Route Before Retrieve: Activating Latent Routing Abilities of LLMs for RAG vs. Long-Context Selection

Recent advances in large language models (LLMs) have expanded the context window to beyond 128K tokens, enabling long-document understanding and multi-source reasoning. A key challenge, however, lies in choosing between retrieval-augmented generation (RAG) and long-context (LC) strategies: RAG is efficient but constrained by retrieval quality, while LC supports global reasoning at higher cost and with position sensitivity. Existing methods such as Self-Route adopt failure-driven fallback from RAG to LC, but remain passive, inefficient, and hard to interpret. We propose Pre-Route, a proactive routing framework that performs structured reasoning before answering. Using lightweight metadata (e.g., document type, length, initial snippet), Pre-Route enables task analysis, coverage estimation, and information-need prediction, producing explainable and cost-efficient routing decisions. Our study shows three key findings: (i) LLMs possess latent routing ability that can be reliably elicited with guidelines, allowing single-sample performance to approach that of multi-sample (Best-of-N) results; (ii) linear probes reveal that structured prompts sharpen the separability of the "optimal routing dimension" in representation space; and (iii) distillation transfers this reasoning structure to smaller models for lightweight deployment. Experiments on LaRA (in-domain) and LongBench-v2 (OOD) confirm that Pre-Route outperforms Always-RAG, Always-LC, and Self-Route baselines, achieving superior overall cost-effectiveness.

LASAR: Latent Adaptive Semantic Aligned Reasoning for Generative Recommendation

Large Language Models (LLMs) have demonstrated powerful reasoning capabilities through Chain-of-Thought (CoT) in various tasks, yet the inefficiency of token-by-token generation hinders real-world deployment in latency-sensitive recommender systems. Latent reasoning has emerged as an effective paradigm in LLMs, performing multi-step inference in a continuous hidden-state space to achieve stronger reasoning at lower cost. However, this paradigm remains underexplored in mainstream generative recommendation. Adapting it reveals three unique challenges: (1) the gap between prior-less Semantic ID (SID) symbols and continuous latent reasoning - SIDs lack pre-trained semantics, hindering joint optimization; (2) representation drift due to a lack of reasoning chain supervision; and (3) the suboptimality of applying a globally fixed reasoning depth. To address these, we propose LASAR (Latent Adaptive Semantic Aligned Reasoning), an SFT-then-RL framework. First, we bridge this gap via two-stage training: Stage 1 grounds SID semantics before Stage 2 introduces latent reasoning, ensuring efficient convergence. Second, we mitigate representation drift through explicit CoT semantic alignment. Step-wise bidirectional KL divergence constrains the latent reasoning trajectory using hidden-state anchors extracted from CoT text, while a Policy Head predicts per-sample reasoning depth. Third, during the GRPO-based RL phase, terminal-only KL alignment accommodates variable-length reasoning, and REINFORCE optimizes the Policy Head to dynamically allocate steps. This nearly halves the average latent step count while simultaneously improving recommendation quality. Experiments on three real-world datasets demonstrate that LASAR outperforms all baselines. It adds marginal inference latency and is roughly 20 times faster than generating explicit CoT text.

Policy Improvement Reinforcement Learning

Reinforcement Learning with Verifiable Rewards (RLVR) has become a central post-training paradigm for improving the reasoning capabilities of large language models. Yet existing methods share a common blind spot: they optimize policies based on instantaneous group-level or batch-level statistics without ever verifying whether the resulting update actually improved the model. This open-loop design -- updating in isolation at each step, guided only by within-group (batch) reward signals -- means optimization can drift or collapse with no mechanism to detect and correct these failures. We argue that the missing ingredient is policy improvement feedback: the ability to measure and optimize inter-iteration progress directly. To this end, we introduce Policy Improvement Reinforcement Learning (PIRL), a framework that replaces surrogate reward maximization with the explicit objective of maximizing cumulative policy improvement across iterations, and prove this temporal objective is perfectly aligned with maximizing final task performance. Building on PIRL, we propose Policy Improvement Policy Optimization (PIPO), which implements closed-loop optimization through retrospective verification. At each iteration, PIPO evaluates whether the previous update yielded genuine improvement against a sliding-window historical baseline, then actively reinforces beneficial updates and suppresses the harmful ones -- transforming an open-loop process into a self-correcting one. We provide theoretical analysis showing that PIPO performs ascent on the PIRL objective in expectation, and experiments on mathematical reasoning benchmarks demonstrate improved stability and performance over GRPO and its variants.

Heterogeneous Agent Collaborative Reinforcement Learning

We introduce Heterogeneous Agent Collaborative Reinforcement Learning (HACRL), a new learning paradigm that addresses the inefficiencies of isolated on-policy optimization. HACRL enables collaborative optimization with independent execution: heterogeneous agents share verified rollouts during training to mutually improve, while operating independently at inference time. Unlike LLM-based multi-agent reinforcement learning (MARL), HACRL does not require coordinated deployment, and unlike on-/off-policy distillation, it enables bidirectional mutual learning among heterogeneous agents rather than one-directional teacher-to-student transfer. Building on this paradigm, we propose HACPO, a collaborative RL algorithm that enables principled rollout sharing to maximize sample utilization and cross-agent knowledge transfer. To mitigate capability discrepancies and policy distribution shifts, HACPO introduces four tailored mechanisms with theoretical guarantees on unbiased advantage estimation and optimization correctness. Extensive experiments across diverse heterogeneous model combinations and reasoning benchmarks show that HACPO consistently improves all participating agents, outperforming GSPO by an average of 3.3\% while using only half the rollout cost.

UniFAR: A Unified Facet-Aware Retrieval Framework for Scientific Documents

Existing scientific document retrieval (SDR) methods primarily rely on document-centric representations learned from inter-document relationships for document-document (doc-doc) retrieval. However, the rise of LLMs and RAG has shifted SDR toward question-driven retrieval, where documents are retrieved in response to natural-language questions (q-doc). This change has led to systematic mismatches between document-centric models and question-driven retrieval, including (1) input granularity (long documents vs. short questions), (2) semantic focus (scientific discourse structure vs. specific question intent), and (3) training signals (citation-based similarity vs. question-oriented relevance). To this end, we propose UniFAR, a Unified Facet-Aware Retrieval framework to jointly support doc-doc and q-doc SDR within a single architecture. UniFAR reconciles granularity differences through adaptive multi-granularity aggregation, aligns document structure with question intent via learnable facet anchors, and unifies doc-doc and q-doc supervision through joint training. Experimental results show that UniFAR consistently outperforms prior methods across multiple retrieval tasks and base models, confirming its effectiveness and generality.

UniARM: Towards a Unified Autoregressive Reward Model for Multi-Objective Test-Time Alignment

Multi-objective alignment aims to align LLM responses with multiple human preference objectives. Among existing methods, guiding the generation of frozen LLMs through autoregressive reward models (ARMs) to accomplish multi-objective test-time alignment is a low-cost solution. However, these methods typically rely on independent parameters for each preference objective, either by training ARMs independently across preference dimensions, which neglects interactions among preference features, or by training a single ARM with separate feature extraction modules for each preference, which can cause feature entanglement. Both strategies can result in misalignment between generated outputs and user preferences. To address this limitation, we propose Preference-Modulated \& Shared Low-Rank Adaptation (MoSLoRA) for ARM training, which first extracts shared features via a preference-agnostic module and then applies affine transformations to shared features via a preference modulation module conditioned on mixed preference vectors. This design mitigates feature entanglement and enables precise control over preference trade-offs during inference. Building on this, we introduce the Unified Autoregressive Reward Model (UniARM), a novel framework for multi-objective test-time alignment. UniARM jointly models all preference dimensions in a single parameter space, eliminating the need for independent parameters for each preference objective. es on larger-scale LLMs, enhancing its practical usability.

Weak-Driven Learning: How Weak Agents make Strong Agents Stronger

As post-training optimization becomes central to improving large language models, we observe a persistent saturation bottleneck: once models grow highly confident, further training yields diminishing returns. While existing methods continue to reinforce target predictions, we find that informative supervision signals remain latent in models' own historical weak states. Motivated by this observation, we propose WMSS (Weak Agents Can Make Strong Agents Stronger), a post-training paradigm that leverages weak checkpoints to guide continued optimization. By identifying recoverable learning gaps via entropy dynamics and reinforcing them through compensatory learning, WMSS enables strong agents to improve beyond conventional post-training saturation. Experiments on mathematical reasoning and code generation datasets show that agents trained with our approach achieve effective performance improvements, while incurring zero additional inference cost.

Does Your Reasoning Model Implicitly Know When to Stop Thinking?

Recent advancements in large reasoning models (LRMs) have greatly improved their capabilities on complex reasoning tasks through Long Chains of Thought (CoTs). However, this approach often results in substantial redundancy, impairing computational efficiency and causing significant delays in real-time applications. Recent studies show that longer reasoning chains are frequently uncorrelated with correctness and can even be detrimental to accuracy. In a further in-depth analysis of this phenomenon, we surprisingly uncover and empirically verify that LRMs implicitly know the appropriate time to stop thinking, while this capability is obscured by current sampling paradigms. Motivated by this, we introduce SAGE (Self-Aware Guided Efficient Reasoning), a novel sampling paradigm that unleashes this efficient reasoning potential. Furthermore, integrating SAGE as mixed sampling into group-based reinforcement learning (SAGE-RL) enables SAGE-RL to effectively incorporate SAGE-discovered efficient reasoning patterns into standard pass@1 inference, markedly enhancing both the reasoning accuracy and efficiency of LRMs across multiple challenging mathematical benchmarks.

Contextual Rollout Bandits for Reinforcement Learning with Verifiable Rewards

Reinforcement Learning with Verifiable Rewards (RLVR) is an effective paradigm for improving the reasoning capabilities of large language models. However, existing RLVR methods utilize rollouts in an indiscriminate and short-horizon manner: responses of heterogeneous quality within each prompt are treated uniformly, and historical rollouts are discarded after a single use. This leads to noisy supervision, poor sample efficiency, and suboptimal policy updates. We address these issues by formulating rollout scheduling in RLVR as a contextual bandit problem and proposing a unified neural scheduling framework that adaptively selects high-value rollouts throughout training. Each rollout is treated as an arm whose reward is defined by the induced performance gain between consecutive optimization steps. The resulting scheduler supports both noise-aware intra-group selection and adaptive global reuse of historical rollouts within a single principled framework. We provide theoretical justification by deriving sublinear regret bounds and showing that enlarging the rollout buffer improves the achievable performance upper bound. Experiments on six mathematical reasoning benchmarks demonstrate consistent gains in performance and training efficiency across multiple RLVR optimization methods.

Real-Time Aligned Reward Model beyond Semantics

Reinforcement Learning from Human Feedback (RLHF) is a pivotal technique for aligning large language models (LLMs) with human preferences, yet it is susceptible to reward overoptimization, in which policy models overfit to the reward model, exploit spurious reward patterns instead of faithfully capturing human intent. Prior mitigations primarily relies on surface semantic information and fails to efficiently address the misalignment between the reward model (RM) and the policy model caused by continuous policy distribution shifts. This inevitably leads to an increasing reward discrepancy, exacerbating reward overoptimization. To address these limitations, we introduce R2M (Real-Time Aligned Reward Model), a novel lightweight RLHF framework. R2M goes beyond vanilla reward models that solely depend on the semantic representations of a pretrained LLM. Instead, it leverages the evolving hidden states of the policy (namely policy feedback) to align with the real-time distribution shift of the policy during the RL process. This work points to a promising new direction for improving the performance of reward models through real-time utilization of feedback from policy models.