-guided conversations (p < 0.001). An auxiliary analysis on 1,509 conversations from a general US population panel validated that these results generalize beyond wearable device users. We used SymptomAI diagnoses as labels for all 13,917 participants to analyze over 500,000 days of wearable metrics across nearly 400 unique conditions. We identified strong associations between acute infections and physiological shifts (e.g., OR > 7 for influenza). While limited by self-reported ground truth, these results demonstrate the benefits of a dedicated and complete symptom interview compared to a user-guided symptom discussion, which is the default of most consumer LLMs.

Human-AI collaboration requires AI agents to understand human behavior for effective coordination. While advances in foundation models show promising capabilities in understanding and showing human-like behavior, their application in embodied collaborative settings needs further investigation. This work examines whether embodied foundation model agents exhibit emergent collaborative behaviors indicating underlying mental models of their collaborators, which is an important aspect of effective coordination. This paper develops a 2D collaborative game environment where large language model agents and humans complete color-matching tasks requiring coordination. We define five collaborative behaviors as indicators of emergent mental model representation: perspective-taking, collaborator-aware planning, introspection, theory of mind, and clarification. An automated behavior detection system using LLM-based judges identifies these behaviors, achieving fair to substantial agreement with human annotations. Results from the automated behavior detection system show that foundation models consistently exhibit emergent collaborative behaviors without being explicitly trained to do so. These behaviors occur at varying frequencies during collaboration stages, with distinct patterns across different LLMs. A user study was also conducted to evaluate human satisfaction and perceived collaboration effectiveness, with the results indicating positive collaboration experiences. Participants appreciated the agents' task focus, plan verbalization, and initiative, while suggesting improvements in response times and human-like interactions. This work provides an experimental framework for human-AI collaboration, empirical evidence of collaborative behaviors in embodied LLM agents, a validated behavioral analysis methodology, and an assessment of collaboration effectiveness.

As LLM-based agents are increasingly interacting in multi-party settings, they need to properly handle information asymmetry, i.e., knowing when and to whom to disclose information is appropriate. Yet, existing benchmarks fail to measure this ability in realistic multi-party settings. Thus, we introduce SOTOPIA-TOM, a multi-dimensional benchmarking framework to evaluate LLM agents' ability to successfully navigate information asymmetric and privacy sensitive multi-party interactions. We create an interaction environment which enables both public (broadcast) and private (direct message) communication, and craft 160 human-reviewed scenarios across eight industry sectors, each involving 3 to 5 agents with partitioned private knowledge and channel-dependent sharing policies. To measure interaction abilities, we create a multi-dimensional evaluation framework to assess how well agents share useful information, seek missing details, coordinate efficiently, and protect privacy, which we also combine into a composite INFOMGMT metric. Results show that, across 6 LLM backbones and prompting strategies (vanilla, CoT-privacy, and ToM-based interventions), even the largest high-reasoning model (GPT-5) reaches only a 62% INFOMGMT score, which indicates persistent deficiencies in information seeking and privacy-aware decision-making. Additionally, ToM-based interventions more consistently improve the overall coordination-privacy balance (for example, relative to the vanilla baseline, ToM-Coach reduces critical privacy violations on GPT-4o from 9.9% to 2.2% while increasing the composite InfoMgmt score more than 2.5x from 15% to 40%). Overall, SOTOPIA-TOM exposes persistent limitations of current LLM agents in complex, information-asymmetric coordination and provides an extensible testbed for developing more privacy-aware, theory-of-mind capable multi-agent systems.

Long-term LLM agents must compress streams of past interactions into persistent memory before future queries are known. Existing evaluations usually measure final question-answering accuracy, which entangles memory writing with retrieval, prompting, and reader reasoning. We introduce MEMAUDIT, an exact packageoracle evaluation protocol for budgeted long-term memory writing. A MEMAUDIT package fixes an experience stream, candidate memory representations, storage costs, semantic evidence units, future-query requirements, and a budget, turning write-time memory selection into a finite auditable optimization problem with a certified denominator. We instantiate this protocol with a concave-over-modular semantic coverage objective under storage and one-representation-per-experience constraints, and compute exact package optima using branch-and-bound with MILP certification. Across controlled exact packages, validity-heavy stress tests, human-audited natural support slices, and exported Mem0, A-Mem, and Letta stores, MEMAUDIT separates representation quality, validity-state preservation, and budget-aware selection effects that end-to-end QA cannot localize. The resulting artifact provides reusable package generators, certified solvers, natural package exports, external-system scorers, and cached reproducibility metadata for evaluating what memory writers actually preserve under fixed storage budgets.

What if the twelve jurors of Sidney Lumet's 12 Angry Men (1957) were not men, but large language models? Would the one juror who disagrees still be able to change everyone's mind? This paper instantiates that scenario as a multi-agent benchmark for LLM deliberation: twelve agents, each conditioned on a film-faithful persona, debate the film's murder case using multi-agent framework. Two models representing opposite ends of the RLHF spectrum are tested: GPT-4o (closed-source, heavy alignment) and Llama-4-Scout (open-weight, lighter alignment), across three conditions (baseline, open-minded prompt, no initial vote), with N = 3 replications per cell (18 runs total). Three findings emerge. (i) Seventeen of eighteen runs end in a hung jury (a state where the jury fails to reach a unanimous verdict); the film's central event, gradual minority-to-majority persuasion, almost never occurs, indicating that anchoring is the dominant failure mode of current LLMs in this setting. (ii) The two models exhibit sharply different internal dynamics: GPT-4o produces a mean of 1.0 vote changes per run across all conditions, while Llama-4-Scout ranges from 2.0 (baseline) to 6.0 (open-minded prompt), and is the only model to reach a NOT_GUILTY verdict (1 of 3 runs in the no-initial-vote condition). The same ``open-minded'' instruction is internalized by Llama and ignored by GPT-4o. (iii) This asymmetry suggests that the intensity of RLHF alignment training, not model capability, is the primary determinant of deliberative flexibility in multi-agent settings. Flexibility, not capability, tracks human deliberation. The work is framed as an exploratory study and discusses implications for jury-of-LLMs evaluation and multi-agent debate.

Outcome-only evaluation under-specifies whether an evaluated agent profile preserves the commitments required to solve a multi-turn task coherently. NeuroState-Bench is a human-calibrated benchmark that operationalizes commitment integrity through benchmark-defined side-query probes rather than inferred hidden activations. The released inventory contains 144 deterministic tasks and 306 benchmark-defined side-query probes spanning eight cognitively motivated failure families, paired clean and distractor variants, and three difficulty bands. The main 32-profile evaluation contains a fixed 16-profile local subset and a matched 16-profile hosted large-model subset evaluated through the same benchmark pipeline. Human calibration uses the final merged reporting scope: 104 sampled task units, 216 raw annotations, and 108 adjudicated task rows, with weighted kappa = 0.977 and ICC(2,1) = 0.977. Empirically, task success and commitment integrity diverge across this expanded grid: the success leader is not the integrity leader, 31 of 32 profiles change rank when integrity replaces task success, and integrity rankings are more stable under distractor perturbation. The primary confidence-free score HCCIS-CORE reaches 0.8469 AUC and 0.6992 PR-AUC for post-probe diagnostic discrimination of terminal task failure; the legacy full heuristic variant HCCIS-FULL reaches 0.7997 AUC and 0.6410 PR-AUC. Probe accuracy and state drift achieve slightly higher ROC-AUC, 0.8587, and better Brier/ECE, while HCCIS-CORE has substantially higher point-estimate PR-AUC and remains more closely tied to the benchmark's intended construct. The exploratory neural-augmented variant HCCIS+N is weaker overall, and a randomized subspace control approaches chance. NeuroState-Bench therefore contributes a calibrated evaluation axis for exposing commitment failures over a broader model grid than the original local-only subset.

Personas are often employed to guide large language model agents, yet their effectiveness in shaping strategic behavior in social dilemma settings remains uncertain. To address this, we examined the impact of persona prompts in an iterated Split or Steal game where persona-driven agents interacted with a Virtual Human (VH) controlled by a fixed prompt. Agents were instantiated from four open models (Ministral 3:3b, phi4:14b, Gemma3:12b, and Gemma4:e4b) at two temperature settings (0.3 and 0.7) and deterministic decision with zero temperature, while the VH was powered by GPT 4.1 mini. Across 160 sessions of 15 rounds each conducted in European Portuguese, mutual Split outcomes dominated (roughly 74 percent of rounds), with exploitation occurring in fewer than 11 percent of rounds. Model choice significantly influenced behavior: phi4 and Ministral 3:3b remained consistently cooperative across temperatures, whereas Gemma3:12b and Gemma4:e4b exhibited more varied strategies and outcomes. Analyses based on Big Five personality traits indicated that Prosocial and Principled personas were most consistently cooperative, while Analytical personas were more likely to exploit the VH. Topic analysis revealed that friendship-related dialogue aligns with Split decisions, whereas money and vengeance-related content is more prevalent in Steal outcomes; sentiment labels were predominantly neutral or happy and provided limited additional explanatory value. These findings characterize the interaction between persona prompts and model differences in repeated trust games and serve as a baseline for planned virtual reality studies involving human participants interacting with an embodied VH.

Reinforcement learning (RL) trained language model agents with tool access are increasingly deployed in coding assistants, research tools, and autonomous systems. We introduce the Reward Hacking Benchmark (RHB), a suite of multi-step tasks requiring sequential tool operations with naturalistic shortcut opportunities such as skipping verification steps, inferring answers from task-adjacent metadata, or tampering with evaluation-relevant functions. RHB supports independent and chained task regimes, where chain length acts as a proxy for longer-horizon agent behavior. We evaluate 13 frontier models from OpenAI, Anthropic, Google, and DeepSeek. Exploit rates range from 0% (Claude Sonnet 4.5) to 13.9% (DeepSeek-R1-Zero), varying sharply by post-training style. A controlled sibling comparison (DeepSeek-V3 vs. DeepSeek-R1-Zero) shows RL post-training is associated with substantially higher reward hacking (0.6% vs. 13.9%), with consistent gaps across all four task families. We identify six exploit categories and find that 72% of reward hacking episodes include explicit chain-of-thought rationale, suggesting models often frame exploits as legitimate problem-solving. Simple environmental hardening reduces exploit rates by 5.7 percentage points (87.7% relative) without degrading task success. Models with near-zero exploit rates on standard tasks show elevated rates on harder variants, suggesting that production-aligned post-training appears to suppress reward hacking only below a complexity threshold where honest solutions remain tractable.

Formal specification is essential for rigorous program verification, yet writing correct specifications remains costly and difficult to automate. Although large language models (LLMs) and agents have shown promising progress, their true capabilities and failure modes remain unclear. We present the first systematic and contamination-aware study of LLM- and agent-based formal specification generation for C programs. We introduce LiveFMBench, a continuously evolving benchmark of 630 ACSL (ANSI/ISO C Specification Language)-annotated C programs, including 360 newly collected cases designed to mitigate data leakage. Using this benchmark, we evaluate direct prompting with different sampling sizes, reasoning-enabled (thinking mode) inference, the agentic pipeline, and perform a fine-grained failure analysis. Experimental results reveal that naive evaluation substantially overestimates performance because models under direct prompting may exhibit unfaithful behaviors, such as deceiving automated provers or ignoring code-context constraints; after excluding such cases, the true specification generation accuracy drops by approximately 20%. We further find that both increased sampling and thinking mode significantly improve success rates, with smaller models benefiting more from thinking mode. Agentic pipelines are particularly effective under low sampling budgets and on harder datasets. Failure analysis further shows that incorrect loop invariants are the dominant error type, while agentic pipelines notably reduce assertion errors. These results expose fundamental limitations in current LLM-based approaches and suggest they remain far from replacing human-authored formal specifications. We release LiveFMBench at https://huggingface.co/datasets/fm-universe/Live-FM-Bench and all evaluation artifacts to support future research.

This position paper argues that agentic AI systems should be designed and evaluated as marginal token allocation economies rather than as text generators priced by the unit. We follow a single request -- a developer asking a coding agent to fix a failing test -- through four economic layers that today are designed in isolation: a router that decides which model answers, an agent that decides whether to plan, act, verify, or defer, a serving stack that decides how to produce each token, and a training pipeline that decides whether the trace is worth learning from. We show that all four layers are solving the same first-order condition -- marginal benefit equals marginal cost plus latency cost plus risk cost -- with different index sets and different prices. The framing is deliberately minimal: we do not propose a complete theory of AI economics. But adopting marginal token allocation as the shared accounting object explains why systems that locally minimize tokens globally misallocate them, predicts a small set of recurring failure modes (over-routing, over-delegation, under-verification, serving congestion, stale rollouts, cache misuse), and points to a concrete research agenda in token-aware evaluation, autonomy pricing, congestion-priced serving, and risk-adjusted RL budgeting.

Production agentic systems make many model calls per user request, and most of those calls are short, structured, and routine. This raises a practical routing question that existing evaluations do not directly answer: which parts of an agent workflow truly require large frontier intelligence, and which can be handled by smaller models? We introduce AgentFloor, a deterministic 30-task benchmark organized as a six-tier capability ladder, spanning instruction following, tool use, multi-step coordination, and long-horizon planning under persistent constraints. We evaluate 16 open-weight models, from 0.27B to 32B parameters, alongside GPT-5 across 16,542 scored runs. Our results reveal a clear boundary of model necessity. Small and mid-sized open-weight models are already sufficient for much of the short-horizon, structured tool use work that dominates real agent pipelines, and in aggregate, the strongest open-weight model matches GPT-5 on our benchmark while being substantially cheaper and faster to run. The gap appears most clearly on long-horizon planning tasks that require sustained coordination and reliable constraint tracking over many steps, where frontier models still hold an advantage, though neither side reaches strong reliability. We also find that this boundary is not explained by scale alone: some failures respond to targeted interventions, but the effects are model-specific rather than universal. These findings suggest a practical design principle for agentic systems: use smaller open-weight models for the broad base of routine actions, and reserve large frontier models for the narrower class of tasks that truly demand deeper planning and control. We release the benchmark, harness, sweep configurations, and full run corpus.

LLM agents are expected to complete end-to-end units of work across software tools, business services, and local workspaces. Yet many agent benchmarks freeze a curated task set at release time and grade mainly the final response, making it difficult to evaluate agents against evolving workflow demand or verify whether a task was executed. We introduce Claw-Eval-Live, a live benchmark for workflow agents that separates a refreshable signal layer, updated across releases from public workflow-demand signals, from a reproducible, time-stamped release snapshot. Each release is constructed from public workflow-demand signals, with ClawHub Top-500 skills used in the current release, and materialized as controlled tasks with fixed fixtures, services, workspaces, and graders. For grading, Claw-Eval-Live records execution traces, audit logs, service state, and post-run workspace artifacts, using deterministic checks when evidence is sufficient and structured LLM judging only for semantic dimensions. The release contains 105 tasks spanning controlled business services and local workspace repair, and evaluates 13 frontier models under a shared public pass rule. Experiments reveal that reliable workflow automation remains far from solved: the leading model passes only 66.7% of tasks and no model reaches 70%. Failures are structured by task family and execution surface, with HR, management, and multi-system business workflows as persistent bottlenecks and local workspace repair comparatively easier but unsaturated. Leaderboard rank alone is insufficient because models with similar pass rates can diverge in overall completion, and task-level discrimination concentrates in a middle band of tasks. Claw-Eval-Live suggests that workflow-agent evaluation should be grounded twice, in fresh external demand and in verifiable agent action.

Large Language Models (LLMs) are increasingly used as proxies for human perception in urban analysis, yet it remains unclear whether persona prompting produces meaningful and reproducible behavioral diversity. We investigate whether distinct personas influence urban sentiment judgments generated by multimodal LLMs. Using a factorial set of personas spanning gender, economic status, political orientation, and personality, we instantiate multiple agents per persona to evaluate urban scene images from the PerceptSent dataset and assess both within-persona consistency and cross-persona variation. Results show strong convergence among agents sharing a persona, indicating stable and reproducible behavior. However, cross-persona differentiation is limited: economic status and personality induce statistically detectable but practically modest variation, while gender shows no measurable effect and political orientation only negligible impact. Agents also exhibit an extremity bias, collapsing intermediate sentiment categories common in human annotations. As a result, performance remains strong on coarse-grained polarity tasks but degrades as sentiment resolution increases, suggesting that simple label-based persona prompting does not capture fine-grained perceptual judgments. To isolate the contribution of persona conditioning, we additionally evaluate the same model without personas. Surprisingly, the no-persona model sometimes matches or exceeds persona-conditioned agreement with human labels across all task variants, suggesting that simple label-based persona prompting may add limited annotation value in this setting.

This paper examines how large language model (LLM) agents perform on scientific visualization (SciVis) tasks that require generating visualization workflows from natural-language instructions. We compare three representative agent designs: domain-specific agents with structured tool use, computer-use agents, and general-purpose coding agents, across 15 benchmark tasks, evaluating visualization quality, efficiency, robustness, computational cost, and the impact of persistent memory. We further study interaction modalities, including code scripts, model context protocol (MCP) or API calls, command-line interfaces (CLI), and graphical user interfaces (GUI). Our goal is to characterize the tradeoffs among representative SciVis agent configurations used in practice. The results reveal clear tradeoffs across agent designs and interaction modalities. General-purpose coding agents achieve the highest task success rates but incur greater computational cost, whereas domain-specific agents are more efficient and stable but less flexible. Computer-use agents perform well on individual operations but struggle with multi-step workflows. Across both CLI- and GUI-based settings, persistent memory improves performance over repeated trials, but its effectiveness depends on the interaction mode and the quality of feedback. These findings suggest that future SciVis systems should combine structured tool use, interactive capabilities, and adaptive memory mechanisms to balance performance, robustness, and flexibility.

With the advancement of multimodal large language models (MLLMs) and coding agents, the website development has shifted from manual programming to agent-based project-level code synthesis. Existing benchmarks rely on idealized assumptions, especially for well-structured, information-rich inputs and static execution settings. In contrast, real-world development is constrained by a critical bottleneck: the semantic misalignment between ambiguous, low-quality instructions from non-expert users and model understanding, which results in a failure mode that we term blind execution. To address this gap, we introduce InteractWeb-Bench, the first multimodal interactive benchmark for website generation under non-expert low-code user conditions. InteractWeb-Bench introduces four types of user agents and persona-driven instruction perturbations to systematically simulate diverse user behaviors, including ambiguity, redundancy, and contradiction, grounded in requirement engineering defect taxonomies. We develop an interactive execution environment for agents, featuring a unified action space comprising Clarify, Implement, Verify, and Submit, enabling iterative intent refinement, code synthesis, and visual feedback-based validation. Extensive experiments and analysis reveal that frontier MLLM-based agents remain trapped in blind execution, exposing limitations in intent recognition and adaptive interaction.

Democratic discourse analysis systems increasingly rely on multi-agent LLM pipelines in which distinct evaluator models are assigned adversarial roles to generate structured, multi-perspective assessments of political statements. A core assumption is that models will reliably maintain their assigned roles. This paper provides the first systematic empirical test of that assumption using the TRUST pipeline. We develop an epistemic stance classifier that identifies advocate roles from reasoning text without relying on surface vocabulary, and measure role fidelity across 60 political statements (30 English, 30 German) using four metrics: Role Drift Index (RDI), Expected Drift Distance (EDD), Directional Drift Index (DDI), and Entropy-based Role Stability (ERS). We identify two failure modes - the Epistemic Floor Effect (fact-check results create an absolute lower bound below which the legitimizing role cannot be maintained) and Role-Prior Conflict (training-time knowledge overrides role instructions for factually unambiguous statements) - as manifestations of a single mechanism: Epistemic Role Override (ERO). Model choice significantly affects role fidelity: Mistral Large outperforms Claude Sonnet by 28pp (67% vs. 39%) and exhibits a qualitatively different failure mode - role abandonment without polarity reversal - compared to Claude's active switch to the opposing stance. Role fidelity is language-robust. Fact-check provider choice is not universally neutral: Perplexity significantly reduces Claude's role fidelity on German statements (Delta = -15pp, p = 0.007) while leaving Mistral unaffected. These findings have direct implications for multi-agent LLM validation: a system validated without role fidelity measurement may systematically misrepresent the epistemic diversity it was designed to provide.

Forecasting benchmarks produce accuracy leaderboards but little insight into why some forecasters are more accurate than others. We introduce Bench to the Future 2 (BTF-2), 1,417 pastcasting questions with a frozen 15M-document research corpus in which agents reproducibly research and forecast offline, producing full reasoning traces. BTF-2 detects accuracy differences of 0.004 Brier score, and can distinguish differential agent strengths in research vs. judgment. We build a forecaster 0.011 Brier more accurate than any single frontier agent, and use it to evaluate agent strategic reasoning without hindsight bias. We find the better forecaster differs primarily in its pre-mortem analysis of its blind spots and consideration of black swans. Expert human forecasters found the dominant strategic reasoning failures of frontier agents are in assessing political and business leaders' incentives, judging their likelihood to follow through on stated plans, and modeling institutional processes.

We study reliability in autonomous language-model agents that translate user mandates into validated tool actions under real capital. The setting is DX Terminal Pro, a 21-day deployment in which 3,505 user-funded agents traded real ETH in a bounded onchain market. Users configured vaults through structured controls and natural-language strategies, but only agents could choose normal buy/sell trades. The system produced 7.5M agent invocations, roughly 300K onchain actions, about $20M in volume, more than 5,000 ETH deployed, roughly 70B inference tokens, and 99.9% settlement success for policy-valid submitted transactions. Long-running agents accumulated thousands of sequential decisions, including 6,000+ prompt-state-action cycles for continuously active agents, yielding a large-scale trace from user mandate to rendered prompt, reasoning, validation, portfolio state, and settlement. Reliability did not come from the base model alone; it emerged from the operating layer around the model: prompt compilation, typed controls, policy validation, execution guards, memory design, and trace-level observability. Pre-launch testing exposed failures that text-only benchmarks rarely measure, including fabricated trading rules, fee paralysis, numeric anchoring, cadence trading, and misread tokenomics. Targeted harness changes reduced fabricated sell rules from 57% to 3%, reduced fee-led observations from 32.5% to below 10%, and increased capital deployment from 42.9% to 78.0% in an affected test population. We show that capital-managing agents should be evaluated across the full path from user mandate to prompt, validated action, and settlement.

Given the increased use of LLMs in financial systems today, it becomes important to evaluate the safety and robustness of such systems. One failure mode that LLMs frequently display in general domain settings is that of sycophancy. That is, models prioritize agreement with expressed user beliefs over correctness, leading to decreased accuracy and trust. In this work, we focus on evaluating sycophancy that LLMs display in agentic financial tasks. Our findings are three-fold: first, we find the models show only low to modest drops in performance in the face of user rebuttals or contradictions to the reference answer, which distinguishes sycophancy that models display in financial agentic settings from findings in prior work. Second, we introduce a suite of tasks to test for sycophancy by user preference information that contradicts the reference answer and find that most models fail in the presence of such inputs. Lastly, we benchmark different modes of recovery such as input filtering with a pretrained LLM.

Multiple myeloma is managed through sequential lines of therapy over years to decades, with each decision depending on cumulative disease history distributed across dozens to hundreds of heterogeneous clinical documents. Whether LLM-based systems can synthesise this evidence at a level approaching expert agreement has not been established. A retrospective evaluation was conducted on longitudinal clinical records of 811 myeloma patients treated at a tertiary centre (2001-2026), covering 44,962 documents and 1,334,677 laboratory values, with external validation on MIMIC-IV. An agentic reasoning system was compared against single-pass retrieval-augmented generation (RAG), iterative RAG, and full-context input on 469 patient-question pairs from 48 templates at three complexity levels. Reference labels came from double annotation by four oncologists with senior haematologist adjudication. Iterative RAG and full-context input converged on a shared ceiling (75.4% vs 75.8%, p = 1.00). The agentic system reached 79.6% concordance (95% CI 76.4-82.8), exceeding both baselines (+3.8 and +4.2 pp; p = 0.006 and 0.007). Gains rose with question complexity, reaching +9.4 pp on criteria-based synthesis (p = 0.032), and with record length, reaching +13.5 pp in the top decile (n = 10). The system error rate (12.2%) was comparable to expert disagreement (13.6%), but severity was inverted: 57.8% of system errors were clinically significant versus 18.8% of expert disagreements. Agentic reasoning was the only approach to exceed the shared ceiling, with gains concentrated on the most complex questions and longest records. The greater clinical consequence of residual system errors indicates that prospective evaluation in routine care is required before these findings translate into patient benefit.

Markets are a promising way to coordinate AI agent activity for similar reasons to those used to justify markets more broadly. In order to effectively participate in markets, agents need to have informative signals of their own ability to successfully complete a task and the cost of doing so. We propose MarketBench, a benchmark for assessing whether AI agents have these capabilities. We use a 93-task subset of SWE-bench Lite, a software engineering benchmark, with six recently released LLMs as a demonstration. These LLMs are miscalibrated on both success probability and token usage, and auctions built from these self-reports diverge from a full-information allocation. A follow-up intervention where we add information about capabilities from prior experiments to the context improves calibration, but only modestly narrows the gap to a full-information benchmark. We also document the performance of a market-based scaffolding with these LLMs. Our results point to self-assessment as a key bottleneck for market-style coordination of AI agents.