Event Causality Extraction (ECE) aims to extract causal event pairs from text. Existing methods overlook the interplay between causal event pairs and their corresponding textual evidence (e.g., causal event mention pairs), and fail to effectively leverage global causal dependency information. To address these issues, we propose a Mention-Level Causal Evidence and Global Causal Graph Reasoning (MLCE-GCGR) framework to enhance ECE. First, we introduce an auxiliary Event Mention Causality Extraction (EMCE) task, which extracts causal event mention pairs, to provide evidence for the main ECE task, and design a Dual-Level Interaction Enhancement (DLIE) strategy to enhance the bidirectional interplay between event-level and mention-level causality. Second, we develop a Global Causal Graph Reasoning (GCGR) module that simulates human-like multi-turn reasoning, aiming to progressively refine the causal graph by capturing global dependencies among event mentions, types, and arguments. Experiments on four benchmark datasets show that our method outperforms state-of-the-art approaches. Moreover, by extracting causal event mention pairs as supporting evidence, our approach improves the interpretability of structured causality extraction.

On-policy distillation (OPD) has become a key paradigm in LLM post-training, yet its training dynamics remain poorly understood. We present a systematic study examining the role, pathologies, and regulations of OPD. We first clarify the role of OPD as an exploration catalyst: it steers the student toward correct reasoning paths via dense token-level guidance, without expanding capability ceiling. We confirm this by showing that prompt diversity matters more than per-problem sampling numbers, and critically, that the effectiveness of OPD hinges entirely on the quality of its guiding signal. This dependency exposes two pathologies that derail exploration. The Student-Teacher Mismatch occurs when a large teacher-student distributional gap causes the guiding signal to misalign with task correctness, steering exploration in counterproductive directions. Length Exploitation arises when the aggregated token-level objective creates length-dependent shortcuts, allowing the student to game the reward landscape through response truncation or redundant padding, exploring degenerate length modes rather than reasoning strategies. To tame these pathologies, we investigate lightweight signal regulations: advantage clipping and log-scale compression, ensuring exploration is guided by faithful signals. Experiments across seven benchmarks demonstrate that these regulations alleviate length exploitation and enable effective distillation, stably surpassing OPD variants and RLVR baselines, thereby confirming that well-regulated signal quality, rather than mere teacher scale, governs successful exploration in OPD.

Large language models (LLMs) can call tools effectively, yet they remain brittle in multi-turn execution: after a tool-call error, smaller models often fall into repetitive invalid re-invocations instead of interpreting the feedback and recovering. This failure mode persists because current training paradigms do not explicitly teach models how to recover from execution errors. In particular, standard reinforcement learning (RL) collapses rich failure experience into sparse negative rewards, while pre-collected error-correction datasets become mismatched to the policy’s evolving failure modes. To bridge this gap, we propose Fission-GRPO, a framework that converts execution errors into on-policy corrective supervision within the RL training loop. Our core mechanism fissions each failed trajectory into a new training instance by augmenting it with diagnostic feedback from a fine-tuned Error Simulator, then resampling multiple recovery rollouts on-policy. This enables the model to learn from the precise errors it makes during exploration, rather than from static, pre-collected error cases. On BFCL v4 Multi-Turn, Fission-GRPO improves the error recovery rate of Qwen3-8B by 5.7% absolute and overall accuracy by 4.0% (from 42.75% to 46.75%), outperforming both RL baselines and specialized tool-use agents. The method further generalizes to TAU-Bench and TAU2-Bench, achieving leading results across most settings with gains up to +17.4%.

Reinforcement learning for user-centric agents is limited by the cost, latency, and risk of collecting online feedback, as well as by the lack of counterfactual comparisons under the same user state. In this paper, we propose World Model-Guided Reinforcement Learning via counterfactual user engagement simulation (WMG-RL), a framework in which a frozen user simulator provides reward supervision before real user exposure. Motivated by language world models, we instantiate the simulator as a User Engagement World Model (UEWM), which treats a recommended item as the agent action and the user's heterogeneous feedback as the environment observation. Rather than learning one fixed environment transition, UEWM learns to infer user-specific dynamics from engagement history and apply them to candidate items. In WMG-RL, a downstream policy proposes multiple candidate items for the same history; UEWM predicts the corresponding engagement feedback in parallel; and the simulated feedback is converted into dense rewards for policy optimization. Experiments show that UEWM provides reliable and transferable reward signals across domains, and that WMG-RL enables a compact 1.7B student policy to match or surpass much larger LLMs on downstream recommendation tasks.

Supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR) are two widely used post-training paradigms for improving the reasoning ability of large language models (LLMs). Recent methods attempt to integrate SFT and RLVR in a single stage by reweighting or scheduling their objectives. However, such coupling can be counterproductive because supervised updates are not uniformly beneficial for reward optimization, which can diminish reward gains. To address this, we propose \textsc{BRIDGE}, a scalable framework in which SFT learns to supervise RL by selectively transferring knowledge that improves reward optimization. Specifically, \textsc{BRIDGE} employs two nested optimization loops during meta-training: the inner loop updates base model parameters using a fused SFT--RL gradient. Concurrently, the outer loop updates a lightweight low-rank adapter (LoRA) to coordinate the two objectives by maximizing a reward-gap signal, defined as the reward of joint SFT--RL training over an RL-only baseline. Across three model scales and five reasoning benchmarks, \textsc{BRIDGE} consistently outperforms two-stage cold start, naive mixing, and representative single-stage integration baselines, yielding over three points average absolute improvement and more stable training dynamics.

Large language models (LLMs) have demonstrated remarkable performance on standard benchmarks, yet it remains largely unexplored whether they truly meet user expectations. Existing evaluation approaches, relying on model heuristics, expert rubrics, or user simulation, fail to capture the diversity and subtlety of real human expectations, causing models to appear competent while misaligning with what users actually seek. we present the first systematic study of user expectations in real-world LLM interactions, proposing a principled procedure to extract semantically rich expectations and introducing ExpectBench, a benchmark grounded in real user expectations. Analyses reveal that current LLMs struggle to satisfy and anticipate what users hope to obtain, highlighting a fundamental source of misalignment. Building on these observations, we propose LENS, a lightweight latent expectation–aware response generation framework. LENS enables models to internalize user expectations and generate better-aligned responses, consistently improving expectation satisfaction and underscoring the importance of explicitly modeling user expectations for realistic human–AI alignment.

Large Reasoning Models (LRMs) have demonstrated impressive capabilities in many tasks, yet they struggle with reliably following multiple instructions, either by failing to satisfy individual constraints or by struggling to balance competing constraints simultaneously. We formalize this challenge as the Constraint Adherence Problem (CAP). This paper introduces a novel framework that addresses CAP by representing instructions as a structured knowledge graph of constraints. Our approach, Constraint Relationship Graph Completion (CRGC), explicitly models relationships between constraints, identifies adherence challenges, and discovers "bridge constraints" that help the model better focus on and reconcile requirements. Bridge constraints act as auxiliary instructions that make primary constraints more salient and compatible. Unlike existing approaches that enhance instruction following through general training methods, CRGC specifically improves constraint satisfaction by leveraging the model's own knowledge to create better pathways for generation. Experiments across three popular instruction following datasets demonstrate that our approach reduces constraint violations by 39% compared to standard prompting while maintaining reasoning abilities of large reasoning models.

Processing long-context information remains a major challenge for large language models (LLMs). Retrieval-Augmented Generation (RAG) alleviates this issue by fetching relevant passages, yet it often wastes retrieval budget on near-duplicates and disperses mutually relevant evidence across an oversizes prompt. Cognitive studies suggest that human readers instead maintain only a few topical chunks in working memory and selectively reactivate details when needed. Inspired by this principle, we propose Topic-Graph-based RAG (TGRAG) for long document question answering (LDQA). Specifically, TGRAG (1) segments a document into sentence-level units and assigns each a high-level topic label; (2) connects sentences via lightweight topic-and similarity-based edges to construct a sparse topic graph that approximates coarser topical clusters; and (3) activates the k most relevant sentences together with their strongest neighbors at query time. This graph-guided retrieval preserves cross-sectional context while reducing redundancy, thereby improving the activation of relevant information for long-term reasoning. Experiments on three LDQA benchmarks demonstrate that TGRAG outperforms competitive RAG and GraphRAG baselines, improving over the strongest baseline by 0.8, 1.0, and 2.8 absolute accuracy points on NarrativeQA, HotpotQA, and MuSiQue, respectively (averaged over strict and lenient LLM-based rating criteria), which validates its effectiveness.

When asked what a meme or sarcastic post means, Large Vision Language Models (LVLMs) tend to describe what the image shows rather than what the author is trying to communicate. Standard instruction tuning entangles a post's literal content with its pragmatic meaning, letting surface-level details contaminate the final response. We reframe meme understanding as a problem of literal-pragmatic decomposition and propose Intent Projection, a framework that separates the two signals at the representation, output, and objective levels within a single LVLM backbone. At the representation level, an orthogonal projection module removes dominant unimodal directions from the fused image-text representation, retaining only the pragmatic residual, while a surface-real affect classifier anchors the decoder with a discrete tag that names the polarity gap. At the output level, the model externalizes a structured reasoning chain, and at the objective level a contrastive reward explicitly penalizes answers that restate the literal description. Across six multimodal benchmarks, Intent Projection consistently outperforms open-source baselines and narrows the gap to proprietary models, with the largest gains on high-divergence posts where literal collapse is most damaging.

Long Document Question Answering (LDQA) remains a critical challenge for LLMs, as lengthy contexts often exceed context window and contain irrelevant distractors. We propose Topic-Chain Compression, a novel prompt compression framework that organizes documents into topic chains. The method first splits the document into sentences, each labeled with a topic derived by a Large Language Model (LLM). Sentences sharing the same topic are linked sequentially, forming coherent topic chains. Relevance of a chain to a given query is assessed via a Cross-Encoder at the sentence level, with chain scores obtained by aggregation. The most relevant chains are retained to form a compressed prompt for final inference. TCC balances structural coherence with query sensitivity, enabling more accurate and efficient long-document reasoning in LLMs. Empirical results demonstrate the effectiveness of the framework compared with baseline methods.

Integrating memory components into large language models (LLMs) can improve the generation quality for long-term conversations. However, existing memory management methods largely overlook the cognition and regulation of the memory process, lacking the capability to dynamically manage and utilize memory on demand. To address this challenge, this paper approaches Meta-Memory for Memory Management (M$^{4}$), a novel paradigm that equips LLMs with the ability for self-monitoring and self-reflective memory management. In long-term conversations, where dialogue history accumulates continuously, the meta-memory capability of M$^{4}$ enables LLMs to autonomously 1) identify what knowledge needs to be memorized; 2) determine how to construct and store memory; 3) monitor the correctness and validity of the acquired information; and 4) decide when to learn more and how to retrieve information to refine their responses. Experimental results on two long-term conversation datasets and two long-term question-answering datasets demonstrate that our M$^{4}$ significantly enhances the memory management capacity of LLMs in long-term information learning, achieving more efficient storage and higher-quality response generation.

Large language models have improved dialogue systems, but often process conversational turns in isolation, overlooking the event structures that guide natural interactions. Hence we introduce EventWeave, a framework that explicitly models relationships between conversational events to generate more contextually appropriate dialogue responses. EventWeave constructs a dynamic event graph that distinguishes between core events (main goals) and supporting events (interconnected details), employing a multi-head attention mechanism to selectively determine which events are most relevant to the current turn. Unlike summarization or standard graph-based approaches, our method captures three distinct relationship types between events, allowing for more nuanced context modeling. Experiments on three dialogue datasets demonstrate that EventWeave produces more natural and contextually appropriate responses while requiring less computational overhead than models processing the entire dialogue history. Ablation studies confirm improvements stem from better event relationship modeling rather than increased information density. Our approach effectively balances comprehensive context understanding with generating concise responses, maintaining strong performance across various dialogue lengths through targeted optimization techniques.

Recent advances in video editing have been largely driven by large-scale instruction-based datasets. However, existing datasets still suffer from two critical limitations. First, target videos are commonly produced by automatic editing models, which may introduce visible artifacts and unreliable supervision signals. Second, most public datasets rely primarily on textual instructions, while lacking visual references that are crucial for precise, identity-preserving, and controllable editing. To address these limitations, we introduce RefVideo-6M, a large-scale reference-guided editing dataset containing 5 million video editing samples and 1 million image editing samples. To ensure reliable supervision, our dataset uses a construction pipeline that treats artifact-free real videos as editing targets and generates quality-filtered input conditions with multiple editing experts. In addition, it provides approximately 6 million visual references, covering diverse reference types and editing scenarios, thereby enabling models to learn fine-grained visual correspondence beyond text-only instructions. Based on RefVideo-6M, we further train a reference-guided video editing model, Ref-MoT, to evaluate the effectiveness and scalability of the proposed dataset. Extensive experiments demonstrate that RefVideo-6M provides substantially more reliable supervision than existing datasets and enables the training of powerful editing models with improved visual quality, controllability, and reference consistency. The open-source dataset is available at https://huggingface.co/datasets/RefVideo6M/RefVideo6M.

Temporal reasoning over long, multi-session dialogues is a critical capability for conversational agents. As dialogue histories grow in length and accumulate noise, existing long-context models struggle to accurately identify temporally pertinent information, significantly impairing reasoning performance. To address this, we introduce **Memory-T1**, a framework that learns a time-aware memory selection policy using reinforcement learning (RL). It employs a coarse-to-fine strategy, first pruning the dialogue history into a candidate set with temporal and retriever filters, followed by an RL agent that selects the precise evidence. The RL training is guided by a multi-level reward function optimizing (i) accuracy, (ii) evidence grounding, and (iii) temporal consistency. This temporal consistency reward provides a dense signal by evaluating alignment at both the session-level (range proximity) and the utterance-level (evidence density), enabling the agent to resolve subtle chronological ambiguities. On the Time-Dialog benchmark, Memory-T1 boosts a 7B model to an overall score of 67.0\%, establishing a new state-of-the-art performance for open-source models and outperforming a 14B baseline by 10.2\%. Ablation studies show temporal consistency and evidence grounding rewards jointly contributing to a 15.0\% performance gain.Moreover, Memory-T1 maintains robustness up to 128k tokens, where baseline models collapse, proving effectiveness against noise in extensive dialogue histories.

Expert-written natural language skills can improve tool-using agents, yet agent-authored skills perform 8-11 points worse than using no skill. This gap suggests that following procedural guidance and improving it from execution evidence are distinct capabilities. Inference time loops can repair skills but do not improve the model that writes the next one. We study how to organize execution experience from intermediate skills into training states for an optimizer. We introduce WER (Write, Execute, and Refine), a multi-phase framework that trains a Skill Optimizer outside a frozen executor. The optimizer proposes skills, a frozen agent executes each repeatedly, and a programmatic verifier scores the outcomes. The scores provide relative credit and select mixed-outcome records. Matched successful and failed trajectories from these records form the next phase's refinement states, so the optimizer learns from the consequences of its earlier outputs. On BFCL v4 multi-turn and tau2-bench, WER improves average Pass@1 over the no-skill baseline by 7.80 and 3.85 points, respectively. Under an identical refinement workflow, it outperforms the same backbone without optimizer training by 9.35 and 10.29 points. The trained 4B optimizer reaches 76.63 percent on BFCL v4, outperforming all evaluated off-the-shelf general-purpose models used as skill optimizers on average.

Reinforcement learning for coding agents increasingly relies on long-running agent harnesses to manage tool integration, repository contexts, and execution feedback. However, the native execution environments of these harnesses are inherently misaligned with policy-gradient training: environmental crashes and reward hacking corrupt outcome signals, while train-inference discrepancies decouple rollout behavior from policy updates. To address this, we present LEGO-RL, a framework that bridges native coding-agent harnesses with scalable policy-gradient optimization without modifying their internal control flow. LEGO-RL is built upon three pillars: (1) faithful optimization via in-process LLM proxying that captures raw generation streams for token-level alignment and robust trainer-side log-probability recomputation, even under harness-side compaction or re-serialization; (2) reliable execution via scalable sandbox orchestration featuring image caching and stage-wise defenses to mitigate reward hacking; and (3) observable training through an integrated plugin that automates validation and monitoring, paired with a Live UI for granular trajectory diagnostics. We evaluate LEGO-RL by training the sparse MoE model Qwen3.5-35B-A3B with GSPO across three native coding-agent harnesses. LEGO-RL improves Qwen3.5-35B-A3B across OpenHands SDK (64.0

Temporal Knowledge Graph Forecasting (TKGF) predicts future events from observed facts. Current approaches rely on historical information, ignoring crucial contemporaneous facts available at the query timestamp. This oversight degrades prediction performance as the graph evolves over time. Leveraging these facts presents a dual challenge: the pool of candidate facts is dynamic, and the impact of facts depends on the specific query. To address this, we propose Active Temporal Knowledge Graph Forecasting (ATKGF), which selects contemporaneous facts from dynamic candidate pool and integrates them into the prompt of large language model. The method employs a reinforcement learning-guided active selection strategy to dynamically curate the most salient facts for the given query. The agent’s policy is underpinned by a multi-aspect contextual representation, specifically designed to model both the query’s context and each candidate fact’s potential contribution. Experiments across multiple TKGF benchmarks show that ATKGF achieves state-of-the-art performance with substantial gains in accuracy and robustness without fine-tuning.

Generating synthetic datasets via large language models (LLMs) has emerged as a promising approach to improve LLM performance. However, LLMs inherently reflect biases in their training data, leading to a critical challenge: when models are trained on synthetic data, they may propagate and amplify the inherent biases that can significantly impact fairness and robustness on downstream tasks-a phenomenon we term bias inheritance. This work presents the first systematic investigation in understanding, analyzing, and mitigating bias inheritance. We fine-tune LLMs with a combined dataset of real and LLM-augmented data with varied bias ratio as the proportion of augmented data. Through systematic experiments across 10 classification and generation tasks, we analyze how 6 different types of biases manifest. Our results indicate that bias inheritance harms downstream task performance in bias directly-related classification and generation tasks. Then, our analysis identifies three key misalignment factors: misalignment of values, group data, and data distributions. Based on these insights, we propose three mitigation strategies: token-based, mask-based, and loss-based approaches, which can work differently on various tasks and bias, indicating the substantial challenges to mitigate bias inheritance. We hope this work can provide insights to the research of LLM data augmentation.

Capturing spatio-temporal dependencies is the key challenge for applications such as intelligent transportation and energy management. Contrastive learning-based methods have recently emerged as a promising paradigm to address this challenge for their ability to encode discriminant spatio-temporal dependencies. However, most of these methods assign binary labels to contrastive pairs according to their static relationship, which neglect the dynamic dependencies prevalent in data and fundamentally limit their forecasting ability. To address these limitations, we propose a novel Dynamic Soft Contrastive Learning (DynSCon) framework for spatio-temporal forecasting. Specifically, we design a dynamic dependency aware similarity metric through dynamic graph construction and clustering regularization to capture the variation of dependency patterns across time and space dimensions. Contrastive learning is then conducted on node representations across nodes and time supervised by dynamic soft labels derived from this metric to supplement forecasting loss. Thus the learned representation will be able to reveal the underlying dynamic spatio-temporal dependencies and forecasting quality can be enhanced. Extensive experiments on three datasets spanning from traffic flow forecasting to power generation forecasting demonstrates that DynSCon achieves state-of-the-art performance.

The hallmark of Deep Research agents lies in compositional reasoning, the capacity to aggregate distributed, heterogeneous information into coherent logical insights. However, current agentic systems are often retrieval-heavy but reasoning-light, where success is predominantly determined by simple entity-seeking rather than the multi-step aggregation of scattered evidence. To address this, we propose a data synthesis pipeline WebAggregator, designed to shift the agentic paradigm from retrieval-centric to compositional aggregation. Our approach first employs Proactive Explorer to collect interconnected knowledge, then Compositional Logic Proposer to weave knowledge into complex questions using over 12 composition guidelines derived from a rigorous deconstruction of the Deep Research problem setting. Fine-tuning on this corpus fundamentally transforms agent behavior, fostering deliberate composition reasoning and reduced tool redundancy. The resulting WebAggregator-32B surpasses GPT-4.1 and matches Claude-3.7-Sonnet on GAIA, WebWalkerQA, and XBench. To address the lack of benchmarks that emphasize both reasoning and retrieval, we introduce the WebAggregatorQA testbed, which reveals that even with perfect retrieval, top-tier models still underperformed. These results demonstrate that compositional reasoning, not retrieval, is the true performance ceiling for next-generation research agents.