Large Language Model (LLM)-based search agents have demonstrated superior performance in knowledge-intensive tasks through interleaved reasoning and search interactions. Process Reward Models (PRMs) can enhance search agent performance by guiding step selection during inference. However, existing PRMs for search agents face a critical bottleneck: assessment unreliability in their training data. This issue stems from an over-reliance on single-source supervision, forcing a compromise between global solvability and local validity. We propose CoPRM (Collaborative Process Reward Model), a process reward model trained on high-quality preference data generated through our Collaborative Calibration methodology. This approach synthesizes immediate quality assessments with outcome-informed evaluations through a dual-strategy mechanism: strictly enforcing consensus filtering to eliminate unreliable signals, and employing an outcome-informed gating mechanism to resolve heterogeneous cross-type trade-offs. Trained on 10k preference pairs, CoPRM provides universal test-time guidance, enabling diverse search agents to achieve consistent performance improvements across four multi-hop QA benchmarks, demonstrating substantial gains, robust cross-domain generalization, and superior efficiency.

Recent reasoning-focused large language models (LLMs) (e.g., GPT-o1 and DeepSeek-R1) combine chain-of-thought prompting with reinforcement learning to generate explicit, multi-step derivations. However, their practical advantages over standard instruction-tuned models remain underexplored. In this article, we present a comprehensive empirical study to address these gaps. We compare reasoning-focused and non-reasoning LLMs across several critical dimensions: (1) their performance on mathematical and general reasoning benchmarks; (2) their generalization capabilities across 30 sub-disciplines spanning six scientific domains; (3) the qualitative and quantitative differences in their underlying solution processes; and (4) the impact of model compression via distillation on core reasoning capabilities. Crucially, to reliably conduct our in-depth process analysis, we address the critical issue of LLM-as-a-Judge reliability by proposing and validating a multi-level evaluation framework against human experts. Our experimental results demonstrate that reasoning-focused models significantly outperform non-reasoning counterparts on logic-intensive tasks such as mathematics and physics, while the latter remain competitive on factual or observational problems (e.g., biology, geography). Specifically, our validated process analysis reveals distinct styles: DeepSeek-R1 generates more elaborate reasoning processes, correlating with greater effectiveness on harder tasks, despite potential redundancy. In contrast, GPT-o1 exhibits greater conciseness and frequently presents solutions with superior clarity and efficiency, particularly on simpler problems, while struggling with complex scenarios. Additionally, reasoning capabilities can be partially preserved through distillation, with distilled reasoning models consistently outperforming their non-reasoning counterparts across all domains. These findings provide new insights into when and how explicit reasoning contributes to LLM performance, offering practical guidance for model development and deployment.

Fact verification is a critical primitive for ensuring the reliability of Large Language Models (LLMs). While frontier LLMs are effective verifiers, their repeated long-context inference incurs substantial costs; meanwhile, existing specialized verifiers face a fundamental dilemma: label-only models suffer from black-box opacity, while free-form rationale generators, whether trained by distillation or explanation-oriented RL objectives, can produce fluent but not structurally checkable explanations. In this work, we propose AtomicCheck a Protocol-Guided Policy Optimization framework that trains specialized verifiers to produce protocol-guided, locally checkable traces, moving beyond opaque label-only predictions and free-form rationales. Rather than cloning teacher traces or optimizing unstructured explanations, AtomicCheck uses RL to make the verification protocol itself part of the policy objective. To preserve semantic integrity during atomic decomposition, we introduce a Contrastive Decomposition Reward that compares decomposed reasoning with a same-backbone holistic counterfactual and penalizes decompositions that reduce verification accuracy. Extensive experiments demonstrate that AtomicCheck outperforms existing specialized verifiers and approaches strong frontier LLMs while producing structured traces that expose local checks behind the final decision.