Long-horizon LLM agents increasingly rely on retrieved procedural skills as external memory. However, existing skill-augmented RL frameworks typically supervise a bundle of retrieved skills only with trajectory-level outcome rewards, leaving two key questions unanswered: whether each skill was executed and whether it helped task completion. We formulate this as a per-skill credit assignment problem. To address it, we propose SkillRubricsRL, which pairs each skill with an applicability-aware trajectory-level rubric and derives two complementary signals: a fulfillment signal for training correct skill execution and a counterfactual utility signal for retrieval and skill-bank evolution. Across ALFWorld and WebShop, SkillRubricsR achieves 97.7% ALFWorld success, 84.4% WebShop success, and a WebShop score of 92.7, improving over the strongest prior multi-skill baseline by 2.8, 2.1, and 1.6 percentage points, respectively, while increasing skill execution rate by over 50 percentage points. Code is available at https://github.com/gminiheimdallr-a11y/SkillRubricsRL.

Entropy mechanism emerges as a key organizing principle for understanding and improving Large Reasoning Models (LRMs). This survey examines how entropy shapes both their training and inference behavior. On the training side, we take a mathematical perspective: casting existing RL-based reasoning algorithms into a unified objective and using this formulation to derive an entropy-centric decomposition of methods, clarifying how different approaches adjust exploration–exploitation "knobs". On the inference side, based on the characteristic of LRM that trading increased inference tokens yields better performance, we summarize methods that leverage entropy to enhance inference performance or reduce uncertainty. Finally, we discuss open challenges and future directions of entropy-driven research for LRMs. Our repository is available on https://anonymous.4open.science/r/Awesome_LRM_with_Entropy-0503.

Compact open-source language models lag behind their larger counterparts in agentic tool-use reliability, yet standard remedies face fundamental obstacles: supervised fine-tuning suffers from exposure bias, while reinforcement learning is hampered by sparse credit assignment over long tool-interaction trajectories. We introduce Skill-CDPO, a progressive framework that first acquires tool-use skills at inference time through static tool analysis and dynamic strategy refinement, then distills the resulting error-correction signals into parameter updates via Critical Step DPO (CDPO). CDPO identifies the specific trajectory steps where model capability is the bottleneck—through rollout divergence between a local policy and an expert model—and constructs distributional preference pairs from all cross-group rollouts at those steps, weighted by both step-level criticality and pair-level score gaps. This provides dense, fine-grained supervision without requiring a process reward model. We evaluate Skill-CDPO on three medical agent benchmarks—PubMed Search (a new PubMed-based deep research benchmark we contribute), CureBench, and MedBrowseComp—using an 8B-parameter deep research model. Skill-CDPO substantially outperforms SFT and trajectory-level DPO baselines and achieves competitive or superior performance compared to GPT-5.2 on retrieval-intensive tasks. Our code and data are available at https://anonymous.4open.science/r/Skill-CDPO-BF51.

Medical reasoning models remain constrained by parametric knowledge and are thus susceptible to forgetting and hallucinations. DeepResearch (DR) models ground outputs in verifiable evidence from tools and perform strongly in general domains, but their direct transfer to medical field yields relatively limited gains. We attribute this to two gaps: task characteristic and tool-use scaling. Medical questions require evidence interpretation in a knowledge-intensive clinical context; while general DR models can retrieve information, they often lack clinical-context reasoning and thus “find it but fail to use it,” leaving performance limited by medical abilities. Moreover, in medical scenarios, blindly scaling tool-call can inject noisy context, derailing sensitive medical reasoning and prompting repetitive evidence-seeking along incorrect paths. Therefore, we propose DeepMed. For data, we deploy a multi-hop med-search QA synthesis method supporting the model to apply the DR paradigm in medical contexts. For training, we introduce a difficulty-aware turn-penalty to suppress excessive tool-call growth. For inference, we bring a monitor to help validate hypotheses within a controlled number of steps and avoid context rot. Overall, on seven medical benchmarks, DeepMed improves its base model by 9.79% on average and outperforms larger medical reasoning and DR models.

Large language models deployed as multi-step agents are increasingly trained via Direct Preference Optimization (DPO) on trajectory-level comparisons. Yet standard DPO weights every step uniformly, penalizing correct early actions in a losing trajectory as severely as the causal error itself. We propose Divergence-Weighted DPO(dwDPO), a drop-in replacement that detects the first divergence step between competing trajectories using implicit rewards already computed by DPO, then concentrates the loss around it via an asymmetric kernel with a closed-form scale factor that preserves the Bradley-Terry preference model. dwDPO recovers standard DPO as a special case and provably achieves higher gradient concentration at the critical step than both uniform and monotonic-discount baselines. Extensive experimental results across six deep-research benchmarks demonstrate the superiority of dwDPO over state-of-the-art open-source baselines of comparable scale at no additional inference cost.

Multimodal large language models (MLLMs) demonstrate strong capabilities in multimodal understanding, reasoning, and interaction but still face the fundamental limitation of hallucinations, where they generate erroneous or fabricated information. To mitigate hallucinations, existing methods annotate pair-responses (one non-hallucination vs one hallucination) using manual methods or GPT-4V, and train alignment algorithms to improve the correspondence between images and text. More critically, an image description often involve multiple dimensions (e.g., object attributes, posture, and spatial relationships), making it challenging for the model to comprehensively learn multidimensional information from pair-responses. To this end, in this paper, we propose RRHFV, which is the first using rank-responses (one non-hallucination vs multiple ranking hallucinations) to mitigate multimodal hallucinations. Instead of using pair-responses to train the model, RRHF-V expands the number of hallucinatory responses, so that the responses with different scores in a rank-response enable the model to learn rich semantic information across various dimensions of the image. Further, we propose a scene graph-based approach to automatically construct rank-responses in a cost-effective and automatic manner. We also design a novel training objective based on rank loss and margin loss to balance the differences between hallucinatory responses within a rankresponse, thereby improving the model’s image comprehension. Experiments on two MLLMs of different sizes and four widely used benchmarks demonstrate that RRHF-V is effective in mitigating hallucinations and outperforms the DPO method based on pair-responses.

Temporal knowledge graph reasoning (TKGR) is a crucial task that involves reasoning at known timestamps to complete the future facts and has attracted more and more attention in recent years. The current TKGR models are mainly based on graph neural networks or tensor decomposition techniques. Few works in TKGR focus on pre-trained language models (PLMs) which have powerful sequence modeling capabilities to capture the temporal associations between facts. In this paper, we propose a model SALMON: a Structure-Aware Language Model with logicality and densification strategy. Specifically, we design a PLM-based framework with a structure-aware layer inside to jointly capture the temporal evolving pattern and structural information in TKGs. To further enhance the model’s ability to infer causal associations of facts, we propose a logical judging module, which can guide the model to prioritize learning the most relevant evolving information of logical causal associations in TKGs during the training process. Moreover, we propose a densification strategy based on large language models, through a carefully crafted Chain of Thought prompt, to dig out some knowledge necessary for reasoning about fact associations, thereby making the model perform better. Extensive experimental results demonstrate the superiority of our model over the state-of-the-art baselines.