Group-relative policy optimization (GRPO) provides a stable recipe for outcome-based RL on language model reasoning, but its credit assignment is coarse: it broadcasts a single trajectory advantage to every token, regardless of how individual steps contribute to the outcome. We call this the uniform-broadcast failure mode and address it with RLPgR, a dependency-aware rescaling of the existing GRPO advantage. RLPgR parses each rollout into a Progress Dependency Graph (PDG), where each node groups a contiguous span of steps into a milestone-level unit with a locally validatable artifact, and directed edges connect a node to the artifacts it depends on. Using the PDG, RLPgR rescales the group-relative advantage by a small, bounded, sign-preserving factor at the node level. This preserves the outcome-governed rollout direction while each step’s update better reflects its dependency role, avoiding the reward-hacking drift of densely shaped process rewards. To reduce PDG construction cost, we train an open Qwen3-8B progress annotator amortizing a one-time proprietary bootstrap. Across model families and sizes on five math reasoning benchmarks, RLPgR is competitive with strong outcome-RL and process-RL baselines under matched training budget. Three controlled studies (scaling-strength sensitivity, segment-level baselines, and PDG label perturbation) support the view that the gain comes from structured credit assignment rather than incidental tuning.

Verifiers have been demonstrated to enhance LLM reasoning via test-time scaling (TTS). Yet, they face significant challenges in complex domains. Error propagation from incorrect intermediate reasoning can lead to false positives for seemingly plausible solutions, while lacking external grounding makes verifiers unreliable on computation or knowledge-intensive tasks. To address these challenges, we propose Agentic Verifier, a framework that transforms reward modeling into a multi-turn, tool-augmented deliberative process. We introduce complementary forward and backward agents: one traces solutions from premises to conclusions, while the other re-checks conclusions against their underlying premises. This bidirectional process enables a comprehensive, reliable, and interpretable assessment of solutions. To facilitate practical deployment, we propose AgentV-RL. Through proactive exploration and reinforcement learning, the verifier autonomously interleaves tool-use with internal reasoning. Extensive experiments show that Agentic Verifier yields consistent performance gains under both parallel and sequential TTS. Notably, our 4B variant surpasses state-of-the-art ORMs by 25.2%, positioning it as a promising paradigm for agentic reward modeling.

Benchmark performance is the dominant standard for comparing large language models (LLMs), typically summarized by a single aggregate score, such as accuracy, pass rate, or rubric score. This aggregate score provides limited information about a model’s underlying capabilities, yet understanding these capabilities is crucial for model development and evaluation. To this end, we propose Tree-Induced Diagnostic Evaluation (TIDE), which recasts diagnostic evaluation as behavior-driven decision-tree induction to perform fine-grained capability diagnosis. TIDE recursively builds a hierarchical capability tree from benchmark samples and model scores: at each node, an LLM judge proposes candidate splits and performs sample-to-group classification, and behavior-based information gain selects among them. We apply TIDE to five models across four benchmarks to demonstrate its effectiveness, and find fine-grained capability contrasts invisible at the aggregate level: Claude Opus 4.6 achieves 100% on microeconomic market reasoning where GPT-5.4 scores 83.3%, but trails GPT-5.4 on contract law reasoning with 77.8% versus 88.9%. We hope TIDE can be extended to past and future benchmarks so that the rich behavioral signal in evaluation results is fully leveraged for understanding model capabilities.

Reward models (RMs) are the surrogate objectives in reinforcement learning from human feedback (RLHF), and their scores directly steer policy optimization. We show that standard RM training is vulnerable in data subsets where response quality depends only weakly on the context: such instances encourage the RM to ignore the context, leading to context neglect and degraded accuracy. To address this failure mode, we propose Distribution-Aware Reward Modeling (DARM), which augments the RM objective with a conditional mutual information regularizer that maximizes context and the predicted reward conditioned on the response. By explicitly preserving the sensitivity of reward signals to the prompting context, DARM reduces over-reliance on response-only features and improves robustness to contextual variation. Extensive experiments across in-distribution and out-of-distribution settings show that DARM trained RMs deliver more accurate and consistent scoring than strong baselines. We further evaluate its downstream impact in RLHF, where DARM produce better aligned policies. We also demonstrate the necessity of each DARM design component and the impact of key parameters on performance through ablation experiments.

Reinforcement Learning (RL) in real-world environments often suffers from ambiguous or incomplete reward supervision, which undermines policy stability and generalization. Such noise may cause models to ignore key information or even collapse in advantage estimation. We find that a strong value model is essential for absorbing unstable signals and producing reliable advantages, offering denser and more robust supervision than the reward model. To better optimize noisy supervision, we propose VRPO, a framework that enhances value modeling for robust RL in LLM post-training. VRPO integrates (1) auxiliary losses guided by entropy and perplexity from a frozen language model, and (2) a variational information bottleneck, enabling the value model to filter noise and capture key words. This design allows the value model to correct noise rewards and generate more reliable advantage estimates, transforming it from a passive predictor into an active noise regulator. Experiments on multi-turn dialogue, math reasoning, and science QA with both rule-based and model-based rewards show that VRPO consistently outperforms baselines such as PPO and GRPO. Our work highlight the central role of the value model in Robust RL and provide a principled and practical approach to policy optimization under noisy supervision.

Sentence intention matching is vital for natural language understanding. Especially for Chinese sentence intention matching task, due to the ambiguity of Chinese words, semantic missing or semantic confusion are more likely to occur in the encoding process. Although the existing methods have enriched text representation through pre-trained word embedding to solve this problem, due to the particularity of Chinese text, different granularities of pre-trained word embedding will affect the semantic description of a piece of text. In this paper, we propose an effective approach that combines character-granularity and word-granularity features to perform sentence intention matching, and we utilize soft alignment attention to enhance the local information of sentences on the corresponding levels. The proposed method can capture sentence feature information from multiple perspectives and correlation information between different levels of sentences. By evaluating on BQ and LCQMC datasets, our model has achieved remarkable results, and demonstrates better or comparable performance with BERT-based models.