Workflow-GYM: Towards Long-Horizon Evaluation of Computer-use Agentic tasks in Real-World Professional Fields

Benchmarks in AI typically have broader scientific impact as they establish standardized evaluation metrics that drive future research. Workflow-GYM addresses a critical gap in evaluating agents on long-horizon, professional GUI tasks, offering a platform that researchers across the field will likely use to test new models. While Paper 1 provides a strong architectural framework for enterprise security, Paper 2's benchmark will directly facilitate and measure the broader progression of agentic AI capabilities.

StatefulDiscovery addresses a fundamental challenge in AI-driven scientific discovery—evidence-calibrated claim formation—which has broad implications across all scientific fields. Its novel framework for coupling exploration with claim adjudication introduces a methodologically rigorous approach to open-ended discovery, a frontier problem with transformative potential. While Workflow-GYM is a useful benchmark contribution highlighting GUI agent limitations, benchmarks tend to have more incremental impact. StatefulDiscovery's framework for autonomous scientific reasoning is more novel, more broadly applicable, and addresses a more consequential problem for accelerating science.

Paper 2 introduces an automated system to generate dynamic, updatable benchmarks, addressing the fundamental issue of static benchmarks quickly saturating in AI. This methodological innovation has broader, longer-lasting impact across multiple embodied AI fields compared to Paper 1, which, while highly relevant, proposes a single static benchmark for GUI tasks.

Paper 2 introduces a much-needed benchmark for long-horizon, professional GUI tasks, addressing a critical bottleneck in the rapidly expanding field of autonomous AI agents. By revealing significant limitations in state-of-the-art models, it provides a foundational evaluation framework that will likely drive broad future research. Paper 1 offers a valuable but highly specific methodological refinement for LLM unlearning, which, while important for safety, has a narrower scope compared to the foundational impact of a novel, challenging benchmark in agentic AI.

Paper 1 likely has higher scientific impact because it introduces a broadly useful, timely benchmark for long-horizon GUI agent evaluation in real professional software—an area with wide applicability across AI agents, HCI, benchmarking, and enterprise automation. Its finding of low success rates creates a clear research gap and can become a standard evaluation tool, influencing many subsequent works. Paper 2 is methodologically rich and impactful for EDA/hardware design, but its domain is narrower; gains, while meaningful, may affect a smaller research and industry community than a cross-domain agent benchmark.

Workflow-GYM addresses a highly timely and economically valuable problem: long-horizon computer-use agents operating professional software. While Paper 1 offers a novel cognitive perspective on collaboration, Paper 2 aligns directly with the current AI frontier of developing autonomous GUI agents for real-world workflows. Its potential to drive immediate practical applications and highlight critical failure modes in state-of-the-art models gives it broader and more immediate scientific and industrial impact.

Paper 1 presents a novel and concrete methodological contribution—delayed per-step reward attribution with eligibility gating—that addresses a fundamental challenge in multi-agent RL for language models. Its strong empirical validation (an 8B model beating GPT-5 on a competitive benchmark, winning first place at NeurIPS 2025) demonstrates significant practical impact. Paper 2, while addressing an important gap in GUI agent evaluation, is primarily a benchmark contribution that reveals limitations without proposing solutions. Benchmark papers tend to have narrower long-term impact compared to papers introducing methods that achieve state-of-the-art results with dramatically fewer resources.

Paper 1 likely has higher impact because it introduces a new, timely benchmark targeting long-horizon, economically valuable, domain-specific GUI workflows—an evaluation gap with broad relevance to agent research, HCI, and real-world deployment. Benchmarks often become shared infrastructure, shaping research directions and enabling standardized comparisons across models and methods. While Paper 2’s uncertainty-aligned RL for tool-calling is innovative and useful, it is a more incremental algorithmic contribution whose impact depends on adoption and generalization, whereas Workflow-GYM can catalyze a wider ecosystem of methods and evaluations.

Paper 1 likely has higher impact due to stronger timeliness and broad applicability: long-horizon, real-world GUI agent evaluation is a key bottleneck for deploying agents in professional settings, and a well-designed benchmark can become a community standard. It targets economically valuable workflows across diverse domains and provides diagnostic failure modes that can steer multiple research areas (agent planning, UI grounding, robustness). Paper 2 is novel and useful, but idea-generation gains are incremental, evaluation is harder to validate, and downstream real-world adoption is less immediate than a benchmark for agent capability.

Paper 2 has higher potential impact due to a clear algorithmic contribution (residual-centric coding with deterministic guarantees) that materially advances high-fidelity scientific data compression, a critical bottleneck for HPC and climate/CFD workflows. It targets stringent error regimes (NRMSE 1e-6–1e-4), demonstrates sizable gains over strong baselines (GAE, SZ) across multiple real datasets, and emphasizes reproducible, deterministic decoding—important for scientific use. Paper 1 is timely and useful as a benchmark, but benchmark papers often yield narrower, more incremental impact unless they become a dominant standard.