Reinforcement learning for deep-search agents has largely focused on trajectory-level scoring—outcome correctness, citation-aware rewards, and evidence coverage. Yet the actions that expose supporting documents receive no targeted credit, a gap we call the reward-credit mismatch. We propose STAMP, in which a reference-based verifier judges whether each cited document supports an entity or relation in a training-time evidence graph, and first-exposure attribution traces each supported citation back to the action that first surfaced it. This step credit is injected through sign-preserving advantage modulation, which redistributes advantage across steps without changing the trajectory-level reward or the relative ranking of trajectories within each group. On BrowseComp, BrowseComp-ZH, and xbench-DS, STAMP improves the strongest GRPO baseline by +2.0/+5.5/+3.0 points under matched SFT initialization, training data, and search tools, and composes with both outcome-only and citation-rubric base rewards. Component ablations confirm that the provenance-based credit signal and the sign-preserving advantage modulation each contribute to the gains.

The rising demand for long-video understanding presents significant challenges due to massive information and complex temporal dependencies. Direct processing with Large Visual Language Models (VLMs) incurs high computational costs and risks information loss due to frame sampling. Meanwhile, existing agent-based methods often struggle with fragmented contexts and the high overhead of iterative processing. Therefore, we propose VideoNav, which introduces a new paradigm of index-centric navigation, reformulating video understanding as cost-aware structured query processing. VideoNav decouples heavy perception from online reasoning, including offline index construction and online adaptive navigation. In the offline phase, we transform unstructured video data into a structured, multi-granular index tree. During the online phase, the framework executes a top-down, coarse-to-fine navigating strategy. Leveraging the LLM agents, it efficiently locates relevant information and answers the query with navigating guidance, multi-level navigation, dual-path routing and visual inspection. Crucially, this adaptive mechanism enables early termination when high-level information suffices, minimizing redundant computation. Extensive experiments across different benchmarks demonstrate the effectiveness of our framework. VideoNav achieves decent accuracy while delivering 2.5x-5.7x cost savings and a 2.9x-5.8x speedup.

Hybrid language models combine Transformer attention with recurrent, state-space, or online-updated components to lower long-context serving cost while preserving memory fidelity. Existing surveys mostly organize this space by architecture family, chronology, or efficiency technique, leaving workload-specific memory functions underspecified. We propose State-Evidence-Access (SEA) for this constraint: bounded State may lose distinctions needed by future queries not fixed in advance. Architectures therefore allocate active State, addressable Evidence, and a per-step Access budget to jointly support reusable structure, query-conditioned recovery, and effective use. This view helps explain why attention ratios or average perplexity can be misleading: a hybrid can save memory while losing the retrieval, state-update, or access function needed by the task. We survey hybrids by allocation locus—from key-value (K-V) and head-level mechanisms through layer, block, external-memory, and system-cache policies—and characterize their *function-granularity mismatch* risks. For workloads such as long-document QA, SEA attributes failure to missing evidence, unstable state updates, access bottlenecks, or poor integration.

Reinforcement learning with verifiable rewards (RLVR) optimizes LLMs using sparse verifiable final-answer rewards. This sparse anchor reliably verifies whether a trajectory succeeds but provides no direct feedback on the reasoning path that produced it. Before success, prerequisite progress on hard problems receives no reward signal; after success, outcome rewards cannot distinguish well-organized correct trajectories from redundant or locally flawed ones. We introduce SCOPE-RL (Scaffolded Chain Optimization with Process Efficiency), a two-stage framework that densifies this anchor while retaining the GRPO update: Adaptive Scaffolded RL adds prefix-decomposed verifiable rewards on answer-hidden sub-question chains before success, and Quality-Aware Process RL applies correctness-gated process-shape rewards to refine correct trajectories after success. An expert-validated Step-Quality Evaluation Protocol evaluates useful-step density, error localization, and token efficiency beyond final-answer accuracy. On Qwen3-8B-Instruct trained on DAPO-Math and Big-Math, SCOPE-RL improves average accuracy by up to 11.2 pp and reduces reasoning tokens by up to 27.1% over outcome-only GRPO; the gains hold under GSPO and on Qwen3-0.6B-Instruct, indicating that reward-signal densification is complementary to policy-update-level RLVR advances.

Agentic search has recently emerged as a powerful paradigm, where an agent interleaves multi-step reasoning with on-demand retrieval to solve complex questions. Despite its success, how to design a retriever for agentic search remains largely underexplored. Existing search agents typically rely on similarity-based retrievers, while similar passages are not always useful for final answer generation. In this paper, we propose a novel retriever training framework tailored for agentic search. Unlike retrievers designed for single-turn retrieval-augmented generation (RAG) that only rely on local passage utility, we propose to use both local query-passage relevance and global answer correctness to measure passage utility in a multi-turn agentic search. We further introduce an iterative training strategy, where the search agent and the retriever are optimized bidirectionally and iteratively. Different from RAG retrievers that are only trained once with fixed questions, our retriever is continuously improved using evolving and higher-quality queries from the agent. Extensive experiments on seven single-hop and multi-hop QA benchmarks demonstrate that our retriever, termed Agentic-R, consistently outperforms strong baselines across different search agents.

Low-Rank Adaptation (LoRA) is widely used for parameter-efficient fine-tuning, yet merging multiple task-specific LoRA updates without additional training remains challenging. Most existing LoRA merging methods rely on SVD-based alignment, which emphasizes globally shared structure across tasks. In this work, we show that LoRA merging performance can be further improved by combining SVD with CUR decomposition. Through a representation-level analysis, we find that SVD-based decompositions primarily model shared components across tasks, while CUR-based decompositions better preserve task-specific and localized updates. These two perspectives are geometrically misaligned and exhibit complementary advantages, revealing an inherent trade-off between capturing shared structure and preserving task-specific information in LoRA model merging. Guided by this analysis, we propose a training-free merging procedure that explicitly combines the shared structure captured by SVD with the task-specific components preserved by CUR. Experiments on both vision and language benchmarks demonstrate consistent improvements over existing gradient-free LoRA merging methods.

Despite substantial advancements in aligning LLMs with human values, current safety mechanisms remain susceptible to jailbreak attacks. We attribute this vulnerability to the distributional discrepancies between alignment-oriented prompts and malicious prompts. To investigate this, and drawing inspiration from logic-driven NLP tasks, we introduce LogiBreak, a universal black-box jailbreak method that utilizes logical expression translation to bypass LLM safety mechanisms. By converting harmful natural language prompts into formal logical expressions, LogiBreak exploits the distributional gap between alignment data and logic-expressed inputs, preserving the underlying semantic intent and readability while evading safety constraints. Furthermore, to fill the gap of existing benchmarks that lack systematic resources specifically targeting logical expression-based attacks against LLM robustness, we construct a novel multilingual logical expression jailbreak dataset for evaluation. Our evaluations of LogiBreak in five languages demonstrate its effectiveness and generalizability in various linguistic contexts. The code is available at https://github.com/Applied-Machine-Learning-Lab/ACL2026_Logibreak.

Recent advances in synergizing large reasoning models (LRMs) with retrieval-augmented generation (RAG) have shown promising results, yet two critical challenges remain: (1) reasoning models typically operate from a single, unchallenged perspective, limiting their ability to conduct deep, self-correcting reasoning over external documents, and (2) existing training paradigms rely excessively on outcome-oriented rewards, which provide insufficient signal for shaping the complex, multi-step reasoning process. To address these issues, we propose an Reasoner-Verifier framework named Adversarial Reasoning RAG (ARR). The Reasoner and Verifier engage in reasoning on retrieved evidence and critiquing each other’s logic while being guided by process-aware advantage that requires no external scoring model. This reward combines explicit observational signals with internal model uncertainty to jointly optimize reasoning fidelity and verification rigor. Experiments on multiple benchmarks demonstrate the effectiveness of our method. Our code is available at link.

Clinical Decision Support Systems (CDSSs) provide reasoning and inquiry guidance for physicians, yet they face notable challenges, including high maintenance costs and low generalization capability.Recently, Large Language Models (LLMs) have been widely adopted in healthcare due to their extensive knowledge reserves, retrieval, and communication capabilities. While LLMs show promise and excel at medical benchmarks, their diagnostic reasoning and inquiry skills are constrained.To mitigate this issue, we propose (1) Clinical Diagnostic Reasoning Data (CDRD) structure to capture abstract clinical reasoning logic, and a pipeline for its construction, and (2) the Dr. Assistant, a clinical diagnostic model equipped with clinical reasoning and inquiry skills. Its training involves a two-stage process: SFT, followed by RL with a tailored reward function.We also introduce a benchmark to evaluate both diagnostic reasoning and inquiry.Our experiments demonstrate that the Dr. Assistant outperforms open-source models and achieves competitive performance to closed-source models, providing an effective solution for clinical diagnostic inquiry guidance. Project information can be found at: https://github.com/YGswu/Dr.-Assistant.

A fundamental challenge in creative writing lies in reconciling the inherent tension between maintaining global coherence in long-form narratives and preserving local expressiveness in short-form texts. While long-context generation necessitates explicit macroscopic planning, short-form creativity often demands spontaneous, constraint-free expression. Existing alignment paradigms, however, typically employ static reward signals and rely heavily on high-quality supervised data, which is costly and difficult to scale. To address this, we propose UniCreative, a unified reference-free reinforcement learning framework. We first introduce AC-GenRM, an adaptive constraint-aware reward model that dynamically synthesizes query-specific criteria to provide fine-grained preference judgments. Leveraging these signals, we propose ACPO, a policy optimization algorithm that aligns models with human preferences across both content quality and structural paradigms without supervised fine-tuning and ground-truth references. Empirical results demonstrate that AC-GenRM aligns closely with expert evaluations, while ACPO significantly enhances performance across diverse writing tasks. Crucially, our analysis reveals an emergent meta-cognitive ability: the model learns to autonomously differentiate between tasks requiring rigorous planning and those favoring direct generation, validating the effectiveness of our direct alignment approach.

Offline preference optimization aligns reasoning models from fixed chosen-rejected pairs, yet standard methods apply gradient updates from every pair regardless of its training value under the current policy. We argue that this uniform treatment is wasteful and potentially harmful. From the perspective of gradient utility, we show that a pair’s contribution depends jointly on informativeness and stability. Pair utility drifts as the policy evolves, high-gradient samples can coincide with high-curvature regions, leading to noisy and destabilizing updates, and the most effective supervision comes from stable confident errors where the model is reliably wrong yet curvature remains low. These findings motivate **SAGE** (Stability-Aware Gradient Efficiency), which maintains difficulty-stratified candidate pools refreshed during training and selects pairs within each pool by a forward-pass signal-to-curvature score. Only pairs with high current utility receive gradient computation; the rest are excluded from backpropagation. On mathematical reasoning benchmarks across multiple model scales, SAGE outperforms full-data and size-matched baselines while producing substantially smoother optimization trajectories.

On-policy distillation reduces train-inference mismatch by supervising student-generated trajectories, but full-vocabulary token-level distribution matching remains expensive for large language models. A common approximation truncates reverse-KL distillation to the student’s top-k tokens, which preserves efficiency but drops tail-token supervision and ignores support-level mass mismatch between the student and teacher. We analyze these approximation gaps and propose Stochastic Support Sampling with Teacher Mass Reweighting, which samples fixed-size supports from the student distribution and uses teacher support mass as a reliability weight for the conditional distillation loss. Under the same support budget, our method outperforms the strongest baseline by +1.4 hard-math average points on Qwen3-8B→Qwen3-1.7B-Base, and by +1.1 math average points and +2.8 code average points on DeepSeek-R1-Distill-Llama-8B→Meta-Llama-3.1-8B-Base.

Large Language Model (LLM) based listwise ranking has shown superior performance in many passage ranking tasks. With the development of Large Reasoning Models (LRMs), many studies have demonstrated that step-by-step reasoning during test-time helps improve listwise ranking performance. However, due to the scarcity of reasoning-intensive training data, existing rerankers perform poorly in many complex ranking scenarios, and the ranking ability of reasoning-intensive rerankers remains largely underdeveloped. In this paper, we first propose an automated reasoning-intensive training data synthesis framework, which sources training queries and passages from diverse domains and applies DeepSeek-R1 to generate high-quality training labels. To empower the listwise reranker with strong reasoning ability, we further propose a two-stage training approach, which includes a cold-start supervised fine-tuning (SFT) stage and a reinforcement learning (RL) stage. During the RL stage, we design a novel multi-view ranking reward tailored to the multi-turn nature of listwise ranking. Extensive experiments demonstrate that our trained reasoning-intensive reranker ReasonRank outperforms existing baselines significantly and also achieves much lower latency than the pointwise reranker.

Retrieval-Augmented Generation (RAG) is widely used to enhance Large Language Models (LLMs) by grounding responses in external knowledge. However, in real-world applications, retrievers often return lengthy documents with redundant or irrelevant content, confusing downstream readers. While evidence retrieval aims to address this by extracting key information, it faces critical challenges: (1) inability to model synergistic inter-dependencies among evidence sentences, (2) lack of supervision for evaluating multi-sentence evidence quality, and (3) computational inefficiency in navigating exponentially growing search spaces of candidate evidence sets. To tackle these challenges, we propose ETS (Evidence Tree Search), a novel framework that reformulates evidence retrieval as a dynamic tree expansion process. Our approach first constructs an evidence tree where each path represents a candidate evidence set, explicitly modeling inter-sentence dependencies through context-aware node selection. We then leverage Monte Carlo Tree Search (MCTS) to efficiently assess evidence quality and introduce an Early-Terminating Beam Search strategy to efficiently accelerate the model inference. Extensive experiments on five datasets demonstrate that ETS significantly outperforms existing methods across different readers. Our code and datasets will be released to facilitate future research.