Reinforcement learning environments are now a primary lever for improving large language model (LLM) capabilities in post-training, yet most agentic benchmarks remain static: the world moves only when the agent acts, the reward is a terminal verdict, and the pass bar is set arbitrarily. We introduce StoreBench, a live-commerce environment in which an agent runs a mid-size online apparel store on a production-grade commerce backend, testing long-horizon planning and economic judgment under uncertainty. Customers order around the clock, suppliers reprice and fail, and market shocks arrive with partial or no warning. The agent acts through the same 29 merchant tools a human operator would use, under a windowed operation budget that makes simulated time a function of actions taken, so model latency cannot influence simulated time. Pass thresholds are calibrated against scripted anchor policies, the reward is hardened against a catalogue of reward hacks, and every episode replays identically given a sequence of actions. We evaluate seven frontier LLMs on 11 scenarios of 30 to 45 days and a full simulated year, over three world seeds at matched reasoning effort. No model matches the scripted smart-triage policy on average: the best, DeepSeek-V4-Pro, passes 49% of task-seed cells against the heuristic's 97%. Human experts working through the same tools and budgets outscore every model (mean composite 0.708 vs. 0.700). Over a full simulated year under the Claude Code harness, most models show dramatic performance improvement. In a GRPO post-training run, Qwen3.5-27B trained on only five disjoint tasks raises its mean composite on the held-out evaluation tasks from 0.136 to 0.373. We release five example training-split tasks, ten sample trajectories, and the scoring and verification tooling; the full environment and evaluation suite are withheld to keep the benchmark uncontaminated.
Preference modeling aims to capture heterogeneous human values and increasingly relies on learning latent components to represent this variation without fine-grained supervision. What these components actually represent, however, is rarely examined. We audit Mixture-of-Experts (MoE) reward models along three rungs---semantic coherence, functional correspondence, and composability---in controlled and real-world settings, holding the architecture and training data fixed so that differences isolate the effect of the training objective. Our diagnosis is that the components conventional MoE reward models learn are entangled, which limits not only interpretability but also the personalization such models are built for. The entanglement is not intrinsic to the architecture but a property of the training objective: off-the-shelf sparsity and diversity regularizers already substantially reduce it. The resulting configuration, which we call sparse MoE, recovers all underlying semantic categories without domain supervision in a controlled setting, reaching 84.8% expert purity against 60.8% for conventional MoE; on real-world data its experts admit human-interpretable task- and domain-level descriptions aligned with their functional strengths. Targeted interventions on routing weights produce predictable preference changes, showing that these interpretations are behaviorally grounded rather than merely descriptive. Crucially, the disentanglement induced by this objective also makes the experts composable: lightweight router updates yield a 25.81-point improvement in few-shot personalization with 50 examples, roughly 2.7 times the gain of the identical architecture and adaptation procedure applied to entangled experts. Together, these results show that MoE reward models can learn interpretable expert structure, and that the same structure can substantially improve personalization.
Computer-use agents (CUAs) are commonly evaluated against fixed task specifications, but users may omit requirements, add constraints, or require approval before consequential actions. Completing the benchmark instruction does not guarantee serving the person who asked. We introduce PersonaCUABench, which holds 74 tasks across 11 domains fixed and varies only who asks. A persona agent, playing one of 44 personas in four groups, rewrites each instruction in that user's own words and answers the CUA's questions, while the CUA sees only the dialogue, never the profile or the original instruction. Across nine backbones (3,330 runs), mean task completion falls by 36% relative to the original requests. The losses have two sources: in 59.9% of cells that lose points, the agent engages with what the user asked for and still gets it wrong, and 11.2% are stops the user requested that the checker scores as failures. Even when a persona request is fully completed, the user gives the top satisfaction rating in only 62.7% of cells, against 84.1% for the original requests. CUAs should therefore be evaluated not only on task completion but also on whether they meet the requirements and constraints users express. Code: https://github.com/MatrAIx-ai/PersonaCUABench. Project page: https://matraix.ai/research/persona-cua.html