RSPO: Reward-Swap Policy Optimization for Multi-Turn LLM Agents

Reinforcement learning holds significant potential for training large language models (LLMs) to handle multi-turn interactive tasks. However, in long-horizon, multi-turn tasks characterized by sparse outcome rewards, directly training with outcome rewards often results in slow convergence due to the sparsity of signals and the lack of fine-grained feedback. Furthermore, the model may fail to learn successful trajectories that are not sampled during training, thereby limiting its performance. Conversely, while employing customized dense process rewards provides richer signals and accelerates convergence, these surrogate rewards may exhibit potential misalignment with the ground-truth outcome rewards. This inconsistency can bias the training direction and ultimately degrade the model's final performance. In this work, we propose Reward-Swap Policy Optimization (RSPO), a method designed to leverage the rich information from dense process rewards to facilitate training with outcome rewards. By utilizing a reward-swap mechanism, RSPO ensures the diversity of sampled trajectories while guaranteeing consistency between the optimization objective and the true outcome rewards, thereby elevating the performance ceiling of the model. We conduct extensive experiments on two challenging agent benchmarks, WebShop and ALFWorld. By applying our method to various reinforcement learning algorithms, including GRPO, PPO, and GiGPO, we demonstrate that RSPO achieves consistent performance improvements across different baselines and benchmarks.

Bridging Interleaved Multi-Modal Reasoning as a Unified Decision Process

Unified multi-modal models (UMMs) have shown promising interleaved text-image reasoning capabilities, yet effectively optimizing such multi-turn generation via reinforcement learning (RL) remains an open challenge. Existing approaches apply RL exclusively to text steps, relegating image generation to supervised surrogates, preventing policy gradients from propagating through the full interleaved trajectory across heterogeneous modalities. This leaves the potential of RL for UMMs largely untapped. In the paper, we introduce \textbf{BRAID} (\textbf{B}ridging inte\textbf{R}le\textbf{A}ved mult\textbf{I}-modal reasoning as a unified \textbf{D}ecision process), a simple framework that casts multi-turn text-image-text reasoning as a unified Markov decision process (MDP), enabling joint optimization of textual and visual generation via a single, principled RL objective. BRAID computes a shared trajectory-level advantage and propagates it coherently into both text tokens and image denoising paths, each optimized through its modality-native policy gradient mechanism. To further address long-horizon credit assignment, BRAID employs a vision-language model (VLM) judge that scores each intermediate image on its reasoning utility, supplying dense turn-level feedback to sharpen learning at critical visual branches. Experiments on spatial reasoning and visual perception benchmarks show that BRAID consistently outperforms various baselines, confirming that a unified MDP formulation with vision-thinking guidance is essential for effective multi-modal reasoning.

Latent Visual Cache for Video Reasoning

Video reasoning requires Large Multimodal Models (LMMs) to remain grounded in dense evidence, yet existing systems largely adopt "read-once, generate-many" paradigm, in which visual grounding weakens during generation. This phenomenon has been widely observed and is known as Visual Anchoring Decay. To fill this gap, we introduce Latent Video Cache (Latent-VC), a recurrent latent visual cache inserted into the decoder to preserve compact visual memories throughout reasoning. The cache is trained with supervised contrastive cache alignment and vision-grounded GRPO with a latent grounding reward, while maintaining strict train-inference alignment through native decoder hidden states. Built on Qwen3.5-9B, Latent-VC consistently outperforms strong CoT and SFT+GRPO baselines across six video benchmarks, with especially clear gains on grounding-intensive and long-video tasks. In addition, it also achieves higher accuracy with substantially shorter responses, suggesting that latent visual caching improves video reasoning by preserving visual evidence rather than relying on longer textual chains.

OmniView-Space: Reinforcing Spatial Reasoning via Multi-Perspective Spatial Mapping

Spatial intelligence remains a persistent challenge for Multimodal Large Language Models (MLLMs), as it requires coherent spatial scene representations beyond basic object recognition. Existing methods typically build such representations through textual reasoning or 3D reconstruction. However, they often falter during multi-step reasoning, particularly when required to dynamically re-anchor evidence to the specific camera-, object-, or direction-centric reference frames demanded by complex queries. To address this, we propose OmniView-Space, a framework designed to maintain spatial consistency through multimodal egocentric evidence. Our approach consists of three core components: (1) Multi-Perspective Spatial Mapping (MPSM), which re-anchors reconstructed geometry into a query-aligned visual cognitive map and a textual spatial graph; (2) Tool-Guided Egocentric Reasoning, an interleaved policy trained to actively select the ego anchor required by the query and request the corresponding MPSM evidence; and (3) Cognitive-Map Distillation, which uses MPSM-generated trajectories and ego-frame rewards to train the model to reason with self-generated cognitive maps. Experiments on single- and multi-image spatial reasoning benchmarks show that OmniView-Space achieves state-of-the-art performance. Furthermore, the distilled model maintains this performance while reducing reliance on external geometry pipelines.

Internalizing the Future: A Unified Agentic Training Paradigm for World Model Planning

Large language model (LLM) agents have demonstrated strong capability in sequential decision-making, yet they remains fundamentally reactive in long-horizon tasks. Unlike humans who employ "what-if" reasoning to evaluate potential plans before commitment, standard agents lack an internal world model to simulate future outcomes. Therefore, we propose to internalize future-aware planning by training a single autoregressive model to verbalize both a prospective state rollout and a plan-conditioned success estimate-a textual analogue of the Q-value. Crucially, we identify a format-capability gap: simply fine-tuning agents on look-ahead traces during post-training leads to superficial mimicry of foresight without genuine predictive grounding. To bridge this gap, we introduce a three-stage training paradigm: (i) World Model Agentic Mid-Training (WM-AMT) to inject latent predictive capabilities into the policy; (ii) Format-Eliciting SFT (FE-SFT) to structure this injected capability; and (iii) Foresight-Conditioned Reinforcement Learning (FC-RL) to refine the calibration and utility of the generated simulations. Evaluated on search and mathematical reasoning tasks, our approach consistently outperforms other training baselines. Our results demonstrate that effective internal world modeling in LLM agents requires a capability-first training pipeline to achieve grounded and calibrated foresight.

Breaking the Evaluation Paradox: Evaluating High-Entropy Search with Computationally Irreducible Constraints

Evaluating the exhaustive search capabilities of large language models (LLMs) is plagued by a fundamental paradox: verifying completeness requires complete ground truth, yet high-entropy enumeration tasks make such ground truth impossible for humans to create. This causes benchmarks to systematically penalize models for outperforming their human annotators. Despite rapid progress in web-search and deep research agents -- which now issue hundreds of queries, traverse diverse sites, and synthesize long reports -- evaluation still largely relies on partially annotated answer sets, LLM-based judges, or single-answer questions that avoid genuinely exhaustive search scenarios. We break this paradox by shifting the evaluation paradigm from simulating a messy reality to constructing computationally pure challenges. We introduce VERITAS (Verifiable Traversal Assessment for Search), a framework built on the principle of computationally irreducible constraints. By introducing novel, non-optimizable constraints, we create verifiable, sparse-answer search tasks that are computationally equivalent to exhaustive enumeration. These constraints are easy to verify but impossible for LLMs or search engines to optimize, forcing agents to genuinely traverse the entire search space. VERITAS can automatically generate a virtually infinite number of test cases with perfect ground truth and precise difficulty control, with marginal instance cost dominated by hash computations. This provides not only a robust benchmark for evaluating systematic exploration under uncertainty but also a scalable method for generating training data to improve these crucial, yet underdeveloped, capabilities.

From Chatbot to Digital Colleague: The Paradigm Shift Toward Persistent Autonomous AI

Large Language Models (LLMs) are undergoing a fundamental transformation from conversational generators into integrated AI systems capable of reasoning, action, memory, and self-improvement. We conceptualize this transition as a shift from Chatbot to Digital Colleague: from conversational answers to persistent work. We organize this transition along two tightly coupled dimensions. First, at the cognitive core level, LLMs are advancing from Chatbot-era "fast thinking" systems driven by next-token prediction toward Thinking LLMs that leverage inference-time computation, Chain-of-Thought reasoning, reflection, process supervision, and reinforcement learning to support more deliberate and reliable cognition. Second, at the tool-augmented task execution level, LLMs are progressing from tool-calling Agents that invoke external resources in an ad hoc manner toward OpenClaw-style workstation systems (OpenClaw) equipped with persistent Workspaces, skills, verification loops, and governance. The "Workspace + Skill" paradigm makes episodic tool use colleague-like via state persistence, reusable procedures, task closure, and experience reuse. We examine data construction shifts from instruction-response pairs to State-Action-Observation trajectories and evaluation from static benchmarks to sandboxed, auditable, self-evolving AI ecosystems.

Skill Is Not Document: A Query-Conditional Benchmark and Two-Stage Retriever for LLM Agent Skill Routing

LLM agents complete complex tasks by composing multiple skills, and skill retrieval is a front-end stage for agents. Skill retrieval differs fundamentally from traditional document retrieval at the supervision level: top-K joint correctness depends not only on the semantic relevance of each individual query-skill pair, but also on whether the skills retrieved together can collaborate to fulfill the task under the given query. Such "skill compatibility" cannot be derived from independent relevance alone. Yet existing LLM-based data synthesis pipelines can produce a direct supervision signal for "which skills should not be jointly retrieved under this query" -- namely the LLM's own rejection decisions -- and this signal is routinely discarded as low-quality data. To address this gap, we propose Reject-as-Resource Retriever (R3) and construct R3-Skill, a bilingual (Chinese-English) skill retrieval benchmark targeting realistic agent skill routing. R3-Skill spans four language directions, features query phrasings close to real user requests, and is verified through multi-expert cross-checking. On R3-Skill, we build a two-stage retrieval system (R3-Embedding + R3-Reranker) with skill compatibility as an explicit training signal. Gradient analysis shows that the "push-away" signal is diluted by bilateral balancing in the bi-encoder but acts as lossless graded ranking supervision in the cross-encoder -- motivating its placement at the cross-encoder stage, as confirmed by ablations on two datasets. The R3-Embedding + R3-Reranker pipeline attains Hit@1 = 0.7714, NDCG@10 = 0.8327 and Set-Compat = 0.3525 on R3-Skill. The dataset, training code and model weights are released as open source for agent skill routing.

Moment-Video: Diagnosing Temporal Fidelity of Video MLLMs on Momentary Visual Events

Video multimodal large language models (MLLMs) have made rapid progress on general and long-form video understanding, yet their ability to preserve brief answer-critical visual evidence remains underexplored. Many practical questions are determined by momentary visual events: localized actions or state transitions that may last only a few frames. Such evidence can be skipped by sparse frame sampling, suppressed by visual-token compression, or diluted by coarse temporal aggregation, causing failures that language-side reasoning cannot reliably recover. We introduce Moment-Video, a benchmark for diagnosing the temporal fidelity of video MLLMs through momentary visual event understanding. Each question is grounded in a localized, visually observable, and sampling-sensitive event, requiring models to notice, count, describe, or reason about transient evidence rather than rely on persistent objects, global scene context, or language priors. Moment-Video contains 1,000 human-verified video-QA pairs across 7 domains and 25 fine-grained subcategories, covering four task types: Temporal Occurrence, Temporal Counting, Action Description, and Temporal Reasoning. We evaluate 33 proprietary and open-source MLLMs on Moment-Video. The best-performing model, Seed-2.0-Pro, achieves only 39.6% overall accuracy, while most open-source models remain below 25%, revealing a substantial gap in momentary visual event understanding. Diagnostic analyses show that denser frame sampling improves some models but does not eliminate the bottleneck, and longer videos introduce stronger temporal-localization challenges. These findings suggest that current video MLLMs still lack temporally faithful representations for capturing, preserving, and using brief but decisive visual evidence.

MoG: Mixture of Experts for Graph-based Retrieval-Augmented Generation

Retrieval-augmented generation is intensively studied to ground large language models on external evidence. However, retrieving from a unified knowledge base could inevitably introduce irrelevant information that may mislead generation for complex reasoning. Inspired by the conditional computation of mixture of experts (MoE), where a router sparsely selects specialized experts alongside shared ones for each input, we propose \textbf{M}ixture \textbf{o}f experts for \textbf{G}raph-based Retrieval-Augmented Generation, i.e., \textbf{MoG}. It organizes knowledge into two core components: (i) diverse, always-accessible hub graphs that encode semantically and structurally central knowledge and provide contextual clues for expert activation, and (ii) sparsely activated expert graphs that contain domain-specific evidence. MoG first accesses hub graphs to identify general evidence and derive contextual clues. Then, a topology-aware router dynamically activates a limited set of expert graphs conditioned on the query, thereby confining retrieval to a focused evidence subspace. Extensive experiments on challenging benchmarks show that MoG consistently outperforms strong baselines, with over 20\% relative improvement on MuSiQue. Our code is available in https://github.com/DEEP-PolyU/MoG.